Epistemology and Congnition - Alvin I. Goldman

EPISTEMOLOGY A N D COGNITION
EPISTEMOLOGY
and
COGNITION
Alvin I. Goldman
HARVARD UNIVERSITY PRESS

Cambridge, Massachusetts, and London, England
Copyright @ 1986 by the President and Fellows of
Harvard College
All rights reserved
Printed in the United States of America
FIFTH PRINTING, 1995
Publication of this book has been aided by a

grant from the Andrew W. Mellon Foundation.
Library of Congress Catabging in Publication Data

Goldrnan, Alvin I., 1938-
Epistemology and cognition.
Bibliography: p.
Includes indexes.
1. Cognition. 2. Knowledge, Theory of.
I. Title.
(DNLM: 1. Cognition. 2. Philosophy.
BF 311 G619e
BF311.GS82 1986 153.4 85.21906
ISBN 0-674-25895-9 (cloth)
ISBN 0-674-25896-7 (paper)
In rnemmy of Frances and Nathan Goldrnun
Preface
THEIDEA of this book germinated while I was a fellow at the Center for
Advanced Study in the Behavioral Sciences during 1975-76, with support
from the Guggenheim Foundation. Conceived there was the idea of epis-
temics: an enterprise linking traditional epistemology, first, with cognitive
science and, second, with social scientific and humanistic disciplines that
explore the interpersonal and cultural processes impinging on knowledge
and belief. The intention was to enrich epistemology while preserving its
own identity. This book articulates the first part of epistemics: the relation
between epistemology and cognitive science. A sequel is planned to delin-
eate the second part: social epistemics.
In trying to lay a satisfactory conceptual foundation for epistemics, I was
aided at the Behavioral Science Center by discussions with John Perry,
Keith Donnellan, Michael Bratman, and Holly Smith, and by later debates
with Richard Nisbett during a seminar we taught jointly. Initial program-
matic statements appeared in "Epistemics: The Regulative Theory of Cog-
nition," The Journal of Philosophy, 75 (1978):509-523, and "Varieties of
Cognitive Appraisal," Now, 13 (1979):23-38. Other precursors of material
.
in this book include "What Is Justified Belief?" in G. Pappas, ed., J w t i . a -
tion and Knowledge (Dordrecht: D. Reidel, 1979), and "The Internalist
Conception of Justification," in P. French, T. Uehling, Jr., and H. Wett-
stein, ed~.,Midwest Studies in Philosophy, vol. 5, Studies in Epistemology
(Minneapolis: University of Minnesota Press, 1980).Themes from "The Re-
lation between Epistemology and Psychology," Synthese, 64 (1985):29-68
(written in 1981), form the basis for parts of Chapters 4 and 5. "Epistemol-
ogy and the Theory of Problem Solving," Synthese, 55 (1983):21-48, is the
basis for much of Chapter 6. A version of Chapter 11, "Constraints on Rep-
resentation," appears in Myles Brand and Robert Harnish, eds., The Repre-
sentation of Knowledge and Belief (Tucson: University of Arizona Press,
1986).
viii Preface
Writing of the book began in earnest in 1981-82, when I was a fellow of

the National Humanities Center. (I feel fortunate to have enjoyed the mar-
velous research opportunities of two fine centers.) Work continued in 1983
under a grant from the National Science Foundation (SES-8204737).
I owe deep thanks to many colleagues and friends for valuable comments
on portions of the manuscript at different stages of its evolution. First and
foremost, I am indebted to Holly Smith, who read virtually every version of
every chapter. Her probing criticisms and generous advice have signifi-
cantly improved the book. Other reviewers of hefty chunks of the manu-
script were Charles Chastain, Robert Cummins, Robert M. Harnish, John
Pollock, Lance Rips, and Stephen Schiffer. Kent Bach, Fred Dretske, Keith
Lehrer, and Carl Posey commented on some selected chapters. In all cases
chapters emerged somewhat scathed, but (I hope) improved. I have also
benefited from the helpful comments of many students in graduate sem-
inars, both at the University of Illinois at Chicago and at the University of
Arizona.
Contents
Introduction 1
THEORETICAL FOUNDATIONS
The Ekments of E p i s t e m o l o 13
Skepticism 28
Knowledge 42
Justification: A Rule Framework 58
Justification and Reliability 81
Problem Solving, Power, and Speed 122
Truth and Realism 142
The Problem of Content 162
part I1 ASSESSING OUR COGNITIVE RESOURCES
9 Perception 181
10 Memory 199
11 Constraints on Representation 227
12 Internal Codes 252
13 Deductive Reasoning 278
14 Probability Judgments 305
15 Acceptance and Uncertainty 324
16 Belief Updating 344
17 Production Systems and Second-Order Processes 359
Conclusion: Primary Epistemics and Cognitive Science 378
Notes 383
Illustration Credits 421
Author Index 423
Subject Index 429
Introduction
I. 1. Epistemology as a Multidisciplinary Aflair

The aim of this book is to redirect and restructure the field of epistemology.
One central theme is that epistemology should be a multidisciplinary affair,
not the province of pure, a priori philosophy. Though philosophy is the
chief conductor or orchestrator of epistemology, many other disciplines-
including empirical disciplines-are important parts of the ensemble. This
book is the first part of a larger project, and the full conception of episte-
mology I envisage will only emerge from the whole work. In this introduc-
tion I sketch the larger conception, although much of it outstrips what is
actually developed in this volume.
Epistemology, as I conceive it, divides into two parts: individual episte-
mology and social epistemology. Individual epistemology-at least primary
individual epistemology-needs help from the cognitive sciences. Cogni-
tive science tries to delineate the architecture of the human mind-brain,
and an understanding of this architecture is essential for primary epistemol-
ogy. Social epistemology needs help from various of the social sciences and
humanities, which jointly provide models, facts, and insights into social sys-
tems of science, learning, and culture. The connection between primary '
epistemology and the sciences of cognition is treated in this book. The na-
ture and structure of social epistemology will be examined in a future study.
To some readers the interdisciplinary theme may sound banal. Episte-
mology deals with knowledge, which is the property of individual minds. So
of course epistemology must be interested in the knowing mind. Similarly,
most knowledge is a cultural product, channeled through language and so-
cial communication. So how could epistemology fail to be intertwined with
studies of culture and social systems?
Despite these truisms, strong countercurrents in the history of epistemol-
ogy run against the interdisciplinary theme. Here is a sampling of such
countercurrents (not all mutually compatible, being drawn from different
traditions).
(1)As the study of method, epistemology should be autonomous. It should
2 Introduction
be prior to the sciences; so it must not seek help from them. (2) Epistemol-
ogy should only be concerned with the analysis of concepts, specifically
epistemic concepts such as 'knowledge', 'warrant', 'rationality', and the
like. But conceptual or linguistic analysis is the province of philosophy; so
epistemology needs no help from behavioral or social sciences. (3)The true
aim of philosophy is to "show the fly the way out of the fly-bottle,"' that is,
to dissolve puzzles and paradoxes that lead to skepticism. Such dissolution
requires only linguistic analysis, not a model of the mind-brain or empirical
models of intellectual influence. (4) Epistemology is the study of methodol-
ogy, and proper methodology is the province of deductive logic, inductive
logic, probability theory, and statistics. Epistemology reduces to these sub-
jects, all of which are f o m l disciplines. Empirical sciences are not needed.
(5) Epistemology is normative, evaluative, or critical, not descriptive. So
empirical sciences, which are purely descriptive, cannot help epistemology.
The psychology of reasoning, for example, cannot shed light on proper rea-
soning, on logically or scientifically sound reasoning. In view of these me-
taepistemological currents, the proposed interdisciplinary theme is
controversial, not trivial. It needs sustained clarification and defense.
Furthermore, the interdisciplinary theme is only one of my themes. Of
equal or greater importance is the specific epistemological framework to be
proposed, which specifies the particular ways in which disciplinary collabo-
ration should proceed. This framework contrasts with other systems of
epistemology that would equally favor an interdisciplinary orientation. Let
me briefly mention a few components of this framework.
One crucial component is the evaluative mission of epistemology. Along
with the dominant tradition, I regard epistemology as an evaluative, or
normative, field, not a purely descriptive one. This makes it far from obvi-
ous how positive science can have inputs to epistemology. How, exactly, do
facts of cognition or social intercourse bear on epistemic evaluations or
norms?
A few other recent characterizations of epistemology also link it with
psychology. But these characterizations depict the field as a descriptive one.
On a purely descriptive conception it is not surprising that epistemology
should be indebted to psychology-should even reduce to it. Thus, on W: V.
Quine's naturalistic conception, the epistemologist would study how the
human subject responds to certain input; how, in response to various stimu-
lus patterns, the subject delivers a description of the external world and its
history. In studying the relation between this "meager input" and "torren-
tial output," epistemology "simply falls into place as a chapter of psychol-
ogy and hence of natural s ~ i e n c e . Similarly,
"~ Donald Campbell advances a
conception of the field which he calls "evolutionary epistemology." On this
conception epistemology takes cognizance of "man's status as a product of
biological and social ev~lution."~ Campbell explicitly characterizes his
conception as descriptive: descriptive of man as k n ~ w e r . ~
Introduction 3
If epistemology is a branch of psychology, or evolutionary theory, the

field's empirical status needs no clarification. But this approach, though
perfectly tenable, neglects the evaluative strain pervading most of historical
epistemology.5 Epistemologists have traditionally been interested in
whether beliefs about the world are justified or warranted; whether we are
rationally entitled to these beliefs. Epistemologists seek to discover or in-
vent proper methods of inquiry and investigation, often dismissing estab-
lished procedures as irrational. Clearly, 'justified', 'warranted', and
'rational' are evaluative terms; and the advocacy of particular methods is a
normative activity. So traditional epistemology has a strong evaluative-
normative strain. I aim to preserve that strain.
The emphasis on evaluation invites parallels with moral philosophy and
normative social theory, and these will be important elements in my discus-
sion. In studying various epistemic terms of appraisal, I will inquire into
their basis: Is it objective or subjective? Is it deontological or consequen-
tialist? Should epistemic norms be absolutist or pluralist?
My epistemological framework prominently features an objectivist stan-
dard or set of standards. The central epistemological concepts of appraisal,
I argue, invoke true belief as their ultimate aim. So the evaluation of episte-
mic procedures, methods, processes, or arrangements must appeal to truth-
conduciveness, an objective standard of assessment. While this emphasis on
truth is hardly startling, the framework contrasts with many studies of sci-
ence and opinion that explore properties of social systems and institutions.
Specifically,studies in the history and sociology of science characteristically
shy away from considerations of truth.
Truth-linked standards may seem useless, because of circularity or vacu-
ousness. To decide whether such a standard is satisfied, we have to employ
our present beliefs about the truth. Isn't this circular? Or doesn't it imply
automatic endorsement of current procedures, by which our present beliefs
have been formed? Wouldn't it preclude criticism and revision, to which
normative epistemology ought to be committed?
None of these objections is warranted. To be sure, application of a stan-
dard requires recourse to present beliefs. In Otto Neurath's metaphor we
can only rebuild our intellectual ship while floating upon it at seas6But the
same point holds for any standard, truth-linked or not. So there is no objec-
tionable circularity here. Second, criticism and revision are not precluded.
We can criticize (at least some of) our belief-forming processes even with
beliefs created by those very processes. I will treat all these points more
fully in due course.
12. Objects of Evaluation

Granted that epistemology is an evaluative enterprise, its scope and di-
rection depend heavily on its objects of evaluation. Which things are to be
4 Introduction
evaluated? Epistemologistshave often held that the prime objects of evalu-

ation are arguments, or forms of inference. Now one thing meant by 'infer-
ence', or 'argument', is a set of sentences or propositions. If inferences in
this sense are the objects of epistemic evaluation, epistemology need only
enlist the aid of deductive and inductive logic. On this approach epistemol-
ogy need not evaluate psychological processes. Nor are social institutions,
or properties of intellectual communities, proper objects of evaluation.
Against this choice of evaluative objects, a number of points must be
made. First, although epistemology is interested in inference, it is not (pri-
marily) interested in inferences construed as argument forms. Rather, it is
interested in inferences as processes of belief formation or belief revision, as
sequences of psychological states. So psychological processes are certainly a
point of concern, even in the matter of inference. Furthermore, additional
psychological processes are of equal epistemic significance: processes of
perception, memory, problem solving, and the like.
Why is epistemology interested in these processes? One reason is its inter-
est in epistemic justification. The notion of justification is directed, princi-
pally, at beliefs. But evaluations of beliefs, I contend, derive from
evaluations of belief-forming processes. Which processes are suitable can-
not be certified by logic alone. Ultimately, justificational status depends (at
least in part) on properties of our basic cognitive equipment. Hence, episte-
mology needs to examine this equipment, to see whether it satisfies stan-
dards of justifiedness.
The architecture of cognition does not constitute the focus of all episte-
mology, not even all indiuiduul epistemology. But it does constitute 'the
focus of what I call primary (individual) epistemology. The notion of pri-
mary epistemology may be introduced by consideration of a passage from
Francis Bacon. In his Novum Organum, Bacon wrote:
Neither the naked hand nor the understanding left to itself can effect much.
It is by instruments and helps that the work is done, which are as much
wanted for the understanding as for the hand. And as the instruments of the
hand either give motion or guide it, so the instruments of the mind supply
either suggestions for the understanding or cautions.'
Bacon's point should be granted. Just as man's technological achieve-
ments have not been executed by the naked hand, most advances by the
human intellect have not been achieved by the understanding Zeft to itself.
In both cases instruments, tools, or "helps" have been essential. In the in-
tellectual domain such tools include systems of language and notation,
proof techniques in mathematics, methodologies of empirical science, and
all the other symbolic structures and heuristics by which science and schol-
arship have been guided. With this granted, it might seem as if the job of
normative epistemology is to appraise instruments proffered to the mind,
Introduction 5
and to devise new instruments. On this conception logic, probability theory,

and statistics fall within normative epistemology, since they offer important
scientificand intellectual tools. But studying the understanding considered
by itself-the fundamental architecture of human cognition-should be of
no significance to epistemology.
This conclusion, however, goes awry. While few intellectual feats are
achieved without any "helps" for the mind, the intrinsic properties of the
mind still hold significance for epistemology. First, unless the mind has a
suitable structure, it cannot use tools properly or successfully. Second, the
invention, acquisition, and selective retention of intellectual tools must ulti-
mately rest with native cognitive mechanisms, with the understanding
"considered by itself." It is fully appropriate, then, for epistemology to in-
quire into cognitive architecture, to assess its strengths and weaknesses. At a
minimum, this is a job for primary epistemology.
Within primary epistemology, then, the objects of epistemic evaluation
are cognitive processes, structures, and mechanisms. But this only touches
individual epistemology. What are the objects of evaluation in social
epistemology?
Social epistemology is concerned with the truth-getting impact of differ-
ent patterns and arrangements of social intercourse. For example, there are
different possible forms, or styles, of interpersonal argument, debate, or
controversy. How would these affect the resultant beliefs of the partici-
pants? More generally, social intercourse involves a variety of communica-
tions structures and channels. Informal channels, involving face-to-face
conversations, are one end of a continuum. Other channels, such as class-
rooms, courtrooms, and assemblies, are also face-to-face, but more fotmal.
Finally, there are non-face-to-face channels involving the public media.
Given the informational or misinformational impact of communication
channels, the motivations and resources of people involved in channel
transactions are critical. What variables affect which messages a potential
communicator would like to transmit? What variables affect the decisions ,
of channel controllers--such as editors, media owners, or boards of educa-

tion-in their selection of candidate messages? What variables affect the
activities of potential message receivers in their search for messages of in-
terest and in their response to messages they receive?
Interwoven in such structures and acts of communication, various posi-
tions and patterns are found. There are positions of power and authority,
and patterns of cooperation and conflict. The task of social epistemology, as
I conceive it, is to evaluate the truth-conducive or truth-inhibiting proper-
ties of such relationships, patterns, and structures. What kinds of channels,
and controls over channels, comprise the best means to 'verific' ends? To
what degree should control be consensual, and to what degree a function of
(ascribed) expertise, or 'authority'? To what extent should diversity of mes-
6 Introduction
sages be cultivated? Which principles of selectivity are suitable for differ-

ent sorts of channels? What distinguishes the canons of science from other
systems of investigation and communication? These questions are the do-
main of social epistemology.
1.3. Continuities and Discontinuities with Historical Epistemology

How my conception of epistemology converges with, or diverges from, the
subject's history merits discussion. First, does the psychological orientation
mesh with epistemology's history? If we consider its full history, the psycho-
logical orientation fits well. The historical literature is replete with descrip-
tions and classifications of mental faculties and endowments, processes and
contents, acts and operations. Psychological theories that occupied histori-
cal epistemologists also had a motivation similar to mine: to assess the
mind's powers and limits in its quest for knowledge. Many writers empha-
sized the limits of knowledge, urging us to 'rein in' reason's excessive pre-
tensions. Kant is famous for this thesis, but the idea was expressed earlier by
Locke: "If by this inquiry into the nature of the understanding, I can dis-
cover the powers thereof; how far they reach; to what things they are in any
degree proportionate; and where they fail us, I suppose it may be of use to
prevail with the busy mind of man to be more cautious in meddling with
things exceeding its comprehension"? Other writers were more optimistic,
even Hume in certain moods: '"Tis impossible to tell what changes and im-
provements we might make in [the] sciences were we thoroughly ac-
quainted with the extent and force of human ~ n d e r s t a n d i n ~ . " ~
Psychologistic epistemology, then, is in the mainstream of historical
epistemology. My proposal differs in rejecting introspectionism in favor of
experimental psychology. But the relation envisaged between normative
epistemology and mental science closely parallels the classical tradition.
As is well known, however, epistemology of the twentieth century took
an antipsychologistic turn. Influenced by Gottlob Frege, positivism, and or-
dinary-language philosophy, 'psychologism' became suspect in all branches
of philosophy. Frege opposed psychologism in logic and the philosophy of
mathematics. Positivism viewed all legitimate philosophy as the 'logic' of
the sciences. Ordinary-language philosophy stressed the study of natural
language, and often (like positivism) favored a behaviorism that denied the
relevance (or existence!) of interior mental operations. Through much of
this period it was viewed as a confusion to introduce psychology into episte-
mology. Rudolf Carnap separated these sharply, as did Hans Reichenbach
with his distinction between justification and disc~very.'~ Discovery was
relegated to psychology, while epistemology was supposed to address ques-
tions of validity, an issue of logic rather than psychology. EpistemoIogy was
to elucidate the logic of science, or analyze key epistemic concepts. The
Introduction 7
paradigms of epistemology became the logic of confirmation, the analysis of

'S knows that p', and the theory of justification or warrant. None of these
tasks, it was felt, called for collaboration with empirical investigators.
I do not reject these inquiries. I shall myself sketch an analysis of knowing
(Chapter 3),and present a theory of justification (Chapters 4 and 5). What I
do reject is a radical bifurcation of epistemology and psychology. When
correct accounts are given of knowledge, justification, and other pertinent
epistemic concepts, it emerges that the empirical study of belief genesis be-
comes most relevant to epistemology So, at least, I shall argue. While I en-
dorse (a form of) psychologism in epistemology, I do not endorse it in logic.
Validity or invalidity of arguments, as far as I can see, is not a matter of psy-
chology. Truths of model theory, proof theory, and recursive function the-
ory-the main branches of logic--do not depend on psychological truths. I
am open to persuasion that psychology undergirds logic. But I am not in-
clined toward this view, and nothing said here depends on it.
In linking epistemology with cognitive science, I use the term 'cognitive
science' (or 'cognitive psychology') neutrally. Some use the phrase in an
'ideological' way, to advocate a particular brand or style of theorizing. For
example, it may designate a computational approach, or an approach that
focuses on a certain level of abstraction, say, functional as opposed to neu-
ral. In this usage cognitive science contrasts with neural and biological ap-
proaches to cognition. But my usage is not ideological in any of these ways.
It includes any scientific approach to cognition.
Though my general outlook is catholic in principle, I in fact draw mostly
on cognitive psychology in the narrow use of the term. Rather little is said
about neuropsychology. Although I pay attention to neurally inspired
models in Chapter 15, the present state of neural modeling does not permit
very much in the way of epistemological applications.
Little is said in this book about the tenability of this or that theoretical
approach in psychology or cognitive science generally, though the exam-
ples I use doubtless betray my preferences (in this epistemological context).
A detailed treatment of theoretical stances belongs to the philosophy of psy-
chology, which I do not pursue here. To be sure, the philosophy (methodol-
ogy) of psychology is part of the philosophy of science, which is in turn part
of epistemology. But the focus of this book is primary epistemology, which
does not address the special problems of particular sciences. (That would be
treated under secondary epistemology.) Since the choice of psychological
method depends on epistemology, and epistemology depends on psychol-
ogy, there is (in Quine's phrase) "reciprocal containment" between the dis-
ciplines." But nothing is objectionable in this relationship.
Let me turn now to continuities and discontinuities in my approach to
social epistemology. Although modern epistemology-from Descartes to
Kant-was preoccupied with the individual knower, there has been plenty
8 Introduction
of interest in its social dimensions. This interest goes back at least to Plato
and the Socratic elenchus. Probably influenced by the sophists of his day,
Plato treated inquiry as an interpersonal, dialectical process, in which two
inquirers try to sustain their opinions against the disputations of their oppo-
nent. This dialectic was regarded by Plato as an essential epistemic device.
In more recent times the epistemology of science has often emphasized
social elements. Pragmatists like C. S. Peirce and John Dewey stressed the
corporate aspect of scientific inquiry, and positivists also stressed the inter-
subjectivity, or publicity, of scientific knowledge. Thomas Kuhn is identi-
fied with a different social perspective on science, with his emphasis on the
paradigms that scientific communities share and transmit.12 ~istoriansand
sociologists of science are generally concerned with the communities, or
schools, that have marked scientific controversies.
A different social slant to epistemology was introduced by Wittgenstein,
who viewed epistemological criteria as portions of language games, under-
stood as species of social practices.'3 This Wittgensteinian slant underlies
Richard Rorty's recent formulation of a social conception of epistemology.
For Rorty, knowledge is "the social justification of belief," and rationality
and epistemic authority are supposed to be explained by reference to "what
society lets us say.""
Precedents for social epistemology are plainly legion, but most of the
current approaches strike me as unsatisfactory in various ways. Historians
and sociologists of science typically devote their energy to descriptions of
scientific practice and development. Little attention is given to the evalua-
tive issues that are characteristic of epistemology. When evaluative ques-
tions are broached, they are seldom tackled with a firm theoretical
foundation. I shall try to remedy some of these deficiencies in my subse-
quent work on social epistemology.
The territorial instinct may incline some philosophers to view the pro-
posed interdisciplinary conception of epistemology as a regrettable surren-
der of philosophical autonomy. But epistemology's lack of autonomy would
only parallel that of other branches of philosophy. Neither moral nor
social-political philosophy can proceed without recourse to other disci-
plines. How can moral philosophy abstract from contingent facts of the
human psyche: the sources of aggression and sympathy, the sense of fairness
and reciprocity? How can it abstract from our feelings about ourselves, our
offspring, and infrahuman creatures? How can social philosophy neglect the
feasibilities and infeasibilities imposed by economic, historical, and political
contingencies? Admittedly, foundational questions of ethical and social the-
ory may be addressable without appeal to particular social facts. But sub-
stantive conclusions of social theory must rest in part on human nature and
economic and political feasibilities. Therefore, substantive ethics and social
theory must be multidisciplinary affairs.
Introduction 9
The same holds for epistemology. Although foundational questions may

be autonomous, substantive individual and social epistemology need help
from other disciplines. The present book develops this theme for individual
epistemology. Part I addresses foundational issues and is squarely within
philosophy by anyone's conception. But a major conclusion of Part I is that
substantive evaluations of (primary) individual epistemology require inputs
from the psychology of cognition. I present illustrations of this idea in Part
11, but I must stress that these are only illustrations. I do not purport to pre-
sent a finished primary epistemology. My intention is to delineate the gen-
eral contours of primary epistemology, as I envisage it. With continued
growth in the sciences of cognition, we will be able to do the job better and
better. I am merely trying to illustrate the modus operadi.
The multidisciplinary conception of epistemology I like to call episte-
mics, to distinguish it from other, more autonomous, conceptions of the
field. It is not the only possible conception of the subject, but it is poten-
tially the most fertile. There are current interests and developments in the
study of knowledge processes and knowledge systems (using 'knowledge'
loosely) in many disciplines. Epistemics is a suitable way of integrating
those widely shared interests.
THEORETICAL FOUNDATIONS
chapter 1
The Elements of Epistemology
1.1. Beliefs, Assertions, and Propositions

Epistemology deals with affairs of the intellect. In common parlance 'intel-
lect' often refers to the remote reaches of the mind. This is not what I mean.
I mean the whole range of efforts to know and understand the world, in-
cluding the unrefined, workaday practices of the layman as well as the re-
fined, specialized methods of the scientist or scholar. It includes the entire
canvas of topics the mind can address: the nature of the cosmos, the mathe-
matics of set theory or tensors, the fabric of man-made symbols and culture,
and even the simple layout of objects in the immediate environment. The
ways that minds do or should deal with these topics, individually or in con-
cert, comprise the province of epistemology.
It is controversial whether different topics call for fundamentally differ-
ent intellectual approaches or processes. But certain elements are common
to virtually all efforts of the intellect; and these form the basic building
blocks of epistemology. Deployment of the intellect involves either mental
acts or states, or public utterances, frequently both. Among the mental
states, beliefs are usually singled out by epistemologists. Among linguistic ,
acts, the ones of central concern are assertions. Most intellectual endeavors
try to arrive at some belief on a designated topic, or to formulate a state-
ment on the problem at hand. Accordingly, the 'product' of the scientist or
scholar is typically a body of assertions-presumably accompanied by a
body of beliefs. So epistemology naturally focuses on either beliefs or asser-
tive claims. In this book I will largely abstract from natural language and
public speech acts, since there will be enough complexities on our hands
without the complexities these topics introduce. I will concentrate on the
mental side of epistemology, without implying that this is its only side.
Belief is not the mentalistic concept of choice for all epistemologists. Be-
lief is normally a categorical, or binary, concept, an all-or-nothing affair.
You either believe something or you don't. But epistemologists commonly
point out that there are degrees of conviction or confidence in a statement.
14 The Elements of Epistemology
They urge us to do epistemology in terms of degrees of belief, or perhaps

subjective probabilities. In describing someone's view of tomorrow's
weather, you needn't confine yourself to saying 'He believes it will rain',
'He believes it won't rain', or 'He's undecided'. You can describe him as
'believing-to-degree-.78'that it will rain, or something of this sort.
Whatever the exact concept we select-a categorical or graded notion of
belief-some sort of belief-notion seems critical for mentalistic epistemol-
ogy. Let us now place belief states within a larger map of the mental ter-
rain. This map is implicit in what philosophers call 'folk psychology', that is,
a network of assumptions and views about the mind that is expressed in
everyday language. Whether these folk psychological concepts succeed in
picking out real mental entities is a controversial question, to which I will
return shortly and in subsequent chapters (especially Chapter 8). But first
let us see where folk psychology, as presented by philosophers, places the
concept of belief.
Philosophers commonly divide mental states into two sorts: those that
have and those that lack propositional content. The former are proposi-
tional attitudes, and the latter sensations, qualia, or the like. Propositional
attitudes are recognized by the sentences used to ascribe them, the telltale
sign being an embedded 'that'-clause. Typical examples of such sentences
are: 'He wishes that the siren would stop', 'She doubts that it will snow this
weekend', and 'He fears that war will never end'. Embedded that-clauses
apparently pick out contents of the indicated attitudes, contents commonly
referred to as 'propositions'. Hence the term 'propositional attitudes'. Pro-
positional attitude types include wanting, hoping, wishing, fearing, doubt-
ing, suspecting, and believing. A particular mental state of this type consists
in a person having a specific attitude type directed at a specific proposition.
In addition to propositional attitudes, the class of mental states includes
sensations like pains, itchy feelings, and perceptual experiences, all of
which are said to have qualitative character. However, some theorists hold
that even these mental states have propositional content; and a few
theorists try to explain away qualitative character in terms of propositional
content.
Among propositional attitudes, we distinguish those with a conative or
optative attitude toward a proposition-favoring or opposing the proposi-
tion's realization-and those with a purely intellectual assessment. The lat-
ter involve a stance on the question of whether the proposition is true, quite
apart from whether it would be nice if it were true. Such intellectual atti-
tudes include believing, being certain, thinking it likely, doubting, and sus-
pending judgment. These sorts of states are called doxastic attitudes, or
sometimes credal attitudes. They are central to epistemology.
In calling beliefs or other doxastic attitudes mental states, philosophers
do not imply that they are phenomenological states. Beliefs do not have to
The Elements of Epistemology 15
occur in consciousness to count as mental. The belief that you reside at 3748
Hillview Road may be held for many years, though you only 'think about'
your address intermittently during this period.
Many issues about beliefs and belief ascriptions are controversial. Most of
the controversy concerns belief contents. Does a mental state really have a
determinate content? If so, what is the source of this content? Is content
determined exclusively by what's 'in the head', or by external factors as
well? If external factors are relevant, can we really say that belief states (in-
cluding their contents) are purely mental, that is, inner states of the individ-
ual considered in abstraction from his or her environment or causal
ancestry?
Another problem concerns the objects (or relata) of beliefs. The term
'propositional attitude' naturally suggests that the objects in question are
propositions. And this is a common way of talking, both among philosophers
and psychologists. But propositions are problematic entities. As classically
interpreted, they are logical or abstract entities, somewhat akin to Platonic
forms. Philosophers widely regard this sort of ontological status with suspi-
cion. There are other theories of propositions, but none is free from criti-
cism. In place of propositions some philosophers posit sentences of an inner
language, a lingua mentis, as the relata of beliefs. But this sort of posit is also
controversial.
These sticky issues about beliefs are relevant to my enterprise. They are
relevant, first, because it is hard to do any mentalistic epistemology without
using something like the belief construct. Second, they are relevant because
of the intended bearing of psychology on mentalistic epistemology. Beliefs
may not be a kind of entity with which psychology is capable of dealing. Or,
being a construct of folk psychology, scientific psychology may choose to
replace the belief construct entirely and work with different theoretical
resources.
While all these issues are relevant to my undertaking, they cannot be
fully resolved here. That would require a full-scale treatment of difficult
topics in the philosophy of mind and the philosophy of language, which go
beyond the scope of this book. Several of these issues will be addressed
squarely in Chapter 8, but even there they will not receive a full-dress dis-
cussion. Nonetheless, some remarks are also advisable at this juncture, to
indicate roughly where I stand on these issues.
For convenience, I will proceed on the assumption that the objects, or
relata, of beliefs are propositions. The utility of proposition-talk is that it
nicely unifies the treatment of mental attitudes, overt speech, and truth-
value ascriptions. Suppose Alex says 'The moon is round', and Kurt says
'Der Mond is rund'. It is natural to say not merely that both asserted a truth,
but that they asserted the same truth, namely, that the moon is round. But
this single truth cannot be their two sentences, since these were different.
The suggestion is that the two sentences expressed one proposition, which is
the truth in question. Suppose further that Oscar utters no such sentence,
but has a belief he could express as 'The moon is round'. Then it is tempting
to say that he believes the same truth that Alex and Kurt assert. How can
this be explained? Again, proposition-talk can do the job. If Oscar's belief
content is a proposition-the same one expressed by Alex's and Kurt's sen-
tences-then it is easy to identify the single truth with this very proposition.
Thus, by letting propositions be (1)the contents of verbal assertions, (2) the
contents of beliefs, and (3) the bearers of truth, we nicely unify a whole
range of discourse.
Given the utility of proposition-talk, I shall avail myself of this talk, but
only as a f q o n de parler. I regard propositions as a temporary theoretical
posit from which we should ultimately ascend to a better theory. Since such
a theory may be quite complex, and its details probably would not seriously
affect my project, it is an issue to which I will devote little attention.
The bearing of psychology on the belief construct, however, is more cen-
tral to my concerns. So let me briefly anticipate a fuller discussion of this
matter in later chapters. What stance should cognitive science adopt to-
ward beliefs? At least four approaches are possible: reduction, replacement,
neutrality, and refinement.
A reductionist says that beliefs are perfectly fine entities for a serious
cognitive science, only they need to be reduced to, or explained in terms of,
scientifically respectable elements. Since propositions, for example, are du-
biously respectable, beliefs need to be understood as relations to syntactic
entities, such as sentences in a language of thought.
An advocate of replacement holds that beliefs are not scientifically ten-
able posits. As part of a radically false theory, a degenerating research pro-
gram, they will ultimately be discredited and abandoned like phlogiston.
An advocate of neutrality holds that cognitive science can remain neutral
about beliefs. It can proceed perfectly well without them, but this does not
imply their overthrow or illegitimacy. Cognitive science may not be in a
position to ascribe (propositional) content, or to say how such content
should be ascribed. But it may allow that a suitable theory of interpreta-
tion-a good 'psychosemantics'-can ascribe content, or show how content
is standardly ascribed.
The refinement position I have in mind is not an alternative to the pre-
ceding approaches, mainly because it does not address the problem of con-
tent. Content is not the only dimension of interest in the belief construct.
Psychology should also be interested in the range of contentful states. Cog-
nitive science could endorse the notion of content but stress the need for
acknowledging a richer array of distinct content-bearing states. Many dif-
ferent states, it may hold, are all lumped under the folk-psychological label
of 'belief, and these need to be sorted out for the purposes of an adequate
science of the mind.
I favor a combination of the neutrality and refinement approaches. I do

not expect cognitive science to force the total abandonment of content
states. But I do expect it to foster a more fine-grained set of descriptive re-
sources. This richer set of resources should help epistemology; not by mak-
ing it easier, but by making it (psychologically) more realistic.
1.2. Truth-Values and Knowledge

Beliefs are commonly said to be true or false, and so are assertions. But,
strictly, it is not acts of assertion, or states of belief, that are true or false. It
is the contents of these acts or states. Since we are taking propositions to be
the contents of beliefs and assertions, they have the role of bearers of truth
or falsity. What is primarily true or false is a proposition. A belief qualifies
as true only derivatively: when its content is a true proposition.
Ignoring assertions, then, there are three basic categories that need to be
distinguished. First, there are psychological, or mental, states. Second, there
are propositions, which are the contents of (certain of) these states. Third,
there are truth-values, such as true, false, and perhaps indeterminate. (The
possible need for a third truth-value arises because some propositions may
be neither true nor false, for example, propositions with vacuous referring
terms, like the one expressed by 'The present king of France is bald'.)
It is essential to distinguish carefully between these categories. In gen-
eral, the mere fact that someone believes a proposition does not make it
true, or false. The proposition's truth-value is not determined by its being
the content of some belief state. (Of course, there are exceptions. If the
proposition is 'There are beliefs', then the mere fact that someone believes
it guarantees its truth.) Moreover, there are true propositions that nobody
believes. PresumabIy there is a true proposition saying precisely how many
people will have a toothache in 1993. But nobody believes that proposition
now, and probably nobody will ever believe it.
What makes a proposition have the truth-value it has? The natural an-
swer is: the way the world is. I believe this natural answer is right, and phil-
osophically defensible. Philosophers of this persuasion are often called
realists, so I am a realist (of sorts).
Although a full discussion of truth and realism is deferred until Chapter
7, some preliminary clarificatory remarks must be entered here. In discuss-
ing truth one must distinguish firmly between the question of what makes a
proposition have a certain truth-value and the question of how people can
determine its truth-value. The former question concerns the nature of truth.
The latter concerns the eoidence for a proposition's truth-value, or the
methodology of trying to figure out a proposition's truth-value. This dis-
tinction is elementary, but it still needs emphasis. Some philosophers hold,
to be sure, that a proposition's meaning (or truth-conditions) is a function of
the evidence that would count for or against it. Still, at a minimum, one
must distinguish between a proposition's being true and people hauing eui-
dence in its favor. Surely there are plenty of cases of true propositions for
which nobody now (or perhaps ever) has good evidence. The example of the
1993 toothache sufferers is a case in point.
Since many people, including philosophers, insist on conflating truth
with other notions-mostly related to evidence, justification, or the like-it
is worth dilating on this point at this early juncture. To take a recent exam-
ple, consider Richard Rorty's claim that there is a sense of 'true' in which it
means (roughly) "what you can defend against all comers."' The idea of
"what can be defended against all comers" is some sort of social justification
notion. According to Rorty, there is a sense of 'true' in which it is necessary
and sufficient for a proposition's truth that somebody can defend it, pre-
sumably successfully, against all who argue against it.
There is no such sense of 'true'. To appreciate what's wrong with this def-
inition, suppose you are an unfortunate victim of circumstance and misi-
dentification. A horrible crime has been committed of which you stand
accused. You are totally innocent-such is the truth. But you are a look-
alike of the dastardly criminal, and numerous witnesses come forward to
identify you as the doer of the deed. Sadly, you have no alibis. You were out
for a walk at the time of the act, and nobody can vouch for your where-
abouts. Meanwhile, the real criminal has died in an accident. Given these
facts, the real truth cannot be successfully defended against all comers. You
cannot defend it successfully, for there are too many eyewitnesses to make
your case believable. (Furthermore, we may suppose, you actually did have
a motive for the act, though it did not motivate you enough to do it.) Nor is
there anyone else who could successfully defend your innocence against all
comers. Nonetheless, the truth is: you are innocent. The only correct sense
of 'true' makes truth independent of how well it can be defended. Its de-
fensibility is a separate matter, which may depend on a variety of extrane-
ous circumstances. Any innocent person accused of a crime surely wants the
real truth to emerge; and the real truth is all that is normally meant by
'true'.
The innocent defendant case shows that "what can be defended against
all comers" is not a necessary condition of truth. But neither is it sufficient.
A totalitarian regime may arrange a successful defense of certain false prop-
ositions against all comers (at least all comers in that society); but this can-
not make the false true.
I will examine more sophisticated attempts to explain truth in terms of
evidential, or justificatory, notions in Chapter 7. At this point it is important
to distinguish truth and justification, whether social justification or intra-
personal justification.
It is equally imperative to distinguish truth from knowledge. A proposi-
tion may be true without being known, just as it may be true without being
believed. Let me pause here to make a few remarks about the concept of
knowledge. There is a loose and a strict sense of 'knowledge7. In a loose
sense a person knows something (a proposition) if he believes it and it is
true. If he believes the truth, he is not ignorant of it. He is cognizant of the
fact, and so, in a loose sense, knows it. Many writers, especially in the be-
havioral and social sciences, use 'knowledge' in an ultra-loose sense, to
mean simply belief, or representation of the world. This is probably a mis-
use, though one so common that perhaps it has achieved legitimacy. In
more proper usage, to which I shall cleave, no proposition can be known
unless it is true.
To repeat: in the loose sense of 'know' someone knows proposition p if
and only if he believes p, and p is true. This suffices to show that knowledge
is not equivalent to truth: truth does not require belief, but knowledge does.
There is also, however, a strict sense of 'know', which has much occupied
epistemologists. In this strict sense knowledge requires more than belief and
truth. It requires satisfaction of some third, and perhaps fourth, condition
beyond belief and truth. Since epistemology has often been regarded as 'the
theory of knowledge', many epistemologists have devoted great energy to
analyzing this strict sense of 'know'. I too shall pay it a fair bit of attention,
especially in Chapter 3. But the analysis of knowledge will not be one of my
sustained points of preoccupation. I will use it partly as a jumping-off point
for an examination of the notion of justification, which will be one central
focus of my discussion. For the present, it suffices to note the complexity of
knowledge and to distinguish it from truth.
A principal reason for resisting a realist construal of truth is an alleged
threat to knowledge. Many philosophers fear that if truth is definitionally
prized off from evidence, or justification, it will be impossible for anyone to
know the truth. Truth(s) will be epistemically inaccessible, or unknowable.
But this dire consequence is not at all indicated, at least not without lengthy
argumentation. The mere fact that extramental reality is what makes a
proposition true (or false), as realists maintain, does not imply that no truths
can be known. The mere fact that truth does not (definitionally) require
knowledge, or justification, does not mean that it precludes knowledge, or
justification. (These comments do not convey the most serious historical or
contemporary arguments against realism. I mention them only in an intro-
ductory fashion, to forestall very simple confusions that may befall those
unfamiliar with epistemology.)
I have already had occasion to introduce a central epistemic concept, the
concept of justification. This has been a doubly important term in the field.
First, it is important in its own right, as a term of epistemic evaluation, or
appraisal, which is widely applied. Second, it is often linked to knowledge,
which, as I have noted, is traditionally central to epistemology. Its connec-
tion to knowledge will be discussed later on, in Chapter 3. But I want to
turn now to justification in its own right; more precisely, to the larger fam-
ily of terms of epistemic evaluation of which 'justification' is one member.
1.3. Epistemic Evaluation

The Introduction stressed the fact that epistemology is an evaluative, nor-
mative, or critical discipline. Let me now address the possible scope and
nature of such evaluation. First, what do we mean by 'evaluation', or
'norm'? We mean a judgment that pronounces something to be good or bad,
right or wrong, proper or improper, and the like. Since we are concerned
exclusively with epistemological evaluation, we do not mean to raise ques-
tions of moral, aesthetic, or legal goodness or badness, rightness or wrong-
ness. We mean to ask about specifically intellectual virtues or vices,
proprieties or improprieties.
Notice that the term 'value' is not always tied to the above sense of evalu-
ation. When the mathematician or logician speaks of the values of a vari-
able, he is not talking about a variable's good or bad aspects. He merely
refers to the object or class of objects over which the variable ranges. Simi-
larly, when a mathematician evaluates a function, he does not judge it to be
good or bad; he merely calculates the output of the function for a specified
input. The term 'truth-value', I believe, also uses 'value' in a nonevaluative
sense. To assign a proposition a truth-value does not, as I understand it, in-
volve any evaluative judgment. It does not say that the proposition is good
or bad, intellectually proper or improper, or anything of the sort. It just says
(roughly) that the world is the way the proposition says it is.
But there are many terms and phrases that do register epistemic evalua-
tions. Some of these were mentioned in the Introduction and in the preced-
ing section. But let me discuss these more thoroughly.
On a traffic citation form in Cook County, Illinois, the following clause
appears: 'This complainant . . . states that he has just and reasonable
grounds to believe and does believe that such defendant committed the
above offense'. The phrase 'just and reasonable grounds' is an example of an
epistemic evaluation. An officerissuing such a citation affirms not only that
he believes that the defendant committed a specified offense, but that this
belief is justly and reasonably grounded, and in this sense, is epistemically
proper or in order. More generally, terms like 'justified' and 'unjustified',
'warranted' and 'unwarranted', 'well-grounded' and 'ill-grounded', 'reason-
able' and 'unreasonable', 'rational' and 'irrational', when applied to beliefs
or other doxastic attitudes, typically express epistemic evaluations. They
are used to grade a belief along some evaluative epistemic dimension.
In studying epistemic evaluation, a number of separate categories should
be addressed. Four categories may be usefully distinguished: (1)the domain
of evaluation, (2) terms or dimensions of evaluation, (3)standards of evalua-
tion, and (4) styles of evaluation.
Z h Elements of Epistemology 21
( 1 ) Domain of ewalwltion. Which 'things' are evaluated in epistemic eval-

uations? Prime examples, we have seen, are beliefs and other doxastic atti-
tudes. But these are not the only objects of evaluation. Methods are also
objects of epistemic evaluation. Philosophers of science often debate
whether certain inductive methods are scientifically proper or improper.
The term 'methods', however, is ambiguous. It can refer to purely abstract,
'logical' objects like argument forms (such as modus ponens and universal
instantiation) and principles of confirmation (such as the special-conse-
quence condition and the converse-consequence condition). But it might
equally refer to psychological processes, especially belief-causing, or belief-
generating, processes. Important differences in epistemological directions
might flow from different conceptions of methods. It is important to settle,
therefore, which of these conceptions is most central for this or that episte-
mological purpose. I will turn to this issue in later chapters.
Even among psychological processes, doxastic attitude-formingprocesses
are not the only candidates for objects of evaluation. Other candidates in-
clude hypothesis-forming processes, concept-forming processes, search pro-
cesses, and even second-order processes, for example, processes for forming
new belief-forming processes.
In addition to mental states, processes, and methods, there are other
types of candidates for epistemic appraisal. Certainly speech acts, both oral
and written, should not be exempt. Nor should it be assumed that a verbal
claim, or assertion, must be evaluated on the same basis as a belief. Eviden-
tial standards for a public claim may differ from those for a private belief.
Similarly, constraints on public argumentation may differ from constraints
on intramental inference. Furthermore, single assertions or arguments are
not the only plausible verbal units deserving evaluation. Larger discourses
like speeches, articles, and books are also natural units; in fact, they are
more prevalent units of intellectual evaluation than single assertions or
arguments.
Finally, a variety of institutional arrangements with intellectual import, .
and many social structures and processes, may be objects of evaluation.
Whether these are objects of epistemic evaluation will depend on one's
conception of epistemology. But given my conception of social epistemol-
ogy, as sketched in the Introduction, it should be clear that I regard these as
potentially suitable topics of epistemic evaluation. Most issues in this area,
though, are postponed to a future work.
(2) Dimensions of eualuation. Most evaluations treated in academic
epistemology have been in the category of evidential evaluations (loosely
speaking). I have in mind such evaluations of beliefs as justified, warranted,
well-grounded, reasonable, rational, and the like. It should not be supposed
that these are all equivalent; in fact, I doubt their equivalence. One point of
this section is to highlight the variety of terms of epistemic appraisal, to
prepare for needed distinctions.
Many terms of widespread intellectual evaluation have not figured pro-

minently in analytic epistemology, although there is no good reason why
they should not. Consider the term 'intelligent'. In most walks of life, as
well as in educational and psychological settings, this is the most common
term of intellectual appraisal (a term fraught with difficulties and contro-
versies, to be sure). There is no reason why this should not figure promi-
nently in the mission of epistemic evaluation. Admittedly, precise and cul-
turally acceptable standards for intelligence are needed to make the term
useful, but that does not affect the present point. 'Intelligent' is especially
worth mentioning here because it does not fall (wholly) in the evidential
category. Intelligence does not seem to be just a matter of restricting one's
beliefs (or subjective probabilities) to the evidentially warranted, as I will
discuss in later chapters.
That intellectual evaluations are not exhausted by evidential ones should
be clear from various strands in the philosophy of science, and from social
practices of reward and approbation. Many philosophers of science, Karl
Popper most consistently: have stressed the value of strong, or bold,
theories. A theory that explains more is more desirable than a weaker the-
ory. A scientist who invents and substantiates a more powerful theory re-
ceives more approbation than one who invents, and provides equal
substantiation to, a weaker theory. In science, and in all avenues of the in-
tellect, importance or significance counts for a lot. Nobel prizes are not
awarded simply on the degree of evidence a scientist marshals for his
claims. The importance or significance of these results is a critical compo-
nent. So epistemology should not confine itself to treatment of evidential
dimensions of appraisal.
Still other dimensions of appraisal that figure prominently in intellectual
life are generally neglected in twentieth-century epistemology. For exam-
ple, discourses (and other cultural contributions) are regularly appraised on
the dimension of originality. Originality is not just a matter of evidential ap-
propriateness, argumentative soundness, or even importance or signifi-
cance. Rather, it is a matter (roughly) of how difficult it must have been to
make the contribution against the historical and cultural background of the
contributor, given the intellectual context that existed at the time. I will
explore the possible basis for such evaluations in Chapter 11.
(3) Standards of eualuation. In making an evaluation, a speaker assumes
some factual basis, or grounds, for the evaluation. In judging an athlete to
be outstanding, for example, a speaker presumably relies on some set of ac-
tual feats that the athlete bas performed. In appraising a student as supe-
rior, a speaker bases the judgment on some set of achievements, such as
percentage of correct answers to questions or problems. In general, evalua-
tive status does not enter the world autonomously. It always 'supervenes', as
philosophers sometimes put it, on purely factual states of afTairs. It is often
controversial just which factual states are, or should be, linked with specific
evaluations. This is notoriously problematic in the ethical domain. But few
writers, except perhaps emotivists, doubt that some factual basis (whether
objective or subjective) underlies evaluative judgments, or should underlie
them. A central problem for epistemology, equally, is the factual basis on
which epistemic evaluations are made. Since there are numerous different
terms of epistemic evaluation, it must be expected that different bases will
be paired with different terms. Some terms may be vague, or ambiguous, in
which case a family of different factual bases may be associated with a sin-
gle term.
An illustration from ethical theory should help concretize the idea of fac-
tual grounding for evaluative judgments. 'Right' is a core term of moral ap-
praisal. But what factual circumstances make an action right? A sample
answer is provided by Act Utilitarianism. This theory counts an action A
right if and only if A causes, or would cause, as much net happiness as any
alternative action open to the agent. The rightness of an action is deter-
mined, on this theory, by the relative amounts of happiness that actions in
the agent's alternative set would produce. This is a purely factual matter
(using 'factual' in a broad sense, to include counterfactual states of affairs).
There is a fact of the matter as to what quantities of happiness and unhap-
piness would result from sundry actions. So there is a fact of the matter-
however hard it may be to know this fact-as to whether action A would
produce as much net happiness as other actions in the alternative set. Thus,
Act Utilitarianism provides a factual basis on which rightness (allegedly)
supervenes. It is doubtful, of course, that this theory is correct. But it pro-
vides a clear example of a factual basis.
In the case of epistemic status, we also want to know the factual, or sub-
stantive, conditions on which it supervenes. Consider justifiedness. What
has to be true of a belief for it to qualify as justified? What factual standard
determines justifiedness?
What kinds of answers to this question are legitimate?It is not admissible ,
to answer by using other terms of epistemic appraisal, such as 'rational',

'well-grounded', or the like. We need nonevaluative terms or conditions.
Some principal categories of 'facts', or substantive conditions, are the fol-
lowing: (a) logical conditions, (b) probabilistic conditions, (c) psychological
conditions, (d) 'social' conditions, and (e) metaphysical conditions.
Let me briefly illustrate these categories. A theory might say that a per-
son's belief in proposition p is justified if and only if p is logically implied by
other propositions the person believes. Such a theory only invokes notions
in categories (a) and (c): logical relationships and psychological states. No-
tice that this theory makes justifiedness arise from logical implications
among propositions actually believed, not justifiably believed. A theory that
invoked justifiable belief would be inadmissible, for it would employ a term
of epistemic evaluation where substantive conditions are sought. (This

problem could be averted with a recursive, or inductive, specification of
justifiedness. But then at least the base clause or clauses of the specification
would have to omit terms of epistemic eva~uation.)~
Another admissible theory would let justifiedness arise from the corpus of
a cognizer's beliefs plus probabilistic relationships between the target belief
and the beliefs in this corpus (or rather, the propositional contents of these
beliefs). But here a theorist must tread carefully. The term 'probability' is
notoriously ambiguous, and some of its proposed explications implicitly ren-
der it a term of epistemic evaluation (tied to what an epistemically rational
person would do). For present purposes, 'probability' would have to be re-
stricted to some other meaning, for example, a frequency, or propensity
sense.
Most theories of justifiedness appeal heavily to psychological conditions,
especially the antecedent beliefs of the cognizer. A different sort of theory
would appeal to other people's beliefs, or perhaps their overt statements.
This would fall into category ( d ) :social conditions. The inclusion of socially
accepted practices, or 'language games', would also fall into category ( d ) .
These are admissible as long as it is merely the existence, or acceptance, of
the practices that is invoked.
Under category (e) I include a number of sorts of factual conditions: for
example, conditions invoking the truth or falsity of certain propositions,
conditions invoking the notion of cause, and conditions invoking nomologi-
cal necessity or any other modal notions.
Theories that invoke solely psychological conditions of the cognizer are
naturally called 'subjective', or 'internalist', theories. Theories that invoke
such matters as the actual truth or falsity of relevant propositions are natu-
rally called 'externalist' theories (assuming, at any rate, some realist ap-
proach to truth). The invocation of logical or probabilistic relations may be
harder to classify. These are not purely subjective, internal states of affairs,
but they have often appeared in theories that would normally be classified
as internalist rather than externalist. I shall explore these points later, espe-
cially in Chapters 4 and 5. This is only a preliminary introduction, to pre-
pare the reader for some of the directions the inquiry will take.
My interest in factual standards of epistemic terms sounds similar to an
interest in 'criteria of application', or 'indices' or 'measures' associated with
predicates (as these notions are used, for example, in the behavioral sci-
ences). But all of these terms have a verificationist connotation. They all
seek observational standards, which could directly be used to determine
whether the evaluative term in question applies in specific cases. By con-
trast, when I speak of factual standards, I do not restrict such standards to
observable, or verifiable, standards. I want standards that specify, in gen-
eral, when an evaluative epistemic term does apply, whether or not it pro-
vides a way of telling that it applies.
This contrast can be illustrated with the Act Utilitarian theory of moral
rightness. Such a theory provides substantive conditions of rightness, but
not a verification procedure for determining rightness. For example, the
theory invokes the notion of happiness without offering any observational
reduction of happiness. It presupposes truth-values for various counterfac-
tual statements, without specifying verification procedures for such coun-
terfactuals. These omissions are not objectionable for the purposes in
question. No such verificational, or observational, reductions are being
sought. Similarly, in offering factual standards for epistemic terms, it is ad-
missible to include truth-linked standards, despite the fact that 'true' is not
an observational, or directly verifiable, term. I stress this point because my
own approach will feature truth-linked standards prominently. So it needs
emphasis at the outset that this is quite unobjectionable, given the goals in
question.
(4) Styles of evaluation. Evaluative, or normative, discourse appears in
different styles. Some specimens of such discourse include deontic terms,
like 'obligatory', 'permitted', and 'forbidden', or more familiarly, 'right' and
'wrong'. Often such terms have systems of rules associated with them. Let
us call this kind of normative discourse deontic. There are other forms of
evaluative discourse, however. Some evaluative terms merely appraise cer-
tain qualities as good-making or bad-making features, relative to some suit-
able dimensions. In other words, they pick out certain traits as virtues or
vices. This contrast is found among terms of epistemic appraisal. Epistemo-
logists often try to identify rules of rationality, and they sometimes talk
about what is rationally required. This is deontic language in the epistemic
domain. By contrast, terms like 'careful' and 'original' might be used to
evaluate intellectual performances or traits, but such terms apply no deon-
tic category. They merely identify certain good-making characteristics.
Let me mark the foregoing contrast by saying that evaluative terms can
register either deontic classifications, on the one hand, or virtues and vices,
on the other. Both styles of evaluation will be employed in this book. I will
treat justifiedness (Chapters 4 and 5) as a deontic notion and examine it in a
framework of rules. But other evaluations to be studied will be in the vir-
tue-vice style.
A different division of styles distinguishes regulative and nonregulative
normative schemes. A regulative system of norms formulates rules to be
consciously adopted and followed, for example, precepts or recipes by
which an agent should guide his conduct. A nonregulative system of evalua-
tion, by contrast, formulates principles for appraising a performance or
trait, or assigning a normative status, but without providing instructions for
the agent to follow, or apply. They are only principles for an appraiser to
utilize in judging.' Many epistemologists have sought regulative precepts:
principles to guide a cognizer's reasoning processes.5 But it is also possible
to do epistemic evaluation-even in a rule-based framework-without
seeking action-guiding principles. Indeed, when it comes to fixed or auto-

matic psychological processes, it is pointless to offer principles of guidance.
Nonetheless, such processes can be the subject of epistemic appraisaL6
For the most part, my evaluative approach will be of the nonregulative
variety. This will be true especially when the issue revolves around basic
psychological processes, as it will in the heart of the book. This nonregula-
tive style will be adopted even in talking of justificational rules. However, a
regulative perspective is not entirely excluded, in some contexts of my
discussion.
1.4. Reliability, Power, and Speed

By way of preview, let me introduce three standards of evaluation-all
truth-linked standards. Each can be applied to a variety of objects in the in-
tellectual domain, including both psychological processes and social forms
and institutions.
The first standard is reliability. An object (a process, method, system, or
what have you) is reliable if and only if (1)it is a sort of thing that tends to
produce beliefs, and (2) the proportion of true beliefs among the beliefs it
produces meets some threshold, or criterion, value. Reliability, then, con-
sists in a tendency to produce a high truth ratio of beliefs. This, of course, is
a categorical notion of reliability. But a corresponding quantitative notion
of reliability is easily introduced, proportional to the true-belief ratio. (No-
tice that 'reliability', as used here and in other epistemology literature, cor-
responds to the behavioral scientists' term 'validity'.)
The reliability standard will be invoked in connection with the evalua-
tive notion of justifiedness. More precisely, reliability is one component in a
complex standard appropriate to justification. But reliability is of interest
independently of epistemic justification. It is of interest in the epistemic
evaluation of social arrangements that produce beliefs, even if these ar-
rangements are not judged on the dimension of justifiedness.
To be clear about reliability, it helps to distinguish two different intellec-
tual misadventures: error and ignorance. Error is false belief; ignorance is
the absence of true belief. A reliable process, method, or procedure is an an-
tidote to error. The greater the reliability of one's methods, the smaller
one's proportion of errors. But reliability is no antidote to ignorance. Com-
plete reliability can be achieved by extreme caution, or conservatism: pro-
ducing beliefs only in the 'safest' circumstances, where it is virtually
impossible to go wrong. But such radical caution would probably be pur-
chased at the price of (extensive) ignorance. If hardly any beliefs are pro-
duced, hardly any true beliefs are produced. This spells (extensive)
ignorance.
If the antidote to error is reliability, the antidote to ignorance is (intel-
lectual) power. Power is the capacity of a process, method, system, or what

have you to produce a large number of true beliefs; or, slightly differently,
the capacity to produce true beliefs in answer to a high ratio of questions
one wants to answer or problems one wants to solve. As these alternative
formulations suggest, there are several variants on the notion of intellectual
power. I will focus primarily on the second of these variants, the problem-
solving, or question-answering, conception of intellectual power (see Chap-
ter 6). Although power, like reliability, incorporates the ingredient of true
belief, it is clearly a distinct standard. A method or system can be very reli-
able without being very powerful; and a method or system can be pretty
powerful but not terribly reliable.
I believe that the power standard, especially the problem-solving variant,
is associated with the evaluational term 'intelligent'. One mark of intelli-
gence is a relatively high problem-solving capacity. As indicated earlier,
'intelligence' is an evaluative term that could be central to epistemology.
So it is appropriate to explore the substantive standards linked to it, such as
problem-soIving power. However, I will not attempt a detailed elucidation
of intelligence. That term has too long and controversial a history, and may
be too vague to make analysis profitable.
A third standard I will explore is speed, that is, speed in getting true be-
liefs. In many intellectual tasks, getting a true answer sooner rather than
later is a desideratum. Often there is a specified deadline, after which infor-
mation loses value. So problem-solving speed is a virtue of a cognitive sys-
tem, It is also linked to intelligence. In evaluating the strengths and
weaknesses of the human cognitive system-a task that will occupy me in
Part II-speed is a standard to be considered.
In addition to first-order reliability, power, and speed, I will have occa-
sion to consider second-order reliability, power, and speed. First-order relia-
bility is a property of belief-forming processes (methods, systems, or what
have you). But there are also second-order processes, processes that produce
or modify belief-forming processes or methods. A second-order process may .
be called second-order reliable if the processes it tends to produce are reli-
able, or, alternatively, if the modifications it introduces tend to increase re-
liability. Analogous notions of second-order power and second-order speed
are readily introduced.
I will not attempt to analyze, or specify standards for, all terms of episte-
mic evaluation. The salient omission here is rationality, which has figured
prominently in epistemology. This notion is so vague in ordinary usage, and
so disparately employed by different philosphers and social scientists, that it
has limited usefulness. In portions of the book (especially Chapter 14) I will
address and criticize other people's discussions of rationality, but I will not
try to provide a positive account of my own.
chapter 2
Skepticism
2.1. Varieties and Grounds of Skepticism

I have claimed continuity between my project and historical epistemology,
but some readers may wonder. A chief preoccupation of traditional episte-
mology is the problem of skepticism. Does my conception of the field-
epistemics-come to grips with skepticism? Can psychological investiga-
tions of cognitive processes bear on skeptical puzzles? And what about the
traditional interpenetration of epistemology, metaphysics, and the philoso-
phy of language? The outlines of my conception seem to place little weight
on these connections. Doesn't that belie my claim of continuity? This chap-
ter is devoted to this constellation of issues.
Let me begin with an overview of skepticism, not a historical overview,
but a systematic classification of types of skepticism, grounds for skepticism,
and possible responses to skepticism. This overview will help locate my po-
sition vis-A-vis epistemology's concern with skepticism and some of its other
traditional affiliations.
Skepticism comes in many shapes and colors.' These varieties can be ca-
tegorized by focusing on three dimensions: (a) theme, (b) scope, and (c)
strength. Let us look first at alternative skeptical themes.
Skepticism goes back, of course, to the ancient period. The two most im-
portant ancient schools were the Academics and the Pyrrhonians. The Aca-
demics, such as Arcesilaus and Carneades, maintained that no assertions
about what is going on beyond our immediate experience are certain. The
Pyrrhonians did not endorse the negative dogmatic conclusions of the Aca-
demics, in that they did not deny that knowledge of the nonevident was
possible. Instead, they suspended judgment on the question. The Pyrrhon-
ians attributed the origins of their views to Pyrrho of Elis, who was re-
garded as a model of the skeptical way of life. He tried to avoid committing
himself to doctrines about the nature of reality, while living according to
appearances and attempting to attain peace of mind. The fullest presenta-
tion of Pyrrhonism was given by Sextus ~ m ~ i r i c u s In
. ' response to oppo-
Skepticism 29
nents who portrayed the attitude of skeptics as a definite view, Sextus said
that it was like a purge that eliminates everything, including itself.
Pyrrhonian skepticism, then, might be characterized as the cultivation of
suspension of judgment, or doubt. This kind of skeptical theme does not
make pronouncements on the possibility of knowledge, certainty, or justifi-
cation. It only urges suspension of belief-though it is accompanied by ar-
guments to support this attitude. This form of skepticism contrasts with
other forms, which are constituted by theses about knowledge, certainty, or
other epistemic attainments. (In other historical periods, the term 'skeptic'
was often applied to anyone who rejected, or failed to endorse, widely ac-
cepted views.)
In contemporary epistemology, as well as in much of its history, skepti-
cism is primarily some sort of negative thesis about epistemic attainments. I
will focus on two of these attainments: knowledge and iustificatwn (or war-
rant). Skepticism has often been directed against certainty, but we can in-
terpret this formulation as a special case of either skepticism about
knowledge or skepticism about justification. Certainty can be viewed as
either the highest grade of knowledge or the highest grade of justification.
So skepticism directed against certainty falls either within the knowledge
theme or the justification theme.
Skepticisms differ not only in theme, but in scope. Some skepticisms are
universal, denying all knowledge or justification. This extreme sort of skep-
ticism is rather rare. Even the ancient skeptics tended to exempt proposi-
tions about appearances from the scope of their doubts. Most skepticisms,
then, are somewhat less comprehensive, restricting themselves to selected
domains such as ethics, mathematics, other minds, the past, the future, the
unobserved, and the external world. More global skepticisms tend to focus
on a selected theme, such as knowledge or justification, and try to show
what makes it impossible, or difficult, to achieve in any domain. More local
skepticisms tend to focus on the peculiarity of the selected subject mat-
ter-for example, the external world, or the future-and try to show what
makes it inaccessible to knowledge or justified belief.
Skepticisms vary not only in theme and scope, but in strength. Roughly, a
skepticism's strength is inversely related to the strength of the epistemic
achievement that it disputes. A common form of skepticism, we have seen,
is the denial of certainty. Certainty can be thought of as either the highest
grade of knowledge or the highest grade of justifiedness. Certainty-denying
skepticism disputes the existence of such high-grade epistemic attainments.
But this is a relatively weak sort of skepticism. Significant forms of knowl-
edge or justification may be weaker than certainty. We might be moder-
ately justified in believing something even if not completely justified. Such
lower grades of epistemic accomplishment are not disputed by this kind of
skepticism. A stronger form of skepticism, however, would deny even lower
30 Skepticism
grades of knowledge or justification. It might hold that we are not the

slightest bit justified in believing propositions in the selected domain(s).
A subtopic under strength is modality. Skepticism might be formulated as
a doctrine denying the actual existence of knowledge, or justified belief.
More commonly, however, skeptics dispute the possibility of the selected
epistemic attainment. What is sometimes left unclarified, though, is the na-
ture of this possibility. Is the skeptic denying the 'logical' possibility of
knowledge or justified belief? Is he saying that no being could have such
knowledge or warranted belief? Or is he just saying that it is humanly im-
possible to attain such knowledge or justification?
Some of the most common brands of skepticism address what is attain-
able by humans. Many of these skepticisms are explicitly founded on al-
leged shortcomings of human faculties. Here we can immediately discern
how the psychology of human cognition might be relevant to skepticism.
Where skepticism is predicated on the frailties or inadequacies of human
cognitive powers, the psychology of cognition may address this claim. Of
course, the psychology of cognition might be the foundation for new brands
of skepticism, by unearthing previously undetected human frailties. But
whether it ultimately promotes or impedes the cause of skepticism, psy-
chology can certainly be relevant. This point will be explored further
below.
Another subtopic under strength emerges when one distinguishes direct
and iteratiue skepticism.3 A sample of direct skepticism is the thesis that we
cannot know anything. A sample of iterative skepticism is the thesis that we
cannot know that we know. Although iterative skepticism is an interestingly
weaker variant of skepticism, I will focus my attention on direct brands of
skepticism.
Given this classification scheme for varieties of skepticism, I can now
proceed to survey different kinds of arguments, or grounds, for skepticism.
The single most pervasive ground for skepticism, especially when directed
at the knowledge theme, is the possibility of error. Many famous skepticisms
arise from this source. We can motivate this skeptical strategy by noting
that a person does not know a proposition, in any strict sense, if for some
suitable sense of 'could', he or she could be mistaken. Even if the target
proposition is in fact true, a person's belief in the proposition is not knowl-
edge if it could have been wrong. Suppose the belief was arrived at unsys-
tematically, by a wild guess or by wishful thinking. Then it is just
accidental, just lucky, that the belief is true, and it is not something really
known (in any strict sense of 'know'). Saying that it is just accidental that
the belief is true means that it could easily have been false, and,this error
possibility excludes possession of knowledge. Knowledge requires more
than true belief obtained by lucky guess. Exactly which error possibilities
must be foreclosed is controversial. But in some sense, knowing must rule
Skepticism 31
out the threat of error. So the skeptic's denial of knowledge often takes the
form of insisting that the threat of error is serious.
Three kinds of error possibility are prominent in the skeptical literature:
(1) the fallibility of our cognitive faculties, (2) the relation between the
mind and the objects of cognition, and (3)the logical relation between hy-
pothesis and evidence.
Fallibility of the faculties is the earliest and perhaps most common
ground for skepticism. Foremost is the fallibility of the senses, but the falli-
bility of reuson runs a close second. Deception by the senses was intensively
discussed by Sextus Empiricus, and most later skeptics including Descartes
employed similar ideas. The deceptiveness of reason was commonly demon-
strated by showing how it yields incompatible conclusions. In the seven-
teenth century Pierre Bayle argued that most theories are "big with
contradiction and absurdity," that man's efforts to comprehend the world
rationally always end in perplexities and insoluble diffic~lties.~ Similarly,
Kant undermined the pretensions of Reason with his antinomies, a classical
skeptical device intended to show the equal strength of contrary lines of
reasoning.'
Problems about the relation between mind and cognized objects also cut
across many periods and topics. Some of these dwell on the character of the
objects in question, and some on the nature of the relation. In the former
category there is the problem of mutability or change. Plato, following
Heraclitus, held that physical objects are constantly in flux, and felt that
such flux precludes knowledge. Only unchangeable, immutable objects such
as mathematical entities and Forms can be known. In more recent philoso-
phy just the opposite conclusion has been drawn: only physical objects, not
abstract entities, can be brought within our ken. In both cases, however, the
trouble arises from the possibility of error. Change is a problem because the
real properties of a changing object are difficult to d e t e ~ tAbstractness
.~ is a
problem because it is hard to see how the mind can get reliable information
about an abstract entity, which inhabits a nonspatiotemporal realm. '
The difficulty of knowing about an abstract entity may be assigned to the

mind-object relation, rather than to the composition of the object itself. The
problem is: how can the mind get access to an abstract entity? How can it
reliably form opinions about the entity's properties? A similar worry under-
lies skepticism about the external world, skepticism about other minds, and
other famous skepticisms. If the mind has direct acquaintance only with its
own contents, how can beliefs be reliably formed about physical objects? If
mental phenomena are intrinsically private, how can you form reliable be-
liefs about another person's mind? Whatever external behavior the person
displays, there is always the possibility of erroneous diagnosis of the under-
lying mental causes. Indeed, the assumption of any mental life at all may be
erroneous. The 'person' may be just an elaborately programmed robot. How
32 Skepticism
can mere behavioral access to a humanoid creature preclude this possibility

of error?
The third source of error possibilities is evidential gaps. If evidence fails
to entail a conclusion, alternative hypotheses are compatible with the evi-
dence. So anyone drawing the specified conclusion might be wrong. No evi-
dence available to us, Hume indicated, guarantees that the sun will rise
tomorrow, or that bread will continue to nourish.' None of our evidence,
Bertrand Russell pointed out, rules out the possibility that the world came
into existence just five minutes ago, complete with ostensible memories and
records of a wholly unreal past.8 These evidential relationships can gen-
erate, respectively, skepticism about the future and skepticism about the
past.
It would be a mistake to try to squeeze all arguments for skepticism into
the error possibility category. At least two other arguments may deserve
different treatment. One such argument is that any attempt to achieve jus-
tification incurs the dilemma of either circularity or infinite regress. One
variant of this argument is of ancient vintage, involving an attempt to find a
'criterion' of justification. A belief is justified, we may say, only if it is ap-
proved by a suitable criterion. But how is such a criterion to be chosen? The
Pyrrhonian skeptic claimed that any attempt to identify a criterion involves
either circularity or infinite regress. This dilemma was posed by Sextus
Empiricus:
In order to decide the dispute which has arisen about the criterion, we must
possess an accepted criterion by which we shall be able to judge the dispute;
and in order to possess an accepted criterion, the dispute about the criterion
must first be decided. And when the argument thus reduces itself to a form of
circular reasoning (diallelus),the discovery of the criterion becomes imprac-
ticable, since we do not allow them to adopt a criterion by assumption, while
if they offer to judge the criterion by a criterion we force them to a regress
ad infiniturn. And furthermore, since demonstration requires a demonstrated
criterion, while the criterion requires an approved demonstration, they are
forced into circular reasoning?
The circularity argument appears in a number of epistemological con-
texts; the problem of induction is perhaps the most famous. Is the inductive
method (assuming there is a unique such method) legitimate? Attempts to
validate induction all too readily appeal to induction itself. Yet this, the
skeptic objects, is circular. The proper conclusion, he says, is that induction
is illegitimate. Since there is no other method by which beliefs about the
future (or the unobserved) can justifiably be formed, the circularity problem
appears to doom the possibility of justified beliefs in this domain.
Another skeptical argument that might not go in the error possibility file
is the 'rival hypothesis' argument. Some of the most famous skeptical argu-
ments introduce hypotheses intended to rival ordinary, external world be-
Skepticism 33
liefs. These rivals include Descartes's dreaming and evil genius hypotheses
as well as their modem, brain-in-a-vat counterpart. These rival hypotheses
are all intended to raise doubts about our warrant in believing standard, ex-
ternal world propositions. These ploys might be construed as demonstrating
the possibility of error. But, alternatively, another style of argumentation
can be attributed to this sort of skeptic.
The skeptic might be arguing that a person is justified in believing p only
if there is no other hypothesis q which explains the phenomena as well as p.
The evil genius and brain-in-a-vat hypotheses might be adduced to show
that for any selected external world proposition p, a rival hypothesis can al-
ways be constructed that is equally explanatory. This kind of strategy is
often used in the philosophy of science. If, for any scientific theory T, a
nonequivalent but equally explanatory theory can be constructed, perhaps
we are never justified in believing any scientific theory.''
I suspect that the rival hypothesis stratagem is more commonly moti-
vated, at least in historical writings, by the possibility-of-error worry. Cer-
tainly Descartes, in the first Meditation, seems to have been thinking along
these lines. The other construal of the rival hypothesis stratagem is men-
tioned for the sake of completeness. As the discussion proceeds, however, I
shall pay much greater attention to the error possibility category.
2.2. Responses to Skepticism: Metaphysics, Analysis, and Psychology

There are five kinds of epistemological response to skepticism that are note-
worthy: (1)concession, either total or partial; (2) metaphysical reconstruc-
tion; (3) conceptual analysis; (4) defense of cognitive powers; and (5)
construction of a theory of evidence.
The first category of response is to concede all or part of the skeptic's
claims. At one extreme is concession to global skepticism, even strong
global skepticism, which claims that we know little or nothing, or that we
have not the slightest justification for believing anything. Toward the same
,
end of the continuum is concession to certain important local skepticisms,
such as skepticism about the external world. To concede any of these forms
of skepticism, I think, is to became a skeptic.
We have seen, though, that there are weak forms of skepticism. Indeed,
in some of these cases, it is debatable whether the view merits the label at
all. Suppose you admit that people are never completely justified in believ-
ing any empirical proposition. Or that people are never justified in believ-
ing any scientific theory with certainty. Does this make you a skeptic? Not
in current philosophical parlance. Lots of current philosphers hold these
views but do not style themselves skeptics. So agreement with very miEd
versions of skepticism does not constitute capitulation to skepticism as it is
usually understood.
34 Skepticism
The remaining categories of response to skepticism are all one or another

form of rebuttal. Their suitability, of course, depends on which argument
the skeptic marshals. Where the skeptic attacks the cognitive faculties, the
opponent can defend these faculties; either the very faculties attacked, or
different ones instead. (For example, if the skeptic says that knowledge is
unattainable through the senses, it may be countered that knowledge is at-
tainable through reason or innate faculties.) Where the skeptic claims that
objects of certain kinds are unknowable, the opponent can retort that the
objects do not have the uncognizable properties the skeptic alleges. Where
the skeptic claims that a rival hypothesis precludes justification, the oppo-
nent can construct a theory of explanation to show that the rival is not as
explanatory as the target hypothesis. And so on. Let us now examine these
responses in greater detail, to see exactly where they lead. Which strategies
of response the epistemologist selects will heavily influence the direction of
his epistemology.
We proceed to the second category of response on our list: metaphysical
reconstruction. Metaphysical theorizing has often been a parry to the thrust
of skepticism. When certain domains seemed epistemically inaccessible,
philosophers have sought to avert epistemic catastrophe by inventing meta-
physical theories to reconstitute the target domain, thereby making it more
accessible to the inquiring mind. When it appeared that the physical world
could not be 'reached' by the mind, Berkeley introduced idealism to avert
disaster," and Mill introduced phenomenalism.12By constituting trees and
tables out of 'ideas', or 'permanent possibilities of sensation', Berkeley and
Mill intended to save the knowability of these objects.
Similar stratagems have been employed in mathematics. Kant felt that
mathematics is knowable a priori and with certainty. But if the external
world is what makes mathematics true, this kind of knowledge seemed un-
attainable. Kant's solution was a metaphysical theory that interiorized
mathematical facts. Mathematical truths were said to be a function of the
mind's own structure, which imposes determinate forms on all experience.
This mental structure is knowable by transcendental means, and hence is
knowable a priori and with certainty.
Even in this century philosophers of mathematics have advanced ontolo-
gies designed to cope with skepticism. Mathematical platonism breeds the
puzzle of how the mind can know about abstract, nonspatiotemporal ob-
jects. Formalism, logicism, and constructivism are all (construable as) re-
sponses to this worry. If, as formalism claims, mathematical truths are
reducible to truths about marks, or inscriptions, they should be more readily
knowable, since marks are perceivable (or introspectible) entities. If, as lo-
gicism claims, mathematical truths are reducible to truths of logic, they are
more readily knowable, since logic (it was assumed) does not have so prob-
lematic an epistemic status. If, as constructivism claims, mathematical
Skepticism 35
truths are a matter of proofs that have been or can be constructed, then
these truths should be within human ken.
The general pattern here is plain. Metaphysical reductionism is attrac-
tive because the reducing entities are thought to be more epistemically ac-
cessible than the sorts of entities being reduced or eliminated. The
reduction is intended to avoid the skeptical conclusion that we cannot have
knowledge of the specified domain. The epistemic theme, however, is not
always knowledge or justification. It might feature other epistemic notions,
even a quasi-epistemic notion like reference. Hilary Putnam advocates a
metaphysical view he calls "internal realism," which seems principally de-
signed to avoid a skeptical threat of no determinate reference relation.13
The attractiveness of the metaphysical reductionist strategy depends on
two things: first, how plausible the advocated reduction is; and second,
whether the advocated reduction really facilitates knowledge of the target
domain. Most historical reductionisms have foundered on the first point.
The reduction of physical objects to ideas, or sense data, has not fared well;
nor did Kant's efforts to interiorize the domain of mathematics. In general,
reductive metaphysics has usually proved more costly in overall philosophi-
cal perplexity than the epistemological benefits it promised. So cost-benefit
analysis has turned the tide of philosophical opinion against it. Frequently,
moreover, the epistemological fruits have not proved as sweet as billed.
None of the twentieth-century movements in the philosophy of mathemat-
ics, for example, has been the epistemological panacea its proponents
proclaimed.
I would not make the sweeping claim that all metaphysical reductionism
is doomed to failure, that it is never advisable to use such a strategy to cope
with skepticism. But the track record of these strategies is not good, and my
own inclination is to steer in other directions. This is one reason why my
epistemological program does not venture deeply into metaphysics.
Let us turn next to the third category of response: conceptual (or linguis-
tic) analysis. Two different strains of this response must be distinguished: (a)
analysis of the sentences (or statements) in various special domains, and (b)
analysis of key epistemic terms, such as 'knowledge' and 'warrant'.
The first kind of conceptual analysis breeds many of the same kinds of
theory as metaphysical reductionism, only the theories assume a linguistic
form. Thus, linguistic phenomenalism of the twentieth century sought the
same style of refuge from skepticism about the external world as Berkeleyan
idealism and Millian phenomenalism. Similarly, logical behaviorism was a
linguistic reductionism, which (among other things) aimed to salvage the
knowledge of other minds. Another kind of conceptual reduction was
Hume's treatment of cause and effect. By tracing the idea of causation to
constant conjunction, Hume may have hoped to preserve the possibility of
knowledge of causation (though this interpretation does not sit well with
36 Skepticism
Hume's skeptical tendencies). At any rate, twentieth-century Humeans who

offerregularity analyses of causation do so partly because they hope to vin-
dicate the prospects for knowledge of causation.
My doubts about this use of conceptual analysis parallel my misgivings
about metaphysical reductionism, Most linguistic reductionisms of this sort
have been abortive. Linguistic phenomenalism proved no more successful
than its metaphysical forebears. And logical behaviorism was a similar flop.
Again, I do not oppose all uses of conceptual analysis of strain (a). Clearly
we do need analysis of causal statements, for example, before claims to
knowledge of causation can be made good. But I will not pursue ventures of
this sort in detail.
My feeling about the second kind of conceptual analysis is quite different.
Here I feel no reservations. Any adequate response to skepticism must in-
volve, or presuppose, analyses (or 'accounts') of key epistemic terms like
'knowledge' and 'justification'. Such analyses will be important parts of my
undertaking. Only the specific accounts offered of such epistemic terms will
rationalize the relevance of psychology to epistemology. (Actually, certain
details of my accounts are not crucial to the relevance of psychology. But if
correct accounts of these notions run along very different lines, psychology
is much less relevant, or entirely irrelevant.)
To illustrate the pertinence of analysis to the resolution of skepticism,
consider the mind's access to various classes of objects. The fear that physi-
cal objects may be inaccessible to the mind presupposes some sort of ac-
count of access. I will not try to say what accounts of access (explicit or
implicit) have led to skeptical worries. One way to allay these fears, though,
is to give an adequate account of access, which shows how the mind does, or
can, have access to the external world. Causal theories of knowledge try to
do just this. They suggest that a causal, or causal-subjunctive, relationship
between external objects and the mind, mediated by the perceptual sys-
tems, can be a sufficient condition for acquisition of knowledge about those
objects. Whereas the access metaphor conveys the impression that the mind
must somehow make its way 'to' the external world, the spirit of the causal
theory is that it suffices for the objects to 'transmit information' to the mind,
via energy propagation, sensory transduction, and the like. In short, the di-
rection of epistemic access is not from the mind to the object, but from the
object to the mind. If this is roughly right, there is no need to reconstruct
objects out of materials that are 'closer' to the mind, namely, the mind's
own contents. Knowledge is possible as long as it is possible for objects and
minds to stand in the right sort of causal relationship.
Of course, there are problems of detail with a purely causal theory of
knowledge.14 Nonetheless, the causal approach clearly illustrates how an
analysis of knowledge can, in principle, resolve skeptical worries without a
revisionary metaphysics.
Skepticism 37
Let me further illustrate how conceptual analysis can help answer skep-
ticism, and how different proposed answers guide epistemology down alter-
native paths of inquiry. One approach to justification says (roughly) that a
belief is justified if the believed proposition better explains what it purports
to explain than any rival proposition.15Furthermore, this 'best-explanation'
approach may be presented as a theory of evidence. That is, it may claim
that evidential support for a proposition is to be understood in terms of the
position of that proposition in an explanatory network. Two comments are
in order about this argument. First, if the account is right, it stands a good
chance of resolving certain skeptical challenges, specifically, those that
arise from rival skeptical hypotheses. To make this response stick, though,
the epistemologist must show that ordinary physical object beliefs are bet-
ter explanations of the phenomena than, say, the Cartesian demon hypoth-
esis or the brain-in-a-vat hypothesis. This will necessitate a satisfactory
theory of explanatoriness. (How does simplicity, for example, affect degree
of explanatoriness?How is simplicity measured? And so on.) In addition, the
epistemologist will be called upon to defend the view that a best-explana-
tion account of evidence is the right sort of theory of evidence. My second
comment, then, is that this sort of theory of justification not only bears on
skepticism but also bears on the direction of one's epistemological inquiries.
Obviously, this approach to justification points more in the direction of the
theory of evidence (the fifth response listed above) than in the direction of
psychological scrutiny of cognitive mechanisms.
By contrast, suppose an account of justification asserts that a belief is jus-
tified if it results from reliable cognitive processes. Such an account is rele-
vant to various forms of skepticism, specifically those forms of skepticism
that impugn our cognitive faculties. If this account of justification is correct,
then the skeptic may be right in insisting that our beliefs are justified only if
our cognitive faculties are not excessively error-prone. A full account of
justification would shed light on what is excessive or tolerable in the way of
fallibility. (Total infallibility is presumably a misplaced demand.) Finally, .
there would be the question of whether human cognitive processes are suf-
ficiently reliable. This question, in turn, breaks into two components: (a)
What cognitive processes do human beings in fact have? and (b) What is the
reliability of these processes? At least question (a) falls in the province of
psychology. Hence this account of justification would lead epistemology
down a path that fosters collaboration with cognitive psychology. And it
leads to a version of the fourth response to skepticism (the defense of cogni-
tive powers).
Conceptual analysis is sometimes deployed against skepticism in a much
more direct fashion. Some philosophers try to show that skepticism can be
defused by purely conceptual dissection of the skeptical position. I have in
mind here so-called transcendental arguments, at least semantic versions
38 Skepticism
thereof. Recent transcendental arguments have tried to show the absurd or

paradoxical nature of skeptical questions. They try to show that if the skep-
tic's claim makes sense at all, it must be false.
For example, Sydney Shoemaker has argued against the other-minds
skeptic as follows.16To assert, as the skeptic does, that it is logically impos-
sible for one person to know of another that he's in pain is to imply that the
word 'pain' has no established meaning. But if the word 'pain' has no estab-
lished meaning, then the putative statement that it is logically impossible
for one person to know of another that he's in pain has no established
meaning either. Therefore, either what the skeptic says has no established
meaning, or it is false.
Barry Stroud has critically examined transcendental arguments of this
sort and shown that they depend on a very dubious semantic principle,
which he calls "the verification principle."'7 I am inclined to agree with
this diagnosis. Attempts to rebut skeptical challenges by such a prio?.i means
rest on questionable semantic assumptions. So I do not see conceptual anal-
ysis as undermining skepticism quite so easily. Still, it is relevant to the pos-
sible defeat of skepticism in ways I have sketched above.
Since I have stressed the role of analysis in response to skepticism, let me
digress a bit to comment on conceptual analysis. Most philosophers these
days are cautious of this enterprise, principally in view of Quine's attacks on
analyticity and general problems about the notion of meaning.18
Three comments on this issue are in order. First, while there are doubt-
less severe theoretical problems concerning the notions of meaning and
synonymy, there must be some substance to the commonsense notions sug-
gested by these terms. Certainly we can distinguish better and worse defini-
tions of a given word, whether dictionary definitions or definitions offered
by casual speakers. Moreover, it is clear that the speech of a child or a
foreigner can often profit from a reasonable definition, and the better the
definition, the better the learner's resulting usage. So there must be some
phenomenon of meaning that remains to be clearly elucidated.
Second, although many philosophers preach the abandonment of analyt-
icity, their practice sometimes belies their preaching. People do things very
much like conceptual analysis even if they officially reject it. It is hard to do
much in epistemology (or other branches of philosophy) without feeling
constrained to do something like conceptual analysis.
Third, people often mean by 'conceptual analysis' the presentation of
strictly necessary and sufficient conditions for a term. But given the long-
standing failure of philosophers (or lexicographers) to devise accurate defi-
nitions of this kind for very many (if any) words, it is doubtful whether such
analyses are indeed possible. I share theorists' doubts about this form of def-
inition." In the meantime, however, 1 will follow the working practice of
trying to give necessary and sufficientconditions for important epistemolo-
Skepticism 39
gical terms. This can be regarded as a first approximation to a better way of

treating meaning, if such a better way can be devised.
2.3. How Important Is Skepticism?

It is clear that epistemologists cannot respond satisfactorily to skepticism
without accounts of key epistemic terms, especially 'knowledge' and 'justi-
fication'. This is one reason I shall devote the next three chapters to these
topics. However, the nature of knowledge and the nature of justification are
matters of independent interest, quite apart from the tenability of skepti-
cism. Thus, while I remain mindful of skepticism in these discussions, it does
not wholly dominate them. Skepticism is only one concern of epistemology;
its significance should not be blown out of proportion.
To help keep my treatment of skepticism in proper perspective, let me
amplify this point. The two principal questions of (global) skepticism are: is
it (humanly) possible to have knowledge, and is it (humanly) possible to
have justified belief? On the main accounts I shall offer of knowledge and
justified belief, the attainment of both knowledge and justified belief de-
pends critically on the use of sufficiently reliable cognitive processes.
Hence, whether knowledge and justified belief are humanly possible hinges
on the availability or nonavailability of sufficiently reliable processes. Sup-
pose that this matter-the availability or nonavailability of such pro-
cesses-is settled, whether in the affirmative or in the negative. Would it
follow that nothing is left for epistemology to do? Certainly not.
Suppose first that the threat of skepticism were thwarted. Suppose we
were guaranteed that there are sufficiently reliable processes to make
knowledge and justified belief possible. Still, much work would be left.
Epistemology should be concerned with not only whether there are reliable
processes, or methods, but which ones these are. Epistemology should be
interested in specifying the good procedures, not simply ensuring that there
are such. To be assured that knowledge or justified belief is possible is to be ,
assured (on this theory) that some satisfactory processes are available. But
how many such processes are there? And which ones are they?
Notice that even in the history of epistemology, the central issues have
not always been couched in terms of skepticism. Consider the rationalist-
empiricist controversy, for example. This was a debate not so much over
whether knowledge can be attained as over how it can be attained: by
principal reliance on the senses, or by greater reliance on reason? Rational-
ists questioned the ability of the senses and induction to deliver the goods;
they advocated the power of innate concepts and principles, intuition and
demonstrative proof. Empiricists challenged the existence of innate ideas
and a priori principles; they doubted the ability of pure ratiocination to
achieve intellectual ends. This debate continues today among neorational-
40 Skepticism
ists and their opponents. Both parties to the dispute can cheerfully admit
that knowledge is attainable (or perhaps that knowledge in a strong sense is
not attainable). The dispute centers on the means by which knowledge--or
true belief-is to be attained.
Now let us assume that global skepticism wins the day. Would that spell
the end of epistemology? Not necessarily. Even if there are not any suffi-
ciently reliable cognitive processes to qualify a person for either knowledge
or justified belief, there might still be differences among processes in de-
grees of reliability. A worthy epistemological task is to identify the cornpar-
atiuely reliable processes, to discriminate the better from the worse. This
would indicate how to maximize one's truth ratio, or maximize the proba-
bility of getting a high truth ratio. Even if epistemic honors like knowledge
or justified belief are beyond our grasp, that does not mean all distinctions
should be abandoned.
The next point to notice is that knowledge and justifiedness, at least on
my accounts, critically invoke one particular standard of appraisal: reliabil-
ity. But that is not the only standard of epistemological interest. Even if no
cognitive processes are sufficiently reliable to confer knowledge or justi-
fiedness, indeed, even if all are equally unreliable so that none stands out
above its peers, distinctions could still be drawn according to other stan-
dards. I have already mentioned the standards of power and speed. Proce-
dures might differ in the range of questions to which they can deliver
correct answers and in the speed of getting these answers. Such differences
are certainly of epistemic interest, even if the reliability of all processes is
too low to qualify for knowledge or justifiedness.
Another dimension of appraisal possibly orthogonal to skepticism is
rationality. This is a widely cited epistemic desideratum, one that may be
feasible for human beings even if both knowledge and justifiedness are not.
It may be possible to have rational beliefs even if knowledge is unattain-
able. This point is often taken for granted in the dominant philosophy of sci-
ence tradition, which commonly pursues questions of scientific rationality
without concern for standard problems of skepticism.
It should be noted, in this connection, that the weapons in the skeptical
arsenal are usually best equipped for demolishing the epistemic qualifica-
tions of full-fledged beliefs. When it is claimed that we cannot know, this is
often on the grounds that we do not have a right to be sure or completely
confident. It is more difficult for the skeptic to provide grounds for saying
that we have no right to even partial beliefs, say, to subjective probabilities
of .60. Theories of rationality often try to lay down conditions for rational
subjective probabilities. And such conditions may be important for episte-
mology even if prospects for knowledge are f o r e c ~ o s e dI. ~shall
~ not myself
offer any theory of rationality. But I shall tentatively tender an account of
Skepticism 41
justifiedness for partial beliefs. This would be of interest even if skepti-

cism-at least in the knowledge theme--wins the day.
The issue of skepticism, then, is not the be-all and end-all of epistemol-
ogy. A balanced epistemology should have a place for it without allowing it
to overshadow other issues. This book tries to present a balanced perspec-
tive. But since one of my aims is to redress the imbalance of previous epis-
temologies, which have exaggerated the significance of skepticism, the
space devoted to skepticism will be comparatively small.21
chapter 3
Knowledge
3.1. Knowledge and Causal Processes

What is knowledge? More specifically, what is propositional knowledge:
what is it to know that something is the case? To know a proposition p is to
know that it is true. But you cannot know that p is true unless it is true. So a
necessary condition for knowledge is truth. Equally, you cannot know that
p unless you are of the 'opinion' that p is true, unless you believe p. So be-
lief, like truth, is necessary. But true belief is not suffcient for knowledge, at
least not in the strict sense of 'know'. If it is just accidental that you are
right about p, then you do not know that p, even if you are correct in be-
lieving it.
Suppose you wake up in a foul mood one morning and think to yourself,
'Today is going to be a miserable day'; and lo, a miserable day ensues. It
does not follow that you knew in the morning that it was going to be a mis-
erable day. It is just a fluke if such a feeling is right, and flukes are not suffi-
cient for knowledge.
One traditional proposal for strengthening the requirements is to add a
justification requirement. A true belief is not knowledge unless it is a justi-
fied belief. I do not reject this suggestion, but it is a complex matter to es-
tablish conditions for justifiedness. So let us see what can be said about
conditions for knowledge without yet invoking a notion like justifiedness.
A first factor to consider about knowledge is a causal factor. Whether a
true belief is knowledge depends on why the belief is held, on the psycho-
logical processes that cause the belief or sustain it in the mind. This is illus-
trated by the miserable day example. In that example the only cause of the
belief is your foul mood, which produces a gloomy forecast. But this kind of
cause is not adequate for knowledge. If we change the cause, the same be-
lief could qualify as knowledge. Suppose you get a call from an officemate,
who reports on excellent authority that the boss plans to fire half the staff
today, yourself included. You believe as before-that it's going to be a mis-
erable day--but now you know it (assuming, at any rate, that things go as
Knowledge 43
reported). In this variation, the cause of your belief is suitable for

knowledge.'
The next question to ask is: What makes a cause, or causal process, the
right kind of process for producing knowledge? What distinguishes knowl-
edge-producing causes from other causes? Why isn't a feeling, or mood, an
appropriate kind of cause? The natural answer seems to be: because belief
formation based on mere feelings, or moods, can easily go wrong. It would
be easy to be in a bad mood in the morning although the day is not going to
be miserable. So if a belief gets formed in this fashion, it has a very good
chance of being false. The belief does not qualify as knowledge-even if it
happens to be true-because the style of belief production is error-prone, or
unreliable. If, however, the belief-producing process is reliable, that helps
qualify the belief for knowledge.
We have here a sample motivation for a reliability approach to knowl-
edge. Coupled with the earlier suggestion that the cause of the belief is
crucial, we have a causal reliability approach. But causal reliabilism is not
the only form of reliabilism. Let me briefly mention and discuss an alterna-
tive brand of reliabilism: the reliable-indicator approach.
D. M. Armstrong has proposed such an account of knowledge.2 The
model on which he bases this account--especially the account of nonin-
ferential knowledge-is the model of a reliable thermometer. In a reliable
thermometer the temperature reading is a reliable indicator of the actual
ambient temperature. Similarly, he proposes, a noninferential belief counts
as knowledge when it is a reliable indicator of the true state of affairs. Reli-
able indicatorship is explicated in terms of a nomological, or lawlike, con-
nection between the belief and the state of affairs that makes the belief
true.3 If the having of the belief is nomologically sufficient for the beliefs
being true, then the belief is a reliable indicator and qualifies as a piece of
knowledge.
Let B stand for a particular brain state, and suppose that whenever a
human being believes he is in brain state B, this nomologically implies that
he is in brain state B. This might happen because the only way to realize
belief in this proposition is to be in brain state B. It follows from the reli-
able-indicator account that whenever any person believes he is in brain
state B, and hence the content of that belief is true, this true belief qualifies
as knowledge. But this result is readily susceptible to counterexample. Sup-
pose that what causes a given person to believe he is in brain state B is not
any genuine understanding or information concerning brain states, in par-
ticular, no knowledge of the cited nomological fact. Further, suppose that
he has no autocerebroscope to monitor his brain states and give readouts
about them. He just has a firm hunch that he is in brain state B (it is a state
he heard described in a recent lecture, but he dozed through most of that
lecture and knows nothing of what was said about the state). Surely, he does
44 Knowledge
not know he is in brain state B, although his belief to this effect is a reliable
indicator of the truth.4
It is worth stressing that even a stronger variant of the reliable-indicator
condition would not suffice to capture knowledge. Even if the indicator re-
lation is strengthened from nomological to logical sufficiency, the desired
result is not secured. Even if the having of a given belief 'logically' guaran-
tees the truth of the belief, the belief may not qualify as knowledge.
Suppose young Humperdink considers a proposition having fifteen dis-
juncts, the eighth of which is 'I exist'. Schematically, then, the proposition
is: '1 or 2 or . . . or I exist or . . . or 15'. Now Humperdink is not logically
sophisticated enough to realize that this proposition is entailed by 'I exist',
so he does not believe it for this reason. Nor does he believe it because he
already knows (or believes) one of the other fourteen disjuncts to be true.
He only believes it because he is inclined to believe any complex proposi-
tion with ten or more disjuncts. The very thought of such a complex propo-
sition overtaxes his meager mental powers and induces belief. If these are
the facts, one would hardly say that Humperdink knows this fifteen-disjunct
proposition to be true. Nonetheless, his believing it is logically sufficient for
its truth. This is because his believing it logically guarantees that he exists,
which in turn logically guarantees that the disjunctive proposition is true.
Thus, his belief is a logically reliable indicator of the truth. Yet this does not
raise the belief to the status of knowledge.
Once again, the process that causes the belief is critical to the knowledge
question. Since the causal process in Humperdink's case is just being over-
whelmed by sheer complexity, and since this is an unreliable process, the
resulting belief fails to qualify as knowledge. The fact that the belief hap-
pens to be a reliable indicator of the truth is irrelevant to the knowledge
question.
3.2. Alternative Reliable-Process Theories

Given the attractions of causal reliabilism as an approach to knowledge, it is
worth pursuing further. In particular, let us focus on the attempt to account
for knowledge in terms of reliable processes. The reliable-process approach
was first formulated, in kernel form, by Frank ~ a m s e ~There
.' are now sev-
eral variants under discussion, and we need to explore the prospects of the
various alternatives.
Three pairs of options can help generate a field of alternative reliable-
process theories. (To repeat, these are theories of knowledge. There are also
reliable-process theories of iustifiedness, to which I shall turn in Chapter 5.)
First, we may distinguish a process's global reliability and local reliability.'
(This has nothing to do with the global-local distinction for types of skepti-
cism; that distinction does not enter here.) The difference lies in the range
Knowledge 45
of uses for which the process is reliable. Global reliability is reliability for
all (or many) uses of the process, not just its use in forming the belief in
question. Local reliability concerns only the reliability of the process in the
context of the belief under assessment. However, this might include its reli-
ability in certain counterfactual situations centered on the target belief.
This brings us to a second distinction among reliable-process approaches:
the cu:twl-counterfcu:tual distinction. Some approaches might invoke the
process's reliability only in actual applications. Other approaches might in-
voke its reliability in counterfactual situations as well.
If counterfactual applications are allcwed, a third distinction comes into
play. On the one hand, the admissible counterfactual situation may be re-
stricted to the one determined by the pure subjunctive: what would happen
if proposition p were false? Alternatively, more counterfactual situations
could be invoked, for example, all situations involving relevant alternatives
to the truth of p.
Combinations of options chosen from this menu determine an array of
reliable-process approaches. Let us see how current versions of reliabilism
fit into the resulting classification. Robert Nozick's version of reliabilism-
what he calls the "tracking" theory--can be seen as a choice of local relia-
bility, in counterfactual applications, as specified by the pure subjunctive.'
In first approximation his account features the following four conditions for
knowledge: (1)p is true, (2) person S believes that p, (3)if p weren't true, S
wouldn't believe that p, and (4) if p were true, S would believe that p. No-
zick adds the proviso that beliefs in the counterfactual situation(s) must re-
sult from the same method M, used in the actual situation.
Condition (3)is obviously the critical one in this account, and introduces
the local subjunctive as the linchpin of the analysis. Unfortunately, the local
subjunctive is too weak. Here are a few examples to show why.
Suppose a parent takes a child's temperature and the thermometer reads
98.6 degrees, leading the parent to believe that the child's temperature is
normal, which is true. Suppose also that the thermometer works properly,
so that if the child's temperature were not 98.6, it would not read 98.6 and
the parent would not believe that the temperature is normal. This satisfies
the first three conditions of Nozick's analysis. Presumably the fourth condi-
tion is also satisfied: in (close) counterfactual situations in which the child's
temperature is normal, the parent would believe that it is normal.
But now suppose there are many thermometers in the parent's medicine
cabinet, and all but the one actually selected are defective. All the others
would read 98.6 even if the child had a fever. Furthermore, the parent can-
not tell which thermometer is which; it was just luck that a good thermome-
ter was selected. Then we would not say that the parent knows that the
child's temperature is normal, even though Nozick's analysis is satisfied.
Another case will solidify the point. Suppose S sees a smiling stranger on
46 Knowledge
the street, and concludes that the man just won a lottery. His only basis for
this belief is that the stranger is smiling. S knows nothing else about him
that forms the basis for this belief. As it happens, this stranger would not be
smiling unless he had just won a lottery; the only thing with the power to
make him smile is a lottery prize. So if he had not won a lottery prize, the
cognizer would not believe that he had. According to Nozick's account, S
knows that this man just won a lottery. But intuitively that is not right.
Prior to Nozick, Fred Dretske offered a very similar analysis, although
the formulation of his analysis is slightly more complex.s However,
Dretske's analysis succumbs to the same sorts of counterexamples as I have
just presented.'
Another version of reliabilism was presented in my paper "Discrimina-
tion and Perceptual ~ n o w l e d ~ e . "That
' ~ paper also invokes counterfactual
situations in accounting for knowledge attributions; but it does not employ
the pure subjunctive. It says (roughly) that a true belief fails to be knowl-
edge if there are any relevant alternative situations in which the proposition
p would be false, but the process used would cause S to believe p anyway. If
there are such relevant alternatives, then the utilized process cannot dis-
criminate the truth of p from them; so S does not know.
This relevant alternatives account works better than the pure subjunctive
one. The pure subjunctive account is just too permissive. Suppose Sam spots
Judy across the street and correctly believes that it is Judy. If it were Judy's
twin sister, Trudy, he would mistake her for Judy. Does Sam know that it is
Judy? As long as there is a serious possibility that the person across the
street might have been Trudy rather than Judy (even if Sam does not realize
this), we would deny that Sam knows. This is handled properly by the rele-
vant alternatives account, since the possibility of its being Trudy would
presumably qualify as a relevant alternative. But the pure subjunctive ac-
count would not necessarily give the correct verdict. We may suppose that
what would be the case if Judy weren't across the street is that nobody
would be across the street; or we may suppose that if Judy weren't there,
her colleague in the mathematics department would be there. (Say they had
flipped a coin to see who would fetch lunch, and Judy lost. If the colleague
had lost, he would be the person walking down the street.) In these coun-
terfactual situations, Sam would not believe that Judy is across the street. So
Sam's belief would survive the pure subjunctive test, and that account
would incorrectly say that Sam knows.
Notice that the no-relevant-alternatives condition can handle Gettier and
post-Gettier cases." Consider Brian Skyrms's case in which a pyromaniac
believes that the next Sure-Fire match he strikes will light.'' It is true that it
will light, but not because of the usual chemical condition of previously en-
countered Sure-Fire matches. This particular match is defective, but it will
light because of a coincidental burst of Q-radiation. Here the pyromaniac
Knowledge 47
does not know that the match will light. Why not? He does not know be-
cause there is a relevant alternative situation in which the pyromaniac's
evidence is exactly the same but the match does not light, that is, a situation
in which the match is defective (as in the actual situation) but there is no
burst of Q-radiation. Other Gettier and post-Gettier cases can be handled
similarly, as I will show in section 3.4.13
In "Discrimination and Perceptual Knowledge" I emphasized what I
here call a local reliability requirement for knowledge. But I did not mean
to endorse local reliability to the exclusion of global reliability. Indeed, in
other papers I have made the global reliability requirement quite explicit.14
Notice that local and global reliability are not mutually exclusive. A theory
can require both, and I mean to require both in a theory of knowing. To
qualify as knowledge, a true belief must result from a generally reliable pro-
cess, not just one that is (counterfactually) reliable for the case in question,
Another variant of reliabilism, by contrast, selects global reliability to the
exclusion of local reliability. This is Colin McGinn's recent adaptation of
the discrimination approach.'5 McGinn's theory also contrasts with the two
kinds of theories discussed thus far in eschewing counterfactuals. In terms of
our classification, it selects global reliability; purely factual, not counter-
factual, situations; and the no-relevant-alternatives option (although this
third option does not invoke counterfactual alternatives).
The discrimination theme I stressed in "Discrimination and Perceptual
Knowledge" calls attention to the fact that one sense of 'know' listed in the
Oxford English Dictionary is "to distinguish (one thing) from (another)."
Examples of such a sense are "I know a hawk from a handsaw" (Hamlet),
"We'll teach him to know Turtles from Jayes" (Merry Wives of Windsor), 'I
don't know him from Adam', and 'He doesn't know right from left', as well
as other current sayings that readily come to mind. The conjecture is that
the propositional sense of 'know' is related to this underlying meaning, in
that knowing that p involves discriminating the truth of p from relevant al-
ternatives. On my development of this theme, the alternatives are counter-
factual alternatives. McGinn develops the discrimination theme slightly
differently. He wants us to consider the truth-discriminating power of the
belief-forming process (or 'method') within some relevant class R of propo-
sitions, where R is not restricted to proposition p and contraries of p. We are
to look at the propensity of the process to form true beliefs across a range
of propositions whose truth-values are taken as fixed in the actual w o r ~ d . ' ~
One of McGinn's examples that motivates his approach, especially his es-
chewal of counterfactuals, is this: You visit a hitherto unexplored country in
which the inhabitants have the custom of simulating being in pain. You do
not know that their pain behavior is mere pretense, and so you form the be-
lief, of each person you meet, that he or she is in pain. In this way you ac-
quire many false beliefs. One person, however, is an exception to the cus-
48 Knowledge
tom of pain pretense: this hapless individual is in constant pain and shows
it. You also believe of this person, N, that he is in pain. In this case your true
belief is not, intuitively, knowledge. But the relevant counterfactual 'If N
were not in pain, you would not believe that N was in pain' is true, since if
N were not in pain, then (unlike the pretenders), he would not behave as if
he was, and so you would not believe that he was. Since this counterfactual
is true, it cannot accommodate the absence of knowledge in this case. So a
noncounterfactual version of reliabilism, McGinn surmises, is needed to
handle it.
Notice, though, that this case is only a problem for the pure subjunctive
variant of the counterfactual approach. The relevant alternatives variant
can handle it. A relevant alternative possibility is one in which N is not in
pain but is a pain simulator, like the other people around him. In this coun-
terfactual situation you would falsely believe that he is in pain, which is
why you do not know in the actual situation.
McGinnYsdismissal of counterfactual conditions is unconvincing for other
reasons. First of all, he does not offer a fully formulated analysis. He does
not tell us exactly what the discriminative range of the belief-forming pro-
cess must be in order to be capable of knowledge production. Moreover,
there seems to be a general pitfall for his strategy. A process might be
wholly reliable in actual situations. This would apparently imply that every
true belief it produces counts as knowledge. But that seems implausible:
there might still be cases in which counterfactual possibilities defeat a
knowledge ascription. My worry on this score cannot be nailed down in the
absence of a fully formulated analysis from McGinn; but the outlines of his
approach incur this probable difficulty."
Although I differ from McGinn on the appropriateness of a counterfac-
tual requirement, especially a local requirement, we are in agreement on
the need for a global reliability requirement. This comes out clearly for
knowledge of necessities, such as the truths of mathematics. Can such
knowledge or nonknowledge be handled by invoking such counterfactuals
as 'if it were not the case that 7 + 5 = 12, S would not believe that 7 + 5 =
12'? One problem is that we have no adequate account of counterfactuals
with impossible antecedents. Another problem is that it is not clear that
such a counterfactual is true in the case of someone who does know that 7 +
5 = 12. So Nozick's third condition might be violated by someone who
knows this proposition. (The no-relevant-alternatives approach would not
be imperiled, though, because there may be no alternative in which 7 5 +
does not equal 12.) Also, while Nozick says that his fourth condition is de-
signed to handle true beliefs in necessities that do not qualify as knowledge,
McGinn shows that this claim does not succeed."
How should we handle true beliefs in necessities that do not qualify as
knowledge? My suspicion is that they are mostly cases in which the belief-
Knowledge 49
forming processes, or methods, are not globally, or generally, reliable. Gen-

eral reliability refers to the notion of reliability explained in Chapter 1,that
is, a statistical or dispositional property of a belief-forming process type. We
may take it to refer to the truth ratio of beliefs generated by the process on
all actual occasions of use, and probably some possible occasions as well. So
global reliability may also involve counterfactuals.
Leaving the topic of necessities, let us reflect more on general reliability.
As indicated, general reliability is probably best understood as a propensity
rather than a frequency. This avoids the possibility that actual uses of the
process are numerically too limited or skewed to represent an intuitively
appropriate ratio. If we move to the propensity idea, though, there is the
problem of specifying the range of possible uses that should be counten-
anced. I do not know exactly how to do this, but it seems plausible to re-
strict possible uses to situations rather similar to those of the real world. I
will return to this theme in Chapter 5.
Another problem for the global reliability approach is the 'generality
problem'. The approach speaks of the (psychological) process that causes a
belief. But commitment to a unique process is problematic. Global reliabil-
ity is a ratio among instances, so, strictly speaking, it only holds of a process
type. But whenever a given belief is produced, the process token that gen-
erates it may be described in different ways.'' Correlated with these differ-
ent descriptions are different process types, and these types may have
different reliability properties. The question is: which of these many types
should be used in fixing reliability? Should the relevant type be sliced
broadly or narrowly? In the case of a perceptual belief-forming process, for
example, should the relevant type simply be 'perceptual causation'? Or, at
the other extreme, should it include detailed features of the retinal stimula-
tion peculiar to this specific instance of belief causation?
Before addressing this problem, I must note that it is not peculiar to the
reliability approach, but probably faces any process account of knowledge.
Since some such theory, I have argued, is essential, the problem is not pe-
culiar to reliabi~ism.~~
The generality problem was identified in my earlier paper "What Is Jus-
tified Belief?" but the dilemma posed by the problem has been emphasized
by others.21If type selection determines very broad types, there is the No
Distinction problem. Every case of perceptual belief causation will be ca-
tegorized the same way, including both cases where the object is far away
and seen only briefly and cases where the object is close at hand, observed
at leisure, and in optimal viewing conditions. This seems wrong because our
temptation to credit the cognizer with 'knowledge' differs in these cases,
and this ought to be traced to general reliability differences. But if type se-
lection determines extremely narrow types, there is the Single Case prob-
lem. When the type is extremely narrow, there may be only one actual
50 Knowledge
instance, namely, the instance in question. Since this instance by hypothesis

yields a true belief (otherwise it would not even arise as a serious candidate
for knowledge), the type will have a truth ratio of 1.Intuitively, it might not
be a case of knowledge, but the reliability approach would not have the ma-
terials to imply this judgment.
Now the Single Case problem arises only if global reliability is deter-
mined exclusively by actual frequencies. As suggested above, however, a
propensity approach is preferable. Since we can thereby put aside the Sin-
gle Case problem, we are in a position to favor a narrow principle of type
individuation. Certainly narrow types are needed to draw the desired dis-
tinctions between processes, those that intuitively do yield knowledge and
those that do not.
But how is it determined, in each specific case, which process type is crit-
ical? One thing we do not want to do is invoke factors external to the cog-
nizer's psychology. The sorts of processes we're discussing are purely
internal processes. Let me advance a conjecture about the selection of pro-
cess types, without full confidence. The conjecture is: the critical type is the
narrowest type that is causally operative in producing the belief token in
question.
To illustrate this idea, suppose (purely for illustrative purposes) that there
is a template mechanism for forming perceptual beliefs. The mechanism
takes feature inputs from the sensory systems and tries to match them to
various templates, each template representing some category C, for exam-
ple, the category of dogs, or the category of cars, or the category of chairs.
Suppose the mechanism so functions that there is a value T such that if the
input features match the template of category C to degree T or more, then
the mechanism generates a belief that the stimulus object belongs to C.
Now if the value of T is very low, then even when a stimulus is highly de-
graded-partly obscured, or seen under adverse lighting conditions-fea-
ture inputs resulting from the stimulus may prompt a belief that the
stimulus belongs to C. Suppose such a belief on a given occasion is true.
Should we call it knowledge? Doubtless we would be leery of doing so. Our
reluctance can be explained by pointing to the unreliability of the mecha-
nism. A mechanism of the sort postulated, with a low threshokE for match-
ing, will tend to be quite unreliable. So as long as the value of the matching
threshold is included in the chosen process type, we will get the right an-
swer on this knowledge ascription case.
But notice that the mechanism has lots of different matching properties.
It has the property of producing a belief when the degree of match is T; the
property of producing a belief when the degree of match is T + .l; the
property of producing a belief when the degree of match is T + .2; and so
on. Is the appropriate process type always one that includes the first of
these properties, namely, the minimal degree of match sufficient for belief?
Knowledge 51
Presumably not. For consider a case in which the perceptual input is not
degraded, and the actual degree of match to the template is, let's say, .99
(on a scale of 0 to 1).Then presumably we will want to say that this is ade-
quate for knowledge (if everything else goes well). But if the selected pro-
cess type still includes the minimal value T, the type as a whole may not
have sufficient reliability.
My proposed account would handle this case by noting that the mecha-
nism's property of having T (say, .70) as minimally sufficient is not causally
operative in this case. The property of having this threshold value does not
play a critical causal role in eliciting the belief. The degree of match in this
case is actually .99. So the critical aspect of the mechanism's functioning
that produces the belief is the propensity to produce a belief when the de-
gree of match is .99. If this property, rather than the others, is included in
the selected process type, an appropriate degree of reliability is chosen that
meshes with our knowledge ascription intuition.
Clearly, this proposal needs to be developed and refined, but I will not
try to do that here. I present it only as a promising lead toward a solution of
the generality problem.
3.3. Second-Order Processes

For a belief to count as knowledge, I am arguing, it must be caused by a
generally reliable process. Exactly how reliable I have not said. Nor do I
think this can be answered with precision. The knowledge concept is vague
on this dimension, and an analysis need not impose more precision than the
common sense concept contains.
I want to turn now to another wrinkle in the reliable-process theory. So
far, attention has been restricted to immediate causes of a belief, not remote
cognitive ancestry. But the attainment of knowledge may depend on such
extended ancestry.
Knowledge may depend on remote cognitive ancestry for two reasons.
First, when a person believes p by inference from antecedently held beliefs,
whether he knows p depends on more than the reliability of this final infer-
ence procedure. It also depends on the antecedent beliefs and how they
were derived. Those beliefs must themselves be known, or at least believed
justifiably, and this implies (as I shall argue in subsequent chapters) that
they must have been caused by reliable processes. Since those beliefs' causal
ancestry may be remote, so may be the relevant ancestry of the belief in p.
A second reason why knowledge depends on cognitive ancestry concerns
second-order processes, that is, processes used in acquiring processes. To il-
lustrate, suppose our friend Humperdink has attended a series of talks on
mathematics by a certain Elmer Fraud. These talks are not under the aus-
pices of any certified educational institution, and Humperdink has been
52 Knowledge
warned that Fraud has no credentials in mathematics. Humperdink hears

Fraud enunciate numerous principles and algorithms, almost all of them
defective. Nonetheless, being a complete novice-and a gullible one at
that-Humperdink blindly accepts and applies them all. In one case, how-
ever, Fraud happens to teach a perfectly correct algorithm. Humperdink
internalizes this one along with the others, and applies it to a relevant class
of problems. In using this algorithm to solve a problem, Humperdink gets
the answer right and forms a true belief in the answer. This belief is the re-
sult of a reliable process, namely, the algorithm. (Later I will call such algo-
rithms methods, rather than processes, but the distinction will be waived for
now.)" Clearly, though, Humperdink should not be credited with knowl-
edge. He does not know because it is quite accidental that he has hold of a
reliable algorithm. Using blind faith to acquire algorithms is notoriously
unreliable; an algorithm so acquired cannot transmit knowledge, though
the algorithm itself may be perfectly reliable.23
This example suggests that a further requirement for knowledge is
needed. Not only must the belief result from a reliable process, or method,
the process or method used must have been acquired (or sustained) by a
suitable second-order process.
Can this further requirement be dispensed with? Can Humperdink's fail-
ure of knowledge be explained by the fact that his belief in the answer is
based on a belief in the correctness of the algorithm, which is an unjustifid
belief? This might work in the present example, but it will not be generally
adequate. Many of our belief-forming procedures are not believed by us to
be correct; we may not have any belief about them at all. They may be
merely automatized procedures, not objects of 'declarative' belief (in the
terminology of cognitive science.)24If these procedures have been unsuit-
ably acquired, however, they cannot yield knowledge, even if they are com-
pletely reliable. Suppose, then, that in addition to the principle below,
(1) S knows that p only if S's belief in p results from a reliable belief-
forming process,
we also accept the following principle:
(2) An acquired belief-forming process (or method) can generate knowl-
edge only if it is acquired (or sustained)by an appropriate second-order
process.
The question immediately arises: What determines 'appropriateness' for
second-order processes? The reliability theme readily suggests an answer,
already mentioned in Chapter 1. The obvious proposal is that a second-
order process is appropriate just in case it is second-order reliable. But what
is second-order reliability? One answer (also mentioned in Chapter 1)is that
the ratio of reliable processes among the processes it generates must be very
Knowledge 53
high. However, this requirement might be too strong. A weaker require-

ment may suffice: the second-order process need only acquire processes that
are more reliable than previous ones used in the same contexts. Whichever
of these proposals is adopted, second-order reliability seems to be a promis-
ing kind of ingredient for the appropriateness of second-order processes.25
Second-order processes are of special importance to one of the book's
general themes: the link of epistemology to psychology. Many first-order
procedures used to form beliefs are not deep-seated psychological pro-
cesses. They are learned algorithms or domain-specific inference precepts.
But when we ascend (descend?)to second-order processes, or perhaps even
higher-order processes, we eventually come to processes that clearly are
deep-seated psychological processes: parts of the architecture of cognition.
If, as my recent ruminations suggest, such processes are relevant to ques-
tions of knowledge, it becomes clear how portions of psychology that study
basic cognitive mechanisms are relevant to judgments of knowledge. Only if
(some of) our basic cognitive processes are either reliable or higher-ofder
reliable can we qualify as knowers. Therefore, whether we so qualify
hinges, in part, on facts in psychology's bailiwick.
Parenthetically, I might add that similar points hold for other themes of
epistemological interest, though not specially pertinent to knowledge
ascriptions. Take the topic of evidence-gathering strategies. These too may
be learned strategies, which are partly domain-specific.But they are doubt-
less learned with the help of higher-order processes, which may be wired-in
features of our cognitive architecture. Since these processes are also of
epistemologicalsignificance, we have another area in which psychology can
contribute to epistemology.
3.4. Knowledge and Justif;catwn

Still another refinement is needed to achieve a fully adequate account of
knowledge. There are cases in which beliefs are caused by globally and lo-
cally reliable processes, yet the person has reason to believe they are not
reliable. These are cases where the belief, though caused by a reliable pro-
cess, is not justified.
Two examples should be mentioned. In the first Millicent in fact possesses
her normal visual powers, but has cogent reasons to believe these powers
are temporarily deranged. She is a subject of a neurosurgeon's experiments,
and the surgeon faksely tells her that current implantations are causing
malfunction in her visual cortex. She is persuaded that her present visual
appearances are no guide at all to reality. Yet despite this belief, she con-
tinues to place credence in her visual percepts. She ignores her well-justi-
fied belief in the incapacitation of her visual faculty; she persists in
believing, on the basis of visual appearances, that a chair is before her, that
54 Know ledge
the neurosurgeon is wearing a yellow smock, and so on. Now these beliefs
are all, in fact, true. Moreover, they are formed by the usual, quite reliable,
perceptual processes. But are they specimens of knowledge? Intuitively, no.
The reason is that Millicent is not justijied in holding these beliefs; they
contravene her best evidence. It seems, then, that causation by reliable
processes is not sufficient for knowing.
Another example is this: Maurice uses a reliable-but not perfectly reli-
able-heuristic to arrive at a certain belief p. Now the fact that the process
is not perfectly reliable does not by itself preclude knowledge. I do not as-
sume that perfect reliability is required. But there is another heuristic
Maurice knows, which is more reliable than the first (though still not per-
-feet), and Maurice believes it is more reliable. He even suspects, in this case,
that the better heuristic might yield a different result, since it has differed
from the first in similar cases before. But despite these beliefs, Maurice ne-
glects the superior heuristic. Had he used it, it would indeed have led him
to a different conclusion: to believe not-p rather than p. But in this particu-
lar case the first heuristic gets things right: p happens to be true. Does
Maurice know p? Intuitively, no. He does not know, once again, because bis
belief in p is not justified. He should have consulted the superior heuristic.
If knowledge requires justifiedness, as these cases suggest, I cannot com-
plete my account of knowledge without a theory of justifiedness. But does it
follow that the reliable-process theory is on the wrong track entirely? That
justifiedness is a wholly different element, quite out of spirit with the
themes I have stressed until now? That of course is possible. It might bk
supposed that once the justificational element is introduced into knowl-
edge, it will change our picture of knowledge dramatically. But that is not
what I shall argue. On the contrary, my account of justified belief will fea-
ture some of the same ingredients discussed thus far, though couched in a
different format. Further requirements will be added, however, to handle
the foregoing cases (especially the Millicent case).
The reader may wonder whether my account handles the standard exam-
ples in the post-Gettier knowledge trade. What about the original cases in
which the (true) belief in p is (justifiably) inferred from false premises? And
what about the so-called social dimensions, where things known to other
cognizers, but not to S, may keep S from knowing p? These cases can be
handled, I believe, by the local reliability requirement, the no-relevant-al-
ternatives condition.
Suppose S believes that either Jones owns a Ford or Brown is in Barce-
lona, and his justification consists in evidence for Jones's owning a Ford. In
fact, Jones does not own a Ford, but by sheer coincidence Brown is in Bar-
celona. S does not know that either Jones owns a Ford or Brown is in Bar-
celona, although this is true.% Why doesn't he know? s's true belief is not
knowledge because it is defeated by the following relevant alternative,
Knowledge 55
which is compatible with all his evidence: Jones doesn't own a Ford and
Brown isn't in Barcelona. This counterfactual situation is one that S cannot
discriminate (with his evidence) from the actual state of affairs.
Similarly, consider the case in which you apparently see Tom Grabit steal
a book from the library. Unbeknownst to you, Tom's mother has said that
Tom is miles away while his twin brother, John, is in the library. This may
be enough to defeat any claim on your part to know that Tom stole the
book, even if it is true." My explanation of this case is that you cannot dis-
criminate this truth from John's stealing the book. Furthermore, the latter
alternative seems to be relevant, as long as Mrs. Grabit says what she says.
When it transpires, however, that there is no twin John-he's only a fig-
ment of Mrs. Grabit's demented imagination-this ceases to be a relevant
alternative, and you can be credited with knowing. In this fashion, the no-
relevant-alternatives element of my account seems able to handle Gettier
and post-Gettier examples.
3.5. Knowledge and Skepticism

Where does this reliable-process approach to knowledge leave us on the
issue of skepticism? Can it help meet skeptical challenges? Or does it pre-
sent any new avenues to skepticism? Finally, what does it imply about the
relationship between psychology and skepticism?
On the reliable-process account, of course, the critical factor in knowl-
edge is the acquisition of true belief via processes that are both locally and
globally reliable. It seems clear that it is at least logically possible to acquire
true beliefs via processes that satisfy these requirements. So at least the 'log-
ical possibility' of knowledge seems to be secured. Furthermore, this may be
all that can reasonably be expected of an analysis. An analysis cannot be ex-
pected to entail that we do know, or that we know very much or very often.
Indeed, it cannot even be expected to entail that it is humcmly possible to
know. That may depend on human powers-in this case, the possession of
sdciently reliable belief-forming processes. And the existence of such
powers cannot reasonably be secured by an analysis of knowing per se.
However, some analyses might open the door to skepticism wider than
others. For example, an analysis of knowledge that requires the exclusion of
all logically possible alternatives would open the door very wide indeed.
My account, however, which requires the exclusion of only 'relevant' alter-
natives, does not open the door quite so wide. In particular, I assume that I
can know there is a computer keyboard in front of me because I do not have
to discriminate that state of affairs from the possibility that I am merely a
brain in a vat being artificially stimulated to make it appear as if there is a
keyboard in front of me. The brain-in-a-vat alternative just is not a relevant
alternative. (I do not, however, have a detailed theory of relevance.)
56 Knowledge
There is, however, a related argument for skepticism not handled by this
reply: the problem of closure. It is an initially plausible principle that
knowledge is closed under known logical implication. If I know that p and I
know that p logically implies q, then I must know that q. In the present case
let p be that there is a keyboard in front of me, and let q be that it's not the
case that I am a mere brain in a vat . . . Since I know that p logically implies
q, I only know p if I know q. But do I know q? Do I know that the envat-
ment hypothesis is false? It does not seem as if I discriminate that from its
relevant alternatives. So it seems as if I do not know q. But that seems to
imply that I do not know p.
Dretske and Nozick have argued, plausibly, against the principle of clo-
sure.* To legitimate the rejection of closure, though, one needs an account
of knowledge that sustains this rejection. But, as McGinn points out, the dis-
crimination account does sustain n o n c l o s ~ r eDifferent
.~~ propositions carry
with them different requirements as to the discriminative capacities neces-
sary for knowledge of them. And knowledge of logically weaker proposi-
tions can require greater discriminative powers than knowledge of logically
stronger propositions. It can be easier to know p than q although p implies q
(and not vice versa) because q requires more in the way of discrimination
than p. To know there is a keyboard before me I needn't discriminate this
state of affairs from envatment. But to know I am not envatted, I do need to
discriminate this state of affairs from envatment. So I may not know I am
not envatted. Yet this possibility does not preclude my knowing there is a
keyboard before me.
The kind of knowledge skepticism I have principally examined is direct
skepticism, which claims that it is impossible to know. But there is a differ-
ent sort of question, concerning iterative skepticism (see Chapter 2), that
may perturb many people regarding the reliable-process approach.
Granted that such a theory leaves open the possibility of knowledge, does it
allow us to know that we know? Many discussions in epistemology go astray
in failing to distinguish questions of first-order and higher-order knowledge
(or ju~tification).~'But once the distinction is carefully made, there is no
objection to raising questions and doubts about the prospects for higher-
order knowledge, and the implications different theories have for such
knowledge.
Is the reliable-process theory worse off than other theories vis-a-vis
higher-order knowledge? No. At least it isn't worse off than a good theory
ought to be. We have already seen that it makes knowledge a logical possi-
bility. By the same token, it makes it logically possible to know that one
knows. To know that we (sometimes or often) know, we would have to
know that we (sometimes or often) use reliable processes of belief forma-
tion. But since the analysis makes it (logically) possible for us to know what
processes we use, and makes it (logically) possible for us to know all sorts of
Knowledge 57
truths about the world (which is essential for knowing the reliability of our
processes), the analysis makes it possible for us to have higher-order knowl-
edge. To be sure, the analysis does not entail the actual existence of higher-
order knowledge. It does not even entail that if we know, then we know
that we know. But no analysis of knowledge should entail this. A plausible
theory ought to have the property that knowing that one knows is more
difficult than simply knowing. It is a virtue of this theory that it has this
property.
Granted that my analysis allows the possibility of both first-order and
higher-order knowledge, what does it say, or imply, about whether we do in
fact know (very much of the time)? As indicated earlier, the analysis per se
does not speak to this question. It does imply, though, that we do know only
if we have reliable (and higher-order reliable) cognitive processes. It follows
that to find out whether we know, we need to ascertain the properties of
our native cognitive processes. This is where psychology enters the picture.
Psychology can (in principle) tell us about the nature of our cognitive pro-
cesses. When these processes are spelled out, we can try to determine their
reliability. (It is doubtful, though, that psychology alone can determine
their reliability. For this we need help from a variety of intellectual fields,
probably including logic and statistics.)
Notice that psychology is needed not merely to tell us whether we do
know, but whether it is humanly possible to know. The reliable-process the-
ory of knowing entails the logical possibility of knowledge, but it does not
entail that knowledge is humanly possible. It is humanly possible only if
humans have suitable cognitive equipment. And this is something of which
we can best be apprised only with the help of psychology.
In saying that whether we do or can know depends on psychological
facts, I partially concur with Quine when he says that "sceptical doubts are
scientific doubt^."^' But my agreement is only partial. Some, but not all,
skeptical doubts are scientific doubts. Some routes to skepticism arise from
concern over the propriety of crediting someone with knowledge if certain
logically possible alternatives cannot be excluded. The best way to counter
this skeptical maneuver is through a satisfactory analysis of knowledge, not
through psychology or other branches of science. (Of course, any piece of
linguistic analysis is plausibly viewed as applied psycholinguistics, and
hence as part of science. This is not what Quine had in mind, though, given
his strictures about analysis and analyticity.) But this sort of maneuver, as I
have discussed, can be met by the reliable-process account.
chapter 4
Justification: A Rule Framework
4.1. Justification and Rules

There are several reasons for interest in the notion of justification. First, we
may have a purely intrinsic interest in the question of when people's beliefs
are justified. Second, we may be concerned with skepticism, which is some-
times couched in the justification theme: Is it possible to have justified be-
liefs? Third, our interest in justification may derive from an interest in
knowledge. Justified belief, I have shown, is necessary for knowing. So
ascriptions of knowledge, and questions about the possibility of knowledge,
hinge on justifiedness.
These reasons for interest in justification all arise in theoretical contexts.
But judgments of justifiedness can also arise in practical contexts. Suppose I
am trying to determine whether or not p is true. So far, my evidence is very
scant. I now learn that Delphine believes p. How much does that help me in
my determination? Not very much; it is not very decisive unless I know
whether Delphine's belief is justified or unjustified. If her belief is unjusti-
fied, then her believing p is not much of a guide to the truth or falsity of p.
But if her belief is justified, I have much stronger evidence for the truth of
p. I have a practical reason, then, for wanting an assessment of Delphine's
belief on the justifiedness dimension.' (Notice how this point meshes with a
reliabilist theory of justifiedness. It makes sense to regard someone's justi-
fied belief in p as evidence for the truth of p if justified belief is belief
formed by reliable processes.)
There are multiple reasons, then, for interest in the notion of justified-
ness. For these reasons, and because it has played a central role in tradi-
tional epistemological theory, I shall devote two chapters to the topic. It
should be acknowledged from the start that there is much vagueness in the
(epistemic) use of 'justified'. No unique conception of justifiedness is em-
braced by everyday thought or language. This will emerge in Chapter 5,
where I present several distinct accounts of justifiedness, each with some
hold on intuition. However, these accounts form a close-knit family; so
Justification:A Rule Framework 59
there seems to be a core idea of justifiedness, which my theory will seek to

capture.
I approach justification in terms of a rule framework. This is warranted
by purely semantic connotations. Calling a belief justified implies that it is a
proper doxastic attitude, one to which the cognizer has an epistemic right or
entitlement. These notions have a strong deontic flavor, as indicated in
Chapter 1. They are naturally captured in the language of 'permission' and
'prohibition', which readily invite a rule formulation.
There are two additional reasons for the rule framework. First, it helps
formulate a comprehensive classification of theories of justification. It pro-
vides a neutral structure within which competing conceptions can be arti-
culated. Second, it lends itself to instructive parallels between epistemic
evaluation and moral and social evaluation. Questions about objectivity and
subjectivity, absolutism and relativism, are points of lively controversy in all
of these domains. The rule framework helps formulate such questions with
greater theoretical rigor than would otherwise be possible, and thereby
helps us see the issues clearly.
We should be reminded, however, of the caution expressed in Chapter 1.
Talk of rules naturally suggests a regulative conception of evaluation: an
attempt to provide advice, decision guides, or recipes, for making doxastic
choices. But the rules I shall be discussing should not be understood as rules
for guiding a cognizer's intellect. A person need not even understand the
rules, and if he does, he need not be able to apply them in the process of
belief formation. Perhaps rules advanced by some theorists could serve as
(first-person)decision guides; perhaps I too would endorse such rules for
secondary epistemology. But the rules I envisage for primary epistemology
would not have this property.
A good first pass at a rule framework for justification is presented by the
following principle:
(Pl) S's believing p at time t is justified if and only if
S's believing p at t is permitted by a right system of justificational
rules (J-rules).
This principle can be generalized for all sorts of doxastic attitudes, not
just beliefs. But I will not be concerned with this generalization until Sec-
tion 5.7.
Principle (PI) is intended to express a semantic truth about the language
of justified belief. Indeed, it is intended to be a purely formal, largely neu-
tral, principle. Although it is not wholly uncontroversial, I would expect it
to be acceptable, at least as a necessary condition, to epistemologists of very
different persuasions. It is not intended to identify any particular factual
standards for justified belief. To obtain a full theory of justifiedness, one
must go well beyond principle (Pl). Indeed, two additional levels, or stages,
60 Justification: A Rule Framework
of theorizing are needed to get a complete account. I will return to these

levels shortly.
Several other comments about (Pl)are needed. First, a comment about
the term 'right'. Obviously, all sorts of different justificational rules, or sys-
tems of rules, might be proposed. Many candidate rules are quite silly, and
would not be endorsed by any epistemologist (for example, 'A cognizer is
permitted to believe any self-contradictory proposition on Tuesdays'). I
certainly do not want to say that justified belief consists in belief licensed by
any rule, even a silly rule. The only sensible linkage is with correct, or right,
rules of justification, not mere candidate rules. Nor does justifiedness neces-
sarily consist in conformity with widely accepted rules. Widely accepted
rules could be wrong, and if they are, conformity with them does not guar-
antee justifiednes~.~ If justifiedness is to consist in rule conformity, it must be
conformity with right rules.
Second, why do I insert the qualifier 'justificational' before 'rules'? What
is the import of this? It reminds us that there are different terms of episte-
mic appraisal, and a similar rule framework might be used to explicate
those other terms as well. For example, we might explicate 'rational belief
as belief permitted by right rules of rationality. Since 'rational' and 'justi-
fied' are different terms of appraisal, rules of rationality (R-rules)might well
be distinct from rules of justification (J-rules). Hence we need to distinguish
justificational rules, or J-rules, from other possible rules.
Third, (Pl)links justifiedness and conformity with systems of rules. Why
systems, and not simply single rules, are involved cannot be explained now
and is deferred for later treatment.
Fourth, (Pl) links justifiedness with J-rule permission, rather than obliga-
toriness. Why not say that a belief is justified just in case it is ~ e q u i ~ by
ed a
right system of J-rules? The answer is that right J-rules may be quite per-
missive, allowing cognizers leeway in their doxastic attitudes. For example,
if a person has a certain corpus of prior beliefs, the rules might permit him
to infer a further proposition that logically follows from this corpus. But the
rules might not mandate this inference. To take another kind of example, if
there are J-rules for subjective probabilities, they might permit someone in
a certain cognitive state to assign to proposition q a subjective probability
in the interval from .70 to .80.No determinate subjective probability is
mandated; any credence in this interval is permitted. So if a cognizer in this
cognitive state forms a subjective probability of .77 vis-A-vis q, this doxastic
attitude is justified, even though it is merely permitted by the rules, not
required.3
Fifth, systems of J-rules are assumed to permit or prohibit beliefs, directly
or indirectly, as a function of some states, relations, or processes of the cog-
nizer. For example, J-rules might permit a cognizer to form a given belief
Justijication:A Rule Framework 61
because of some appropriate antecedent or current state. Thus, someone

being 'appeared to' in a certain way at t might be permitted to believe p at
t. But someone else not in such a state would not be so permitted. Alterna-
tively, the rules might focus on mental operations. Thus, if S's believing p at
t is the result of a certain operation, or sequence of operations, then his be-
lief is justified if the system of J-rules permits that operation, or sequence of
operations. Here the belief in p would be not directly, but only indirectly,
sanctioned by the rules-because it is an output of licensed operations.
Principle (Pl) does not try to settle just what the J-rules should be. In that
respect, it is quite neutral.
In introducing (PI), I called it a good 'first pass' at a framework principle.
But there is the issue of whether it is strong enough. Although it provides a
necessary condition for justifiedness, whether it provides a sufficient condi-
tion is questionable.
Two possible ways of strengthening it might be proposed. First, in addi-
tion to requiring that the belief be permitted by right rules, it might be re-
quired that the cognizer know that it is permitted, or believe justifiably that
it is permitted, or something along these lines. Of course, there is a problem
of circularity if we build justifiedness, for example, into the framework
principle. For we would then use the concept of justifiedness to set condi-
tions for justifiedness. The same circularity problem arises if we require
knowledge of the permission, for knowledge implies justifiedness. Recall
(from Chapter 1) that we want a theory to give nonepistemic conditions for
ju~tifiedness.~
The circularity problem might be circumvented, though, by proposing a
principle like the following:
(P2) S's believing p at time t is justified if and only if

(a) S's believing p at t is permitted by some right J-rule
system R, and
(b) R permits S to believe at t that the (or a) right J-rule
system permits his believing p at t.
Is (P2) a better principle than (Pl)? Note first that any reason for endors-
ing (P2) as over against (Pl) ought also to be a reason for endorsing further
iteration of the permission requirement, with a (c)-clause that parallels (b),a
(d)-clause that parallels (b) and (c), and so on. Of course, imposing such an
infinite number of requirements could make it humanly impossible to have
justified beliefs. This is precisely the result a skeptic would welcome, so this
sort of strengthening of (Pl) is likely to appeal to epistemologists with a
skeptical streak.
The mere fact that such strengthening would play into the skeptic's
hands should not disqualify it. Perhaps the strengthening is really required
62 ]ustificatwn: A Rule Framework
for a correct account of justifiedness. We must not assume the skeptic to be

mistaken without argument, or without fair examination of the terms or no-
tions in question.
However, it is indeed a level-confusion to replace (PI) with (~2).'Our
analysandum is being justilfied, not justifiably believing that one is justified.
So (P2) is just ill-motivated. To see this, consider young children and other
philosophically unsophisticated cognizers. People in these categories-mst
people--do not even have the concept of J-rules, so they do not have be-
liefs, and presumably cannot be permitted to have beliefs, about whether
their beliefs conform to right J-rules. But it is counterintuitive to suppose
that only philosophers, or other methodological sophisticates, even have a
chance at justified beliefs. To take just one example, why can't ordinary per-
ceptual beliefs be justified without higher-order sophistication? It seems
clear, then, that (P2) is an excessive strengthening of (Pl).
There is a weaker form of strengthening, however, which seems quite
well motivated. Suppose S's belief in p is permitted by a right J-rule system,
but S believes that it is not so permitted. Alternatively, suppose S's belief in
p is permitted by a right J-rule system and S is justijied in believing that it is
not so permitted (whether or not he actually believes that it is not permit-
ted). In either case, it is counterintuitive to regard s's belief in p as justified.
In these cases S's belief, or justification for believing, that the belief in p is
not permitted undmines its permittedness. That is, although the belief is
permitted, this does not intuitively suffice for its justifiedness. A well-moti-
vated strengthening of (Pl), then, is one that requires no undermining of a
beliefs permittedness.
One kind of undermining described in the previous paragraph occurs
when S is justified in believing that her belief in p is not permitted by right
J-rules. Since we are seeking a principle of justifiedness, however, we should
not use the notion of justifiedness in explaining the notion of undermining.
So let us revise the foregoing description by saying that a permission for S to
believe p is undermined if S is permitted (by right J-rules) to have a belief in
the denial of this (lower-level) permission.
I have mentioned two ways, then, in which a beliefs permittedness might
be undermined. First, the belief can be undermined by the cognizer's being
permitted to believe that the belief is not permitted. Second, the belief can
be undermined by the cognizer's believing that the belief is not permitted,
even where the higher-order belief is not itself permitted. But additional
possibilities should be entertained. The cognizer might not have the con-
cept of belief permissibility, or even of right rules. But suppose the cognizer
believes that certain conditions are not satisfied, where those conditions are
in fact necessary for belief permissibility. This is a third way, I think, in
which the beliefs permittedness may be undermined. I will not try to give a
full theory of undermining, but will say more about it in section 5.6 where it
Justijkation: A Rule Framework 63
will play an important role. Meanwhile, I can strengthen (PI) by adding the
no-undermining condition, thereby yielding the following principle:
(P3) S's believing p at t is justified if and only if
(a) S's believing p at t is permitted by a right system of J-rules, and
(b) this permission is not undermined by S's cognitive state at t.
Although (P3) is the real framework principle I mean to endorse, in most
of what follows I will be mainly interested in clause (a), in other words, in
principle (PI). For convenience, then, I will often refer to (Pl) as the frame-
work principle, and I will make reference to (P3) only when it is specially
pertinent.
4.2. Rightness Criteria for ./-Rules

Principles (Pl) and (P3) are formal principles of justifiedness. They do not
give much in the way of substantive standards on which epistemic justified-
ness supervenes. But substantive standards are precisely what we want, as
indicated in Chapter 1. We want to know under what conditions a belief is
justified.
While principle (Pl) does not answer this question, it does tell us some-
thing we need to know to answer it, namely, what system (or systems) of J-
rules is right. How this right system is identified involves a two-stage ap-
proach. First, we need a criterion of rightness. Second, we need to
determine which system of J-rules in fact satisfies the chosen criterion. Now
if we regard (Pl) as a prior stage of theorizing and add these two stages to it,
the theory of justification consists of three stages or levels: (1)the level of
the framework principle, (2) the level of the criterion, and (3) the level of
the J-rule system. Most of my attention to justifiedness in Part I (the remain-
der of this chapter and Chapter 5) focuses on the second stage: the criterion
of rightness. Hefty portions of Part I1 will be devoted to the consequences
of this choice. But I shall not attempt to specify in detail any system of
J-rules that satisfies the chosen criterion.
What is meant by a criterion of rightness? First of all, I mean a standard
of rightness that fulfills the requirements mentioned in Chapter 1. It must
specify factual, substantive conditions for rightness of J-rules. These cannot
be conditions that themselves invoke, or rely upon, epistemic notions. To
illustrate, consider the following putative criterion:
A system R of J-rules is right if and only if
R is the set of rules an ideally rational inquirer would instantiate.
Such a proposal is inadmissible because it employs the term 'rational', itself
an epistemic term. We want only nonepistemic notions to appear in a crite-
rion of rightness.
correct principles of justice. This parallels principle (PI). Second, he offers a

theory of what makes a justice-principle correct; that is, he offers a criterion
of rightness for justice-principles. That criterion is:
A justice-principle p is right if and only if
p would be chosen by parties in the Original Position.
Third, he formulates two specific principles and argues that they satisfy this
criterion, that they would be chosen in the Original Position.
In the area of the law we also find analogous examples, though in this
sphere the criteria are those of a 'genuine', or 'valid', law, not necessarily a
'right' (meaning, morally right) law. Here we have such examples as that of
John Austin (the nineteenth-century legal theorist):
L is a law if and only if
L is a dictate of the sovereign.'
There is also the criterion of so-called legal realism, that the law is what
judges say (or will say) it is. There are two possible variants of this position:
L is what the courts have said is the law,

L is what the court of highest authority will (or would) say is the law.
One of the underlying motivations of legal positivism, and of legal realism,
is a conception of law that makes the law more readily knowable, which is a
severe problem under the natural law conception. The foregoing criteria
were intended (in part) to meet that desideratum. It is questionable, how-
ever, whether they fully succeed in this. For example, suppose the court of
highest authority has not addressed the validity of a given statute, and will
never address it. Then it is not readily verifiable whether the statute is a
valid law.
Issues of verifiability have strongly colored much theorizing in the law.
This is unfortunate from the standpoint of theoretical clarity. Thus, H. L. A.
Hart introduces the notion of a "rule of recognition" in the law, which is
close to what I am calling a '~riterion'.~ Unfortunately, Hart conflates the
two questions that must be distinguished: what makes something a valid
law, and what the procedures are for recognizing, or verifying, that some-
thing is a valid law. Failure to separate these questions can and has led to
some confusion in jurisprudential theory? The same problem exists in
epistemology. It is therefore imperative to understand by 'criterion' of
J-rule rightness only a set of conditions that make (a system of) J-rules right,
not a test, or verification procedure, for J-rule rightness.
correct principles of justice. This parallels principle (PI). Second, he offers a

theory of what makes a justice-principle correct; that is, he offers a criterion
of rightness for justice-principles. That criterion is:
A justice-principle p is right if and only if
p would be chosen by parties in the Original Position.
Third, he formulates two specific principles and argues that they satisfy this
criterion, that they would be chosen in the Original Position.
In the area of the law we also find analogous examples, though in this
sphere the criteria are those of a 'genuine', or 'valid', law, not necessarily a
'right' (meaning, morally right) law. Here we have such examples as that of
John Austin (the nineteenth-century legal theorist):
L is a dictate of the sovereign.'
There is also the criterion of so-called legal realism, that the law is what
judges say (or will say) it is. There are two possible variants of this position:
L is what the courts have said is the law,

L is what the court of highest authority will (or would) say is the law.
One of the underlying motivations of legal positivism, and of legal realism,
is a conception of law that makes the law more readily knowable, which is a
severe problem under the natural law conception. The foregoing criteria
were intended (in part) to meet that desideratum. It is questionable, how-
ever, whether they fully succeed in this. For example, suppose the court of
highest authority has not addressed the validity of a given statute, and will
never address it. Then it is not readily verifiable whether the statute is a
valid law.
Issues of verifiability have strongly colored much theorizing in the law.
This is unfortunate from the standpoint of theoretical clarity. Thus, H. L. A.
Hart introduces the notion of a "rule of recognition" in the law, which is
close to what I am calling a '~riterion'.~ Unfortunately, Hart conflates the
two questions that must be distinguished: what makes something a valid
law, and what the procedures are for recognizing, or verifying, that some-
thing is a valid law. Failure to separate these questions can and has led to
some confusion in jurisprudential theory? The same problem exists in
epistemology. It is therefore imperative to understand by 'criterion' of
J-rule rightness only a set of conditions that make (a system of) J-rules right,
not a test, or verification procedure, for J-rule rightness.
66 Justification:A Rule Framework
It is time to turn to examples of criteria for rightness of J-rule systems.

The examples to follow only give a flavor of the different criteria that might
be advanced, including some variations on popular themes. Obviously, no
attempt is made at exhaustiveness. In all cases the notion for which the cri-
terion is offered is: 'A system R of J-rules is right'.
Possible criteria include:
(Cl) R is a system of rules derivable from logic (and probability theory).
(Cl') R is the system of rules that would be chosen by someone who be-
lieves all truths about logic (and probability theory), but is ignorant
of all contingent facts.
(C2) R is the system of J-rules accepted by the players of one's language
game (~ittgenstein). lo
(C2*) R is the system of J-rules accepted by members of one's disciplinary

matrix (~uhn)."
(C2**) R is the system of J-rules accepted by one's peers (Rorty).12
(C3) Conformity with R would guarantee a coherent set of beliefs.
(C4) R permits doxastic attitudes proportioned to the strength of one's
evidence.
(C5) Conformity with R would maximize the total number of true beliefs
a cognizer would obtain.
In trying to identify an acceptable criterion, what method should be
used? The strategy I endorse is best expressed by the Goodman-Rawls con-
ception of "considered judgments in reflective equilibrium."13 We examine
what rule systems would likely be generated by each candidate criterion.
We reflect on implications of these rule systems for particular judgments of
justifiedness and unjustifiedness.We then see whether these judgments ac-
cord with our pretheoretic intuitions. A criterion is supported to the extent
that implied judgments accord with such intuitions, and weakened to the
extent that they do not. But our initial intuitions are not final. They can be
pruned and adjusted by reflection on candidate rule systems. There are
other tests of a criterion's adequacy as well. Does it generate any rule sys-
tems at all? Does it generate a complete rule system, that is, one that would
imply justifiedness or unjustifiedness for all cases of belief and all doxastic
attitudes?
This procedure need not involve empirical psychology or social science. I
do not claim that psychology plays a role in selecting a criterion of J-rule
rightness. This is not the level at which psychology enters the epistemologi-
cal enterprise. It enters the picture only if and when a criterion is selected
that makes reference to cognitive processes. In other words, psychology is
relevant to stage 3 of a theory of justification, which involves the choice of
particular J-rule systems, not stage 2, which involves the choice of a
criterion.14
Justification: A Rule Framework 67
With these general comments in mind, let us return to the list of candi-
date criteria and look briefly at some of their characteristics and problems.
This is only a preliminary examination, to get a feel for possibilities and
difficulties.
Criteria (Cl) and (Cl") invoke truths of logic, or rules of logic, as the
basis for a J-rule system. This reflects a widespread view that principles of
logical inference provide justificational procedures. On this approach, then,
proper rules of justifiedness are to be generated by truths of logic, or by
someone omniscient about logical truths.
The first point to raise about such an approach concerns completeness. At
most, deductive logic can generate a proper subset of desired rules of justi-
fication; not all appropriate rules can be derived from it. Where nondeduc-
tive inferences are appropriate-and such cases are certainly
legion--truths of deductive logic are of little or no help. Are there truths of
inductive logic to fill the breach? That is very dubious. Many epistemolo-
gists and philosophers of science now question whether there is a body of
truths we can call 'inductive logic'.15 Perhaps probability theory can be ad-
duced to provide a relevant class of truths. But, again, the epistemological
pertinence of probability theory is complicated and problematic.16 Pre-
sumably, some rules of justification should pertain to perceptual beliefs, and
some to beliefs involving memory. Neither truths of deductive logic nor of
probability theory can generate J-rules in these domains. Hence, criteria
(Cl) and (Cl') are inadequate on grounds of incompleteness.
There are, moreover, further fundamental weaknesses in (Cl) and (Cl").
They both assume-(Cl) most explicitly-that some J-rules can be derived
from truths of logic. But that is not so, as we shall see in Chapter 5. A typi-
cal truth of deductive logic is that modus ponens is a valid argument form.
But no (correct) rule of belief formation is derivable from this sort of truth.
Validity of argument forms has no immediate implication for belief-forming
practices.
It might seem as if belief-forming rules can be derived from logical
truths. Such rules could simply permit a cognizer to believe any proposition
in the corpus of logical truths. This is perhaps what (ClO)envisages: a logi-
cally omniscient cognizer formulating rules that allow people to believe, at
a minimum, what the cognizer himself believes, namely, all logical truths.
But this would have unacceptable implications. It would imply that when-
ever a person believes a truth of logic, the belief is justified. That is clearly
wrong. It is possible to believe a truth of logic unjustifiably. I will elaborate
on this point in Chapter 5.
Criterion (ClO)articulates a common stance in epistemology. The epis-
temologist is allowed to appeal to truths of logic, and perhaps probability
theory, in formulating methodological rules; but appeal to contingent prop-
ositions (except perhaps first-person experiential statements) is disallowed.
This idea bears some resemblance to Rawls's conception of the Original Po-
68 Justification: A Rule Framework
sition, the position in which hypothetical parties are to choose principles of

justice. In Rawls's Original Position the parties are behind a "veil of igno-
rance." They do not know what their own situation in society is or will be.
But they are imagined to have ideal knowledge about relevant laws, such as
laws of economics and sociology. Similarly, the hypothetical agent envis-
aged in (Cl*) is supposed to choose a system of J-rules and is endowed with
a certain combination of ideal knowledge and ignorance. He knows all
truths of logic and probability theory, but no contingent truths. So we
might call his position an Epistemologically Original Position.
Why is this the right position from which to choose a system of J-rules? Is
it because truths of logic and probability theory are necessary? Is it because
they are knowable a priori? But why should these properties make the de-
scribed position a suitable Original Position? Is it because such truths are
knowable with epistemic certainty, whereas contingent propositions are
not? This assumption is dubious. Prior to theory, many contingent proposi-
tions seem to be known with as much certainty as many truths of logic
(especially higher-order logic). Finally, the whole idea of equating J-rule
rightness with what would be chosen in an Epistemologically Original Posi-
tion is a questionable one."
Let us turn now to criteria in the (C2) family. The members of this family
all share a social perspective: they make rightness a function of what is ac-
cepted by a society, or culture, presumably that of the target cognizer. On
this approach a person has a justified belief only if his method of forming or
defending the belief satisfies the J-rules of his community (linguistic com-
munity, disciplinary community, or what have you). Any such proposal in-
vites an obvious objection. Why should we assume that what is accepted as
justification-conferring by the members of a particular community really is
justification-conferring? Can't such a community be wrong? Why should
authority be lodged in a cognizer's own community? What is privileged
about it? Generations of scientists and mathematicians have used methods
of inquiry and proof that they regarded as sound but were later shown to be
somehow unsound or defective. Why couldn't the same failing hold for any
arbitrary community of inquirers?
Next let us look briefly at (C3). (C3) presents a coherence criterion of jus-
tifiedness, and this raises the old question of whether coherence (irany of its
interpretations) is a strong enough requirement for justification. I shall
challenge the coherence approach on precisely this point in Chapter 5. At
this juncture, let me merely note that coherence may enter the picture at
other levels as well. Even if coherence is not selected as the criterion of
rightness, a right J-rule system might demand that cognizers be consistent
and use their total evidential corpus. In short, traditional coherentist
themes could be correct even if coherence is not the proper criterion of
J-rule rightness.
Proposal (C4) is next on our list. This criterion has the virtue of being ap-
plicable to all doxastic attitudes, not just beliefs. But it appears to be for-
mally inadmissible, since it employs the term 'evidence', an epistemic term.
This point is correct, but probably remediable. (C4) would need to be re-
vised so as to indicate, in purely factual, nonepistemic terms, what the class
of evidence beliefs are. There are also other objections to be raised against
(C4). I shall present these objections early in Chapter 5.
Criterion (C5)belongs to a family of possible criteria that invoke truth. It
is what I call a 'truth-linked' criterion. One possible objection to (C5) is an
objection against truth-linkedness in general. Is it appropriate to invoke
truth in a criterion of J-rule rightness? Won't this raise a serious problem of
determining which system(s) of J-rules satisfies the criterion? Won't it be
either impossible or circular to use such a criterion? Since I shall ultimately
advocate a truth-linked criterion (or family of criteria), this is an important
issue. But I must defer its full treatment until the end of Chapter 5. The ra-
tionale and plausibility of a truth-linked criterion will be spelled out gradu-
ally as I proceed. It should be recalled, however, that a criterion is not
intended as a test or verification procedure for J-rule rightness. So the pos-
sible difficultyof determining which J-rule system(s) satisfies a truth-linked
criterion is not an insuperable objection.
Apart from truth-linkedness, (C5) is not a proposal I deem acceptable.
The substance of ((25)-maximizing the number of true beliefs-just isn't
the correct ingredient for justifiedness. Also, (C5) omits elements that an
acceptable criterion must include. But these will not be explored until
Chapter 5.
In short, I do not endorse any of the proposals in this sample list. Still,
they are useful illustrations, to give a sense of the options available and the
considerations that may be brought to bear in assessing candidate criteria.
4.3. Objectivism, Relativism, and Related Issues

Discussions of morality and moral rules often provoke the following ques-
tions. Is there a unique moral code, a system of moral rules, valid for every
person and culture? Or are different moral codes valid for different persons
or cultures? Are there objective standards that make a moral code valid, or
is it just a matter of preference, taste, and subjective attitude? Answers to
these sorts of questions give rise to theoretical oppositions. Relativists hold
that valid codes vary from person to person, or culture to culture. Absolu-
tists hold that a single code is valid for all people and cultures at all times.''
Subjectivists maintain that validity, or rightness, of moral rules is just a
matter of subjective opinion. Objectivists believe that objective facts deter-
mine the validity of a moral rule.
Similar questions arise in epistemology. They are most frequently voiced
in discussions of 'rationality',1Qbut have equal force for 'justification'. Is

there a unique system of right J-rules that holds for all inquirers at all times?
Or does the right system of J-rules vary from culture to culture, or even in-
dividual to individual? What about objectivity versus subjectivity? Are
there objective facts that make one system of J-rules right and other systems
wrong?
I propose, first, a change in some of this terminology, and a slight refo-
cusing of problems. Let us first distinguish three views on whether there is a
uniquely right system of J-rules. These views are: monism, nihilism, and
pluralism. Monism says there is exactly one right system of J-rules. Nihilism
says there is no right system of J-rules. Pluralism says there is more than one
right system of J-rules.
Nihilism is a species of skepticism. If there is no right system of J-rules, no
cognizer can ever have a justified belief, because principle (Pl), and (P3),
says that a belief is justified only if it is permitted by a right system of
J-rules. If there is no right system, no such system permits any beliefs. Skepti-
cism is normally combined with another view, that doubt or suspension of
judgment is a justified attitude toward the target class of propositions. This
view cannot be shared by nihilism. For if there is no right system of J-rules,
then no right rules can permit any doxastic attitude whatever, including
suspension of judgment. So nihilism is a species of skepticism, but should
not be identified with all versions thereof.
Turning to pluralism, two species of pluralism may be distinguished. Rel-
ativistic pluralism says that one system of J-rules is right for one person or
culture, while another is right for another person or culture. The cultural
variant would, of course, yield epistemological cultural relativism, analo-
gous to moral cultural relativism. But there is a second, nonrelatiuistic spe-
cies of pluralism. This view holds that two or more systems of J-rules are
each right for every person and culture.
To understand nonrelativistic pluralism, consider a parallel with justice. .
It is plausible that there is a unique criterion of justice or a set of demands
that justice makes for all societies. But many alternative sets of social insti-
tutions could satisfy these demands. Thus, there is a plurality of sets of in-
stitutions each of which would be fully just. For example, justice may
demand some sort of mechanism of equal opportunity, but different mecha-
nisms might work equally well. Again, justice may demand some method of
compensating victims of unjust harms, but different mechanisms might do
this equally well. And perhaps any of these mechanisms would be just in
any society. Similarly, there might be different systems of J-rules, each of
which would be right, as judged by a single criterion of rightness. This
would not imply relativism; for it would not suggest that rightness of J-rule
systems varies from culture to culture, or individual to individual.
My position will favor nonrelativistic pluralism. But the picture will be
clouded somewhat by different conceptions of justifiedness, all of which I

shall regard as defensible. Under some of these conceptions relativistic plu-
ralism will be sanctioned.
Let me turn now to objectivism and subjectivism. Objectivism (on the
topic of justification) holds that the correct criterion of rightness invokes
some sort of objective facts, not merely the opinions or attitudes of individ-
uals or groups vis-a-vis competing schemes of justifiedness. By contrast,
subjectivism holds that the rightness of rules is only a matter of what some
individual or group thinks.
Several variants of subjectivism may be distinguished. Zndividwll subjec-
tivism says that rightness of rules depends on what some individual thinks.
Social subjectivism says that rightness depends on what some group, or
community, thinks. It is common to pair subjectivism with relativism. Thus,
if an individual thinks that a system of J-rules is right, then that system is
right for him-but not necessarily for others. Similarly, if a community
thinks that a system of J-rules is right, then it is right for them-but not nec-
essarily for other communities. Criteria (C2), (C2"), and (C2'") are all vari-
ants of social subjectivism. Although they are not formulated to entail
relativism, this would be a natural interpretation. In other words, if a given
system of J-rules is accepted by a certain linguistic community, then that
system is right for that community-but perhaps not for other communities.
Another variant of subjectivism is indexical subjectivism. Indexical sub-
jectivism is especially common in metaethical theory, but is readily carried
over to the epistemological sphere. Individual indexical subjectivism would
be the view expressible as, 'A system of J-rules is right just in case I think it
is right'. Social indexical subjectivism would be the view expressible as, 'A
system of J-rules is right just in case we (indexing some relevant group) think
it is right'. Like other forms of subjectivism, indexical subjectivism is readily
paired with relativism.
Not all versions of indexical subjectivism are first-person versions (either
singular or plural). They do not all fix rightness by the opinions of the
speaker, or even the speaker's own group. For example, Stephen Stich and
Richard Nisbett propose the following criterion of rightness for J-rules: Rule
R is right if and only if R accords with the reflective practice of the group of
people I (the speaker) think appropriate.20 This criterion uses an indexical
reference to the speaker. But it does not link rightness to the speaker's own
appraisal of the J-rule; rather, it is linked to the appraisal of a designated
group.
Returning to the basic opposition between objectivism and subjectivism,
what kinds of facts should be regarded as objective? Two prime candidates
come to mind. First, the verific, or truth, properties of the procedures the
rules sanction. Assuming some realist conception of truth, the issue of
whether a given system of J-rules promotes truth (that is, belief in truth) is
72 justification: A Rule Framework
an objective matter of fact, which transcends any individual's or group's

opinion about the issue. So truth-linked criteria of rightness are objective
criteria. Second, logical and mathematical properties are objective. So any
criterion that invokes logical properties of propositions, or probabilistic
properties (at least if they are either frequencies or propensities), would
count (to that extent) as objective. For example, a criterion that requires
rule systems only to ensure logical consistency of a belief corpus is
objective.
Although traditional epistemology has not often formulated J-rules very
explicitly, it is clear that the traditional conception of such rules would
have them deal almost exclusively with a person's subjective states.21They
would permit beliefs as a function of present 'appearances', antecedent be-
liefs, and the like. But this does not mean that traditional epistemology is
wedded to a subjectivist criterion of rightness. The content of right rules
may be wholly subjective; but the criterion of rightness may be objectivist.
This point is worth elaborating in connection with coherentism. Coher-
entism can be formulated as a criterion of rightness, as in (C3). Suppose we
interpret 'coherence' there as logical consistency (this is only one possible
interpretation, of course). Then whether a system of J-rules is right, on this
criterion, is an objective matter. It is an objective matter whether confor-
mity with a given system of rules would or would not guarantee logical
consistency.
One could modify (C3) to obtain a subjectivist (and relativist) coher-
entism. Such a criterion follows:
(C3') System R of J-rules is right for cognizer S if and only if
S believes that conformity with R would guarantee the logical
consistency of a belief corpus.
It might be tempting to equate what I have been calling 'objectivism' and
'subjectivism' with what recent epistemologists have called 'externalism'
and 'internalism', respectively. But this is not quite right. Most internalist
epistemologies invoke logical relationships as determinants of justificational
status. But logical relationships are objective matters of fact. So these inter-
nalist theories would also be (at least partly) objectivist theories. Personally,
I am no longer anxious to use the terms 'externalism' and 'internalism',
which readily convey misleading impressions. I find 'objectivism' and 'sub-
jectivism' more useful (though they too can be misleading).
I t should be clear that I favor an objectivist, at least a largely objectivist,
criterion of rightness over a subjectivist criterion. This seems to make best
sense of the strategies employed by most epistemological methodologists.
When a statistician recommends a certain statistical procedure and criti-
cizes others, he usually appeals to the relative rates of error, or probability
of error, that the different methods would yield. Such benchmarks are ob-
jective benchmarks, not simply the statistician's personal opinion. When a

philosopher of science criticizes rules of detachment because they lead to
the lottery paradox, the criticism appeals to the prospect of inconsistency.
Again, this appeals to an objective standard: genuine inconsistency, not sim-
ply what is believed by this individual or group. More generally, when
methodologists have historically argued in favor of this or that piece of
methodology, their arguments have never simply appealed to their own
opinion. They have always rested on such factors as the comparative pro-
pensities of the competing methods to avoid error, to yield greater accu-
racy, or to yield greater logical and/or probabilistic consistency.22
Unless there is some objective criterion, it may be impossible even to
make sense of methodological disagreement. Under relativistic subjecti-
vism, at any rate, individuals and groups do not really disagree, since one
only claims that a given rule system is right for him (or his group), and the
other claims a different rule system right for him (or his group). There is no
disagreement here, only the semblance thereof. This kind of point is famil-
iar from metaethics, which has examined subjectivist and relativist doc-
trines in the moral domain. Many objections to these positions have been
well stated in this literature, and apply mutatis mutandis to the epistemo-
logical domain.23
Despite these points, I think a case can be made for mixed conceptions of
justifiedness, conceptions that include both objective and subjective compo-
nents, especially socially subjective, or culturally subjective, components. I
shall elaborate this possibility a bit in Chapter 5.
Let me turn now to a crucial distinction that will sound similar to some of
the distinctions already drawn, but which is really quite different. What I
have in mind does not mark properties of rightness criteria. Rather, it is a
distinction between senses of 'justified'. I shall mark this distinction with the
phrases 'really justified' and 'apparently justified'. The same sort of distinc-
tion is more commonly marked by the phrases 'objectively justified' and
'subjectively justified'.% This terminology is avoided since I am already .
using 'objective' and 'subjective' to mark different distinctions.
The distinction drawn here parallels one frequently drawn in moral the-
ory, a distinction between actions that are objectively right and subjectively
right. If an agent performs an action that he believes, or even justifiably be-
lieves, to be right, then it is said that the action is subjectively right. It is
objectively right only if it is really right. Similar cases can arise in episte-
mology. Suppose that a given belief of a cognizer is not permitted by any
right system of J-rules, Then according to (P3) this belief is not justified. I
shall say that it is not really justified. But suppose the cognizer believes,
even believes justifiably, that this belief is permitted by a right system of J-
rules. Then I shall say that his belief is apparently justified.
In the moral realm it is a tricky question whether praise or blame should
74 Justificath: A Rule Framework
be allocated as a function of objective rightness or subjective rightness. A

number of writers have suggested that these acts or attitudes should be pre-
dicated on subjective rightness rather than objective rightness. The matter
is far from clear and raises some sticky problems.
Similar issues can arise in legal contexts. Suppose someone violates a law
but justifiably believes at the time of action that it is not a violation. For
example, a prior court ruling had indicated that such actions are legal, and
our agent reasonably infers that this is the law. As it turns out, he is mis-
taken; for a higher court later overturns the earlier ruling. Is the agent cul-
pable? Since he has violated the law, he is legally culpable. But surely his
justifiable belief in the legality of his deed affects our appraisal of him (and
the deed).
Now turn to similar cases in the epistemological realm. Suppose that
J-rule system R is not a right rule system, but a cognizer thinks R is right,
and correctly thinks his belief in p is permitted by R. (However, he does not
infer p from his belief about R; he infers p from other propositions. He just
concurrently thinks that this inference procedure accords with R, which it
does, and that R is right, which it is not.) In fact, his believing p is not per-
mitted by any right J-rule system. According to our framework principle,
his belief in p is not really justified, but it is apparently justified. The ques-
tion is: does this cognizer deserve epistemic praise or blame for this belief?25
On the one hand, by hypothesis his belief is not (really) justified, so it looks
as though he deserves epistemic blame. On the other hand, he is justified in
believing that the belief is justified, so perhaps epistemic blame is ill-de-
served. Indeed, perhaps he deserves epistemic credit for his intellectual
conduct. I think we are pulled in both directions, and this is a proper and
irresoluble tension. The only way to relieve the tension is to acknowledge
the distinction I have drawn and note the category of problematic cases.
Nothing about these problematic cases, I suggest, gives us reason to with-
draw or revise principle (P3). But we should be prepared to understand our
reaction to certain examples concerning justifiedness by calling attention to
the derived sense of 'justifiedness', namely, apparent justifiedness.
4.4. J-Rules as Cognitive-State Transition Rules

The last two sections were devoted to rightness criteria for J-rules, and to
some properties of such criteria. But I have said little about J-rules them-
selves. It is time to rectify this omission. One thing already implied about
J-rules is that they permit or prohibit beliefs (and / or other doxastic atti-
tudes), either directly or indirectly. Let us restrict our attention to permis-
sion rules. A direct form of J-rules would expressly permit certain beliefs, or
would present schemas for belief permission. For example, a rule might
permit belief in any proposition that has a certain type of relation to other
propositions already believed. An indirect form of J-rules would license the
execution of certain belief-forming processes or operations; the rules would

not specify the permitted beliefs, but would indirectly sanction the belief
output of these licensed processes. This is just an indication of two possible
forms of J-rules. We now need to inquire more generally into the form that
J-rules should take.
Two importantly different approaches to J-rules should be scouted first:
interpersonal and intrapersonal approaches. An interpersonal approach says
that J-rules should permit a person to believe things as a function of some
relation to other persons. An intrapersonal approach says that J-rules should
permit beliefs only as a function of some properties intrinsic to the cognizer
himself.
A prime example of the interpersonal approach focuses on reason giving
as an interpersonal activity. The idea here is that justifications are things we
give to other people, which the latter either accept or reject. A person is
justified in holding a certain belief only if he actually gives others, or could
give them, acceptable reasons for the belief. This interpersonal approach is
succinctly characterized by Rorty:
Justification is not a matter of a special relation between ideas (or words) and
objects, but of conversation, of social practice . . . We understand knowledge
when we understand the social justification of belief.26
As this passage indicates, the interpersonal approach may also be called a
'social' approach to justification. But notice that I am addressing here J-rule
contents, not criteria of rightness for J-rules. A social approach to rightness
criteria is an entirely separate matter, which I have discussed in the previ-
ous sections.
On an intrapersonal approach, permission-conditions of J-rules would
make no reference to other people. They would restrict themselves exclu-
sively to the cognizer's own mental contents, such as prior beliefs, percep-
tual field, ostensible memories, cognitive operations, and the like. Most
traditional epistemologies have indeed focused on such mental states; but
they have not been exclusively mentalistic. They have usually included such
conditions as logical relationships between prior beliefs and the target be-
lief. Such logical relationships are not themselves mental. However, since
they make no reference to other people, I shall classify approaches of this
sort as intrapersonal, and I will even speak of them (loosely) as 'mentalistic'
approaches.
There are a number of solid reasons for rejecting the social, or interper-
sonal, approach to J-rules. Before marshaling these reasons, however, let me
remind the reader that I do not mean to neglect social dimensions in my
overall conception of epistemology. These dimensions will be explored in a
future work. But the theory of justilfication-at least the theory of primuy
justification-does not require reference to social dimension^.^'
The first reason I prefer an intrapersonal approach to a social approach is
76 ]ustificatwn: A Rule Framework
that a theory of justification should pertain to all propositions, including

propositions about the existence of other people. A conception of justifica-
tion that presupposes the existence of others leaves no room for the question
of their existence, which a full theory should be capable of addressing. This
is one legitimate reason why the tradition has opted for an intrapersonal
approach.
Second, interpersonal reason giving really depends on intrapersonal rea-
sons. Whatever a person can legitimately say in defense of his or her belief
must report or express mental states: his or her premise beliefs, perceptions,
and the like. What justifies a person is not the statements that might be ut-
tered but the mental states those statements would express (or describe).
Perhaps it will be replied that the cognizer could justify a belief by appeal
to the physical world, not the mental state. A justifying statement might be:
'The meter reads 9.57'. But if the cognizer is to be justified in accepting this
statement, it is not enough that the meter does read 9.57. The meter reading
must have some impact on the cognitive state. The believer's mental condi-
tion, or the mental operations he or she executes, are fundamental for justi-
fiedness. Ability to report these things is, at best, only of derivative
significance.
Third, a prime motivation for a theory of justifiedness is the desire to
complete the theory of knowledge. As discussed in Chapter 3, justified be-
lief is necessary for knowing. But if we are interested in the conception of
justification most closely associated with knowledge, the interpersonal con-
ception of J-rules is too stringent. One can know without being able to give
a justification in the sense indicated by the interpersonal theorist. One may
possess good reasons for belief without being able (fully) to state or articu-
late them in verbal form. Our perceptual beliefs, for example, often depend
on cues of which we are not wholly conscious, or which we could not char-
acterize in words. Many theoreticaI beliefs depend on a host of background
assumptions that are hard, if not impossible, to bring to the surface, much
less itemize e x h a u ~ t i v e l ~ . ~
Fourth, it begs important questions to assume that current justificational
status depends on what the cognizer can say now in support of his or her
belief. Justifiedness may involve historical factors--how prior beliefs origi-
nated-that the cognizer no longer recalls and hence could not deploy in
argument with an interlocutor. Then, too, what people can say in a debate
or defense depends partly on their psychological constitution: how adept
they are at dealing with pressureful situations. Some people who have ex-
cellent reasons for their beliefs might choke or clam up. So dispositions to
say things in an interpersonal situation are not the proper measure of
justifiedness.
Finally, the guiding idea behind the interpersonal conception is that cog-
nizers are justified only if they can show to others that their view is plausi-
ble. I already gave some arguments against this in Chapter 1, but there is
more to say. Which other potential reason-receivers are relevant? 'What
if those persons are too stupid or stubborn to follow one's arguments or
abandon prior prejudices? One's inability to convince them might only
reflect badly on them, not on the believer. If nobody else had been capa-
ble of understanding Einstein's theory of relativity, Einstein would not
have been able to satisfy anyone with his reasons. But it would not follow
that Einstein was not justified in believing his own theory (at least once ap-
propriate observations had been made, such as those concerning the peri-
helion of Mercury). Collectively, these are strong reasons in favor of
the intrapersonal approach, and that is what I shallmainly pursue in this
To help develop J-rules within the intrapersonal approach, let me ini-

tially propose a format involving cognitive state transitions. I shall argue for
a replacement of this format later on, in Chapter 5; but it will be instructive
to use it as far as it seems satisfactory. Since principle (P3) links justifiedness
with permission, we may focus exclusively on permission rules.
On the present proposal a J-rule permits transitions from prior cognitive
states to beliefs. (The generalized version would permit transitions from
prior cognitive states to any doxastic attitudes; but I will continue to con-
centrate on beliefs.) Prior cognitive states may consist of perceptual appear-
ances, ostensible memories, doxastic attitudes, or any combination of such
states. So, for example, a J-rule may say that a person is permitted to make a
transition from a prior state of being appeared to in such-and-such a way to
a successor state of believing that p. Or a J-rule may say that a person is
permitted to go from a prior state of believing propositions q, r, and s to a
successor state of believing propositions q, r, s, and p.
As indicated earlier, we may want J-rules to present schemar for transi-
tion permissions. For example, if what permits a transition is a certain logi-
cal or confirmational relation between antecedent beliefs and succeeding
beliefs, the pertinent logical or confirmational relation may be specified in
'
the rule's antecedent. So we may have a rule-schema: If proposition p is lo-

gically implied (or highly confirmed) by propositions q, r, and s, then a cog-
nizer is permitted to go from a prior state of believing q, r, and s to an
(immediately) succeeding state of believing p.
To facilitate the expression of such rules, let M be the permission opera-
tor, and let / represent a momentary state transition. An expression flanking
/ on its left will represent a prior cognitive state; an expression on its right
will represent a belief---or other doxastic attitude, in the generalized ver-
sion. (Of course, it need not represent the totality of permitted beliefs.) The
whole expression will be enclosed in parentheses. Let B serve as the belief
operator, and let p, q, and so on, be variables. We may then symbolize the
foregoing rule in the following notation:
78 Justijkatwn: A Rule Framewo~*k
If (q & r & s) logically implies p, then M (Bq & Br & Bs / Bp).

An important constraint must be placed on conditions that appear in the
antecedent: they must not include ariy epktetnic terms. Remember (Chap-
ter 1) that we are trying to provide factual, or substantive, standards for
evaluative notions, in this case justifiedness. Working within a rule frame-
work, this means that rules should invoke only factual conditions. There-
fore, epistemic terms must be excluded from rule-conditions.
Some theorists may wish to say that certain beliefs are permitted auto-
matically, no matter what the cognizer's state. Some epistemologists
have held, for example, that a person is always justified in believing a tautol-
ogy, no matter what his prior evidence or beliefs. I can express this in tran-
sition notation by letting V stand for the universal cognitive state: a state
one is automatically in, no matter what other states one is in. Then the
foregoing idea could be expressed with the following rule:
If p is a tautology (or logical truth), then M (V / Bp).
This says that a cognizer is unconditionally permitted to believe any
tautology.
Theories may differ over whether they permit transitions based on par-
tial cognitive states or based on total cognitive states only. The second kind
of theory would assume that all segments of a total cognitive state are rele-
vant to belief permission, whereas the first kind would imply that proper
segments of a total cognitive state can sometimes suffice for belief permis-
sion. Coherentists would probably favor the total cognitive state require-
ment. But my notation can accommodate either approach.
We can now see more clearly the connection between justifiedness and a
system of J-rules. Suppose a cognizer believes p as a result of an inference
permitted by a certain J-rule. This does not guarantee that his belief in p is
justified. For suppose that the prior beliefs (in, say, q, r, and s) were not jus-
tified. Then even though the most recent transition accords with a J-rule,
even a correct J-rule, the belief in p is not justified. At least it is not justified
in a strong sense, the sense most closely associated with knowledge. Making
a permitted inference from uniustijied premise-beliefs does not help yield a
cognizer knowledge. Rather, the prior beliefs must themselves have been
acquired by permitted transitions from still previous states. But these tran-
sitions (if they were permitted) may fall under different J-rules.
In this fashion the justificational status of a belief in p is, first of all, a
function of the beliefs cognitive ancestry, possibly stretching back quite
far. Second, whether or not there has been a licensed, or permitted, history
may well depend on a group of rules, not any single rule. This is one reason
we need to consider rule systems, not just individual rules. Justifiedness de-
pends on conformity with a right system of rules, not ordinarily on confor-
mity with a single rule.
With a picture of J-rules now before u s - a t least one possible picture-
Justijication: A Rule Framework 79
we can reflect on some traditional theories in terms of that picture. Con-

sider the foundationalist and coherentist approaches to justification. These
can now be seen as different views about the substantive contents of a right
system of J-rules. Coherentism was earlier viewed as a criterion of rightness.
But I think it is better viewed as a theory-or family of theories-of right
J-rule contents. So too foundationalism.
The core of foundationalism is a commitment to some special class of be-
liefs-so-called basic beliefs-from which all justification derive^.^' This
would be reflected in a J-rule system that ultimately permits all other be-
liefs only by their relationship to members of this special belief-class. By
contrast, coherentism's rule structure would not feature such a distinctive
status for any belief-class.
Many versions of foundationalism go beyond this core version. For exam-
ple, some foundationalists hold that each member of the special belief-class
is permitted independently of the cognizer's other beliefs. Roughly, their
justificational status is inferentially impenetrable. A second special doctrine
of some foundationalists is that all basic beliefs have the highest possible
justificatory status. Neither of these theses, though, is essential to
fo~ndationalism.~~
Still another inessential feature is the doctrine that justification guaran-
tees truth. This point is worth stressing because some writers attack foun-
dationalism by reference to this feature. In so doing, they perhaps refute a
radical version of foundationalism, but not its core. Both Richard Rorty and
Keith Lehrer launch this form of attack. Lehrer, for example, writes:
The fundamental doctrine of foundation theories is that justification,
whether it is the self-justification of basic beliefs, or the derivative justifica-
tion of non-basic beliefs, guarantees
Rorty makes the same sort of assumption.33In fact, he goes much further.
He assimilates all epistemology to foundationalism-somehow ignoring co-
herentism entirely-and takes himself to refute the possibility of epistemol-
ogy by refuting foundationalism. Now it may well be that historical writers,
such as Descartes and perhaps Spinoza, held some such view. But certainly
not all recent writers have held it. Indeed, it is possible to endorse the core
of foundationalism without saying anything about the relation between jus-
tifiedness and truth, much less implying that justification guarantees truth.
Turning briefly to coherentism, this doctrine might take the form of a
J-rule system with a single rule:
If p coheres better with one's total set T of beliefs than does any competitor
of p, then M (T / Bp).
In spelling out this rule, one would obviously have to spell out (nonepiste-
mically!) the meaning of 'coherence', which admits of various interpreta-
tions. (More on this in Chapters 5 and 9.) Also, there are other possible ways
of formulating coherentism within a J-rule system. Without necessarily ap-

pealing to some special notion of coherence, one could still insist-in a ho-
listic spirit-that every permitted transition depend on the total set of prior
beliefs, never a mere subset of these.
If foundationalism and coherentism are just particular styles of J-rule sys-
tems, each may be compatible with a variety of different rightness criteria.
Different criteria could each be satisfied by a foundationalist rule system;
and different criteria could each be satisfied by a coherentist rule system.
Furthermore, since a single criterion can be multiply satisfied, it is conceiv-
able that the same criterion be satisfied by both a foundationalist rule sys-
tem and a coherentist rule system.3P
It is not a central purpose of my argument to defend either foundationa-
lism or coherentism. I am principally interested in finding a criterion of
rightness; in particular, I shall advance a reliabilist criterion. It will then be
an open question whether J-rule systems satisfying such a criterion would
be foundationalist or coherentist (or both). Which rule system(s) satisfies the
criterion is an empirical question.35 It is part of epistemology, of course, to
address this question. But the prior issue concerns the selection of a crite-
rion. That is the task to which most of Chapter 5 is devoted.
chapter 5
Justification and Reliability
5.1. Logic, Psychology, and J-Rules

In this chapter I gradually wend my way to a criterion-or family of cri-
teria-of J-rule rightness. The strategy is to examine candidate criteria, or
types of criteria, by their implications. I ask which J-rules a criterion would
authorize, and which beliefs would be permitted and hence deemed justi-
fied by those rules. If the implications of a candidate criterion conflict with
intuitions, there is evidence, sometimes conclusive evidence, of its inade-
quacy But I also judge criteria by general plausibility, and I am mindful of
considerations of 'reflective equilibrium'.
Some of the criteria scouted in Chapter 4 appeal to truths of logic, specif-
ically criteria (Cl) and (Cl*). This accords with a widespread assumption
that logic provides us with proper methods. Since 'proper method' is easily
construed as 'justification-conferring method', it is natural to assume that
J-rules-at least those governing reasoning-can be derived from logic. This
assumption is false. As I shall argue, no J-rules may be derived from logic
alone. For this reason, no criterion like (Cl) or (Cl*) is satisfactory.
In Chapter 4 I indicated that a criterion like (Cl) is incomplete: it can-
not generate all needed rules. But what I'm about to point out has nothing
to do with incompleteness. I deny that any correct J-rules can be generated
by logic. First, J-rules cannot be derived from truths of logic in a strict sense
of 'derive'. Second, even if J-rules are loosely derivable, the rules so derived
would be wrong. So logic cannot serve as even a partial criterion of
J-rule rightness.
Are J-rules derivable from logic? Let me first make it clear that by 'logic'
I mean formal logic, which consists of three branches-semantics (model
theory), syntax (proof theory), and recursive function theory--of which I
will concentrate on the first two. Semantics primarily studies validity and
implication, where the validity of a sentence consists of its being true in all
models, and implication of one sentence by another consists of the first
being true in all models where the second is true. Proof theory studies what
82 Jwtif;catwnand Reliability
sentences are provable in various axiomatic systems. An important part of

logic also studies the connections between semantic and syntactic proper-
ties of formal systems.
One thing formal logic does not study, in any form, is psychological
states. Logic is completely silent about states of belief, or other cognitive
subject matter. Yet that is precisely the concern of J-rules, namely, transi-
tions to states of belief (or other doxastic attitudes). There is no way, then, in
which J-rules are literally derivable from, meaning entailed by, truths of
formal logic. Admittedly, I shall shortly find reason to replace the cogni-
tive-state transition format with a slightly different format for J-rules. But
even under the new format it still will not be possible to derive J-rules from
logic.
There is a further reason why J-rules cannot literally be derived from
truths of logic. Truths of logic are purely descriptive, factual statements.
They formulate certain facts-presumably necessary facts-concerning se-
mantic and syntactic properties and relations. By contrast, J-rules are nor-
mative statements. They say what cognitive-state transitions are permitted.
But factual propositions concerning semantic and syntactic matters cannot
literally entail any permissions or entitlements, at least no permissions vis-
i-vis cognitive states. Such normative conclusions just do not follow from
semantic or syntactic truths.
It is widely assumed that logic deals with principles of good reasoning.
Logic is often characterized as the art of reasoning. Unfortunately, such a
billing is a bit of a sham. It isn't that logic courses are not useful for good
reasoning, it's just that there are no well-established principles of good rea-
soning (good cognitive-state transitions), and no satisfactory theory of how
good reasoning is related to formal logic. In short, there is not really a well-
established discipline of informal logic. If there is to be such a discipline, I
think it must be a branch of epistemology. But that is a subject which has
yet to have its foundations firmly established (which is what I am trying to
do here). The relations that would obtain between this branch of epistemol-
ogy and formal logic are in need of delineation. They are not yet well
understood.
The notion that logic yields norms gains credence from the fact that for-
mal logic formulates rules of inference, which appear in both axiomatic
systems and natural deduction systems. But these rules are not belief-for-
mation rules. They are simply rules for writing down formulas. Further-
more, formal logic does not really endorse any inference rules it surveys. It
just tells us semantic or proof-theoretic properties of such rules. This is un-
doubtedly releuant to belief-forming principles, since these properties bear
on whether certain sentences are true if others are. But it does not in itself
tell us whether, or how, such rules may be used in belief formation. This
will become clearer below.
Justificationand Reliability 83
Granted that I-rules are not literally entailed by truths of logic, can't we
construct J-rules on the basis of formal logic? If a certain argument is valid,
isn't a transition from belief in the argument's premises to belief in its con-
clusion justificationally permissible? So aren't (right) J-rules derivable in a
loose sense from facts of validity? No.
Before defending this answer, I need to expand framework principle (Pl),
to make it more determinate. This should be done, I believe, by reformulat-
ing it as a historical principle:
(Pl') A cognizer's belief in p at time t is justified if and only if
it is the final member of a finite sequence of doxastic states of the
cognizer such that some (single) right J-rule system licenses the
transition of each member of the sequence from some earlier
state(s).
Of course, our real framework principle, (P3), should also be recast in this
manner, yielding (P3'). First let us distinguish two sorts of J-rules: belief-
independent and belief-dependent J-rules. Belief-independent J-rules (pos-
sibly an empty class) permit beliefs to be formed independently of prior be-
lief states (or other doxastic states). They might include perceptual J-rules
that permit beliefs as a function of sensory cognitive states only. (However,
perceptual J-rules might make provisions for prior beliefs, thereby becom-
ing belief-dependent rules.) Belief-dependent J-rules permit beliefs as a
function of prior beliefs, and perhaps other prior doxastic states as well.
They would include rules that permit inferences from old beliefs to new,
and rules that permit retention of old beliefs from one time period to an-
other.' Now according to (PIQ),the justificationally relevant history of a
belief may reach back quite far. If a belief is inferred from previous beliefs,
whose permissibility is in turn traceable to still earlier beliefs, and so on, the
justificational history may be extensive.
Let us now return to the question of deriving J-rules from logic. Before
defending my denial of such derivability (even in a loose sense of 'derive'), a
technical point must be noted. Although I have introduced rightness as a .
predicate of J-rule systems, I also want to discuss, especially here, rightness
of individual rules. The relationship is straightforward: an individual J-rule
is right if and only if it belongs to some right J-rule system.2
I have two arguments against (loose) derivability. The first has been
stressed by Gilbert arm an.^ Suppose p is entailed by q, and S already be-
lieves q. Does it follow that S ought to believe p; or even that he m y be-
lieve p? Not at all. Even if he notices that q entails p, it is not clear that an
appropriate move is to add p to his belief corpus. Perhaps what he ought to
do, upon noting that q entails p, is abandon his belief in q! After all, some-
times we learn things that make it advisable to abandon prior belief^.^ We
might learn that prior beliefs have a certain untenable consequence. For
example, you might discover that your prior beliefs jointly entail that you
84 Justijication and Reliability
don't have a head. Instead of concluding that you don't have a head, you
may be better advised to infer the falsity of some prior belief or beliefs.
We see, then, that the mere validity of an argument-for example, < q,
therefore p>--does not imply that a believer in its premise or premises
should become a believer in its conclusion. This should come as no surprise.
An inconsistent set of premises implies every proposition. But this does not
mean that someone who believes an inconsistent set of propositions should
(or even may) form a belief in any random (further)
This point can be explained as follows (though it need not depend on this
explanation). Rules for forming beliefs should promote the formation of true
beliefs. Drawing conclusions by valid inference patterns promotes this goal
if prior beliefs are true. But prior beliefs are not always true; nor is it likely
that they will always be true even if they are formed in accordance with
right rules. (Even right rules are likely to allow some falsehoods to sneak in.)
So it does not follow that right rules should always license the acceptance of
any logical consequence of prior beliefs.
In view of this consideration, J-rules like the following are incorrect
J-rules:
If <q, therefore p> is a valid argument, then M (Bq / Bp).
If this is the envisioned pattern by which J-rules are derivable from truths of
logic (truths of validity), logic would only generate incorrect J-rules.
The second argument for the nonderivability of (right) J-rules from logic
runs as follows. Again let us examine the candidate rule that permits forma-
tion of belief in any logical consequence of prior beliefs. I now raise a differ-
ent kind of difficulty for the rightness of such a rule.
Suppose Claude believes proposition x, at which he arrived earlier by
transitions permitted by a right J-rule system. In other words, given (Pl*),
Claude's belief in x is justified. Suppose further that x entails y, but this en-
tailment is extremely complex (since, say, both x and y are very complex).
Even a sophisticated logician would have a hard time determining that <x,
therefore y> is a valid argument, and Claude is just a logical neophyte.
Claude does not 'see' that it is valid, prove that it is valid, or learn this on
anyone else's authority. Nonetheless, he proceeds to believe y. He 'makes a
transition' from a state of believing x to a state of believing y (as well). Is his
belief in y justified? Intuitively, no. But if the target J-rule were right, the
belief would be justified. Claude would have arrived at his belief in y via a
sequence of transitions permitted by right rules. Since the belief is not justi-
fied, the rule in question is not a right J-rule.'
The problem with the foregoing J-rule is that it permits any old transition
that accords with valid inference patterns. In particular, it places no restric-
tions on how the transition is made. I t does not require the cognizer to un-
derstand why the inference is valid, nor to see the connection between
premises and conclusion. It places no constraints on the process for arriving

at the target belief, no constraints on the specific causal path by which the
belief is formed. As long as the consequent belief 'succeeds' the premise be-
lief, it is permitted. That just seems wrong.
On reflection, the very format of state transitions is a source of the prob-
lem, since state transitions imply nothing about causal processes. To talk of
state transitions is to talk simply of states arrayed in time, without ad-
dressing the processes that may be causally operative in producing this
array. And this leaves a lacuna that underlies the inadequacy of the forego-
ing rule and other rules of its type.
Now the reader might say, OK, let's just understand by 'state transition'
something a bit stronger. Let's understand that the prior state must cause
the successor state. But this does not resolve the problem. There are many
possible causal chains that could connect the two states. Which precise
causal chain connects them is a critical determinant of justifiedness. Like
knowledge, justifiedness is sensitive to specific causal generators of belief.
Principles of validity, by contrast, are entirely silent on such matters. This is
why logic alone cannot provide conditions of justifiedness.
One upshot of this discussion is that we must strengthen our conception
of J-rules in order to arrive at a satisfactory criterion of rightness. Instead of
conceiving of them as specifying mere cognitive-state transitions, we must
conceive of them as specifying cognitive processes, where by 'process' we
mean a determinate kind of causal chain.' This is the sort of thing that logic,
by its nature, cannot provide, but precisely the sort of thing naturally
sought from psychology. Notice there is nothing here that casts doubt on
our framework principle (P3*). The only moral I wish to draw is the need
for a constraint on an acceptable criterion of rightness: no criterion will be
plausible unless the rules it authorizes are permission rules for specific
(types of) cognitive processes.
Is this a rash conclusion? Are there other ways to handle the Claude case?
Wouldn't it suffice to make a small revision in the proposed J-rule? Let us
not only require that an inference <q, therefore p> be valid in fact, but
that a cognizer believe in this validity. Further, let us require that the valid-
ity belief be justified. These requirements will correctly deny justifiedness to
Claude's belief, since ex hypothesi Claude is not justified in believing that
e x , therefore y> is valid. The proposed revision can be written in the fol-
lowing way:
If <q, therefore p> is valid, then M (Bq & B (<q, therefore p> is valid) / Bp).
It is not apparent from this formulation that the validity belief must be jus-
tified. But this requirement is an implication of (PI0).Where a belief results
by inference from others, it cannot be justified unless those other beliefs are
justified (that is, generated in conformity with right J-rules). So inclusion of
86 Justijicationand Reliability
the validity belief in the rule means that if Claude's belief in y is to be justi-
fied, he must not only believe ex,therefore y> to be valid, but that belief
must be justified.
Does this ploy work? No, it has two problems. First, it is too strong, at
least if it is to serve as the prototype for deductive J-rules. Not all inferential
J-rules can include the provision that the cognizer believe in the inference's
validity. Children and adults untrained in formal logic presumably use in-
ference procedures before they acquire the concept of validity. Lacking the
concept of validity, they have no beliefs about validity. Nonetheless, they
presumably form justified beliefs by justification-transmitting inferences.
Second, the revised rule is still too weak, for the same reasons as its prede-
cessor. It still does not impose any requirements on the causal processes that
produce the successor belief states. In fact, it does not even require the
prior belief states to be causes of the successor state. This allows counterex-
amples to be readily constructed. For example, suppose a person (justifi-
ably) believes x and (justifiably)believes that x entails y, but the latter belief
is merely stored in long-term memory and is not presently activated. Under
these circumstances, he may start to believe y on a mere hunch. His belief
may not 'connect up' with his prior beliefs in x and the validity of <x,
therefore y>. If so, his belief in y is not justified, contrary to the implication
of the currently proposed J-rule.
Would it help to amend the rule to say that prior beliefs are 'activated'?
No, this would still leave it open whether a causal connection is established
between predecessor and successor beliefs; and it would leave open the de-
tails of any such connection. By leaving these matters open, similar coun-
terexamples as before are still constructible.
Since this kind of problem is intrinsic to the state-transition framework,
we really do need the previously proposed revision. To be right, J-rules
should not merely permit state transitions; they should permit selected
cognitive operations or processes-processes that sometimes output beliefs.
Thus, right J-rules should take the form of process permissions, which may
be represented schematically as
M (Proc 1), or
M (Proc 1 & Proc 2 & . . . & Proc N),
where the latter represents a complex process formed from elementary
processes by serial concatenation, parallel composition, or other composi-
tional devices.
The conclusions for which I am currently arguing are bolstered by fur-
ther reflection on a logic-based criterion and a state-transition rule format.
In Chapter 4 I noted the possibility of J-rules that grant unconditional belief
permissions. One such J-rule, mentioned in Chapter 4, is this:
If L is any truth of logic, then M (V / B(L)).
Justifcationand Reliability 87
This sort of ]-rule has been endorsed by many confirmation theorists (such
as Camap), though usually formulated in terms of 'rationality'. For exam-
ple, if one assumes that it is rational to place credence in a proposition to
the degree it is confirmed by one's total evidence, and assumes that a logical
truth is automatically confirmed to degree 1no matter what one's evidence,
then it is always rational to place degree of credence 1in any truth of logic.
Translating this into talk of (all or none) 'belief and 'justifiedness', we get
the principle that one is always justified in believing a logical truth, no
matter what one's prior state.
This J-rule runs into the same trouble as its earlier cousins. Since it places
no constraints on how a belief in a logical truth is produced, it implies, to-
gether with (Pl*), that such a belief is justified even if it is arrived at by
sheer hunch or guesswork. No matter how complex the proposition L, no
matter how naive the cognizer, no matter how weak or undeveloped his
cognitive capacity to interrelate L's constituents so as to determine its logi-
cal necessity, his belief in L is justified. This is clearly unacceptable. As be-
fore, the J-rule runs into trouble because it licenses beliefs irrespective of
the cognitive processes that produce them. Correct J-rules can only permit
belief formation via appropriate cognitive processes; this is what the newly
envisaged rule format proposes. But appropriateness cannot be specified by
pure logic. Logic is silent about cognitive processes, so it cannot yield cor-
rect J-rules.
Again, it might seem as if we could improve the foregoing J-rule without
reference to cognitive processes. One of the following two revisions might
be proposed:
If L is any self-euident truth of logic, then M (V / B(L)).
If L is any obvious truth of logic, then M (V / B(L)).

One trouble with the first proposal is that 'self-evident' may be an epistemic
term, in which case it is inadmissible. If it is not epistemic, it must be co- .
vertly psychological; it must refer to what is psychologically easy, or sim-
ple, to ascertain. This is surely what 'obvious', in the next proposal, conveys.
Obviousness, ease, or simplicity are not notions of formal logic. So while
these principles have better prospects than their cousins, they are not
derivable from logic alone.
These cases not only illustrate the impotence of pure logic in generating
proper J-rules, but reinforce the relevance of psychology. To say that a logi-
cal truth is obvious is to say that there are psychologically simple processes
which would disclose its truth. It follows that the proposed belief permis-
sions depend not only on facts of logic but on facts of psychology. What is
simple for one species of cognizer is not necessarily simple for others. Sim-
plicity is relative to innate cognitive architecture, to the class of processes

or operations that are available, or naturally utilized, by the system.' So
what is obvious for human beings is not a function of logic, but of human
psychological structure.
Moreover, it is not enough for a J-rule to permit bqliefs as a function of
obviousness or simplicity. To say that a truth is obvious is to say that it could
be believed as the result of simple-and approved-operations, But sup-
pose that a given cognizer believes it as a result of completely different, and
unapproved, operations. Then his belief is not justified. So an acceptable
J-rule would not grant belief permissions on the basis of the mere 'availabil-
ity' of suitable processes, as the previously mentioned candidates would. It
would only grant belief permissions on the basis of actual utilization of suit-
able processes. Thus, J-rules must specify which processes are permitted.
This returns me to the process format for J-rules, which I earlier saw reason
to endorse.
Before closing this section, let us look at a point some people raise con-
cerning the relation between logic and J-rules. Even if logic alone cannot
generate permissive J-rules, maybe it can generate prohibitive rules. Maybe
it can tell us when it is unjustifiable to believe certain propositions, even if
it cannot say when it is justifiable. The following would be a possible rule:
If proposition N is logically inconsistent, then there is no state K such that
M (K / B(N)).
This rule says that a cognizer is not permitted to believe a logical inconsis-
tency, no matter what his prior condition. But this rule is blatantly mis-
taken. Suppose N is a complex inconsistency, which Stella does not
recognize to be inconsistent. Moreover, Stella is advised by an expert logi-
cian, who is normally careful and honest, that N is actually a logical truth.
Since Stella has good reason to believe in this logician's honesty and exper-
tise, and since she reasons accordingly, she forms a belief in N. Such a belief
is justified, contrary to the rule in question.
Perhaps it would help to weaken the proposal to the following:
For any logical contradictories, q and -q, there is no state K such that
M (K / Bq & B-q).
Even this rule may be objectionable. If a pair of contradictories' are both
quite complex, it may be difficult to tell that they are contradictories. If, as
in the Stella case, the cognizer does not herself appreciate their actual logi-
cal relationship, but has testimonial evidence for their joint truth, she may
be justified in believing them both. Of course, we could weaken the pro-
posal still further:
For any obvious logical contradictories, q and -q, there is no state K such
that M (K / Bq & B-d).
]ustijication and Reliability 89
Inserting the qualifier 'obvious' improves the chances of the rule being cor-
rect. But it also surrenders the hope of providing a purely logic-based rule.
Again it is imperative to invoke a psychological notion, making tacit refer-
ence to human cognitive powers. This again reinforces the unavoidability of
reference to cognitive factors in getting accurate rules of justifiedness or
unjustifiedness.
An analogous point holds concerning possible permissive J-rules. Suppose
we waive Harman's problem about the relation between deductive validity
and right J-rules. Couldn't we then say that while arbitrary valid argument
forms do not generate right J-rules, s a n e valid argument forms do. In par-
ticular, doesn't modus ponens have a right J-rule associated with it? Now
this might be correct. But if it is, it is only because modus ponens represents
a fundamental, primitive cognitive operation. Only this psychological status
can make modus ponens have an associated right J-rule. Once more, what
makes for rightness is not just a matter of logic; it is also a matter of psychol-
ogy. After all, modus ponens is no more valid than other valid argument
forms. It is just psychologically simpler than others?
5.2. Evidence Proportionalism: Confirmation Theory and Statistics

In the first section I introduced the idea that J-rules must be process rules,
rules that permit selected belief-forming processes (or, more generally, dox-
astic attitude-forming processes). The process approach goes against a sub-
stantial tradition concerning the nature of justification (or rationality). So I
want to spend more time looking at that tradition and contrasting my ap-
proach with it.
Perhaps the root idea of the tradition I am opposing can be given the
label evidence proportionalim. It is the idea that justifiedness consists in
proportioning your degree of credence in a hypothesis to the weight of your
evidence. This idea appeared in Chapter 4 as criterion (C4). This criterion
presupposes some way of scaling the weight of evidence, for example, on
the interval from zero to unity. It then suggests that a proper J-rule is only
one that permits a cognizer to fix his degree of belief as a function of evi-
dential weight. If degrees of belief can also be scaled on the unit interval,
then degree of belief should be identical to weight of evidence. It is nor-
mally assumed that weights of evidence are derivable from certain formal
facts together with the cognizer's present evidential corpus. The critical
formal facts are either of a semantic nature-so Carnap viewed statements
of inductive probability, or confirmation theoryl0-r a mathematical na-
ture-belonging to statistics, for example. Logical implication is a special
case of such formal facts, where the premise proposition confers a confir-
mation of degree 1 on the conclusion.
90 Justijicationand Reliability
~f this approach is correct, then all talk of 'processes', especially psycho-

logical processes, seems to be irrelevant. It is therefore important to exam-
ine this approach and see where it goes wrong. Let me stress at the outset
that I shall not be claiming that statistics, for example, is relevant to epis-
temic justification; on the contrary, I shall try to show where it properly
enters the picture. But I shall be arguing that fields like statistics and con-
firmation theory cannot by themselves give us an adequate theory of
justifiedness.
A preliminary point should be mentioned first. Evidence proportionalism
normally assumes that a large range of credal intensities, or 'subjective pro-
babilities', are available to a cognizer. This is a substantive psychological
assumption that is open to question. Indeed, I shall raise doubts about it in
Chapter 15. At this point, though, I will let this facet of evidence propor-
tionalism go unchallenged. Instead, I will assume whatever repertoire of
credal states the evidence proportionalist wishes to postulate, and criticize
the theory on other grounds.
The main point to be made here is essentially that of the previous section.
Whatever the degree of confirmation of evidence on hypothesis, whatever
the degree of statistical or probabilistic support the hypothesis receives
from the evidence, these relationships cannot ~IX the justifiedness of allot-
ting a selected degree of credence vis-8-vis the hypothesis. The question
will always arise: What understanding does the cognizer have of these rela-
tionships? If he lacks proper understanding, then he is not generally justi-
fied in assigning the indicated degree of credence. But 'understanding' can
only arise from use of suitable cognitive processes. So a specification of jus-
tifiedness must make reference to such processes.
To illustrate this general theme, suppose a detective is investigating a
crime, and seven suspects are under study. The detective has a large body of
evidence, including material about individual suspects (their alleged where-
abouts, their past relationships to the victim, and so on) and a great deal of
base-rate data concerning personality types, kinds of interpersonal relation-
ships, and incidence of criminal acts vis-A-vis partners in these relationships.
Now let us suppose there is such a thing as a 'proper' statistical analysis of
all this information, an analysis on which all expert statisticians would
agree. (This is, of course, a fiction, given the differing views among statisti-
cians themselves.) And suppose that ideal statistical analysis would say that
the probability of suspect S1 having committed the crime, given all of this
evidence, is .85. But suppose that the detective himself is ignorant of statis-
tical methodology, and is not even wise enough to consult a statistician.
Suppose he simply eyeballs his evidence and leaps to the conclusion that
the probability of the hypothesis 'S1 did it' is .85; and thereupon forms a de-
gree of belief of .85 in this hypothesis. Is this degree of belief (or subjective
probability) justified? Clearly not. Once again, the psychological processes
Justijication and Reliability 91
actually executed in forming the judgment are critical to the judgment's

justifiedness. Justifiedness is not fixed by an ideal statistical, or confirma-
tional, analysis of the evidence (even assuming there is such a thing, which
is problematic).
At this point the statistical methodologist may object. All right, he may
say, justifiedness is not simply a function of the end results of a statistical
analysis, If the detective forms some degree of belief in Sl's culpability by
sheer guesswork, and this fortuitously coincides with the assignment pre-
scribed by the proper statistical analysis, that does not mean his degree of
belief is justified. But what if the detective actually uses the statistical anal-
ysis? What if he applies the prescribed statistical method, that is,
psychologically 'realizes' or 'instantiates' that method? Wouldn't it then
follow that his resulting degree of belief is justified? That is all the statisti-
cian (or the epistemologist-cum-statistician) claims. As long as the statistical
method is 'used', justifiedness is guaranteed. Admittedly, talk of 'using' the
method implies agreement with the claim that psychological processes are
critical for justifiedness. But this does not mean that statistical method is an
irrelevant, or even secondary, issue. It just means that statistical methods
must be psychologically instantiated in order to yield justifiedness.
That this argument does not succeed is in effect shown by the Humper-
dink-Fraud case of section 3.3. But let me elaborate the point with another
example. Begin with the claim that as long as the proper statistical method
is used, justifiedness is guaranteed. For simplicity, let us move from the sta-
tistical domain to another domain where suitable methods or algorithms are
uncontroversiaI. Take the methodology of deriving square roots. Here there
is a clear algorithm for getting the right 'answer. And it may be suggested
that anyone using this algorithm will be guaranteed not only a true belief in
the calculated answer, but also a justified belief. Is this so?
Suppose Gertrude's mathematical education is seriously deficient: she has
never learned the square root algorithm. One day she runs across the algo-
rithm in a pile of papers written by someone she knows to be a quirky,
unreliable thinker, and no authority at all on mathematical matters. Despite
this background knowledge, she leaps to the conclusion that this rule for
deriving square roots (the rule is so labeled) is a sound method. She proceeds
to follow, and form beliefs in accordance with, this algorithm. She forms
beliefs in propositions of the form 'x is the square root of y'. Are these beliefs
justified? Clearly not, for Gertrude has no adequate grounds for trusting the
results of this algorithm. She herself has inadequate understanding to realize
that it works correctly: nor has she checked to see that the answers are
right; nor does she have good inductive grounds for supposing that a ran-
dom algorithm in this room is correct. So her belief is unjustified; yet she has
used a correct algorithm to arrive at her belief.
This case, like the Humperdink-Fraud case, shows that unless a method is
92 Justification and Reliability
properly acquired (or properly tested after its original acquisition), it can-
not yield a justified belief. The mere fact of usage is not enough. We must
now determine what kinds of acquisition processes are proper. Of course,
some reference may again be made to learned algorithms or techniques
suitable for appraising the soundness of other methods. But the same ques-
tion will arise for them. What is their pedigree in the cognizer's mind?
Were they suitably acquired? Ultimately, the question of justifiedness must
rest on some unlearned processes, on some psychologically basic processes
that are never themselves encoded or stored as recipes, algorithms, or
heuristics. These processes are just executed, not 'applied' in the way that
an algebraic or statistical technique is 'applied'.
We arrive in this fashion at the same conclusion tentatively put forward
earlier. Justifiednesscentrally rests on the use of suitable psychological pro-
cesses; and by psychological processes I here mean basic, elementary pro-
cesses, not acquired techniques that are mentally encoded and applied."
Now suppose statistics indeed provides us with 'ideal' methods. (It is
questionable whether statistics incorporates methods of belief formation, or
credal-state formation. However, various forms of statistical inference do
seem to prescribe credal-state-forming methods. So let us allow this suppo-
sition to pass.) Should we then conclude that unless our native cognitive ap-
paratus instantiates this very set of methods, we never have justified beliefs?
This demand would almost certainly lead to skepticism (in the justification
theme), for it seems highly unlikely that our native cognitive architecture
would realize these methods. In fact, certain psychologists have recently
been engaged in arguing that native inference propensities do not match
those of 'normative' statistical methodology.'2 If our native processes do not
match these statistical methods, then, first of all, ordinary beliefs formed
without the methods cannot be justified. But, worse yet, even beliefs based
on method utilization cannot be justified. Utilization of the methods does
not confer justifiedness unless the methods are acquired by justification-con-
ferring processes. Ultimately such acquisition must utilize native processes
(including second-order processes). But, ex hypothesi, these native processes
do not include the 'ideal' statistical methods. So it will be humanly impossi-
ble to form justified beliefs.
Is the prospect of landing in the skeptical ditch enough to refute the
foregoing proposal? Perhaps not. That would involve too strong an anteced-
ent commitment to skepticism's falsity. But there are independent reasons
for thinking that the requirements of a refined, or optimal, procedure
should not be imposed at every level of cognition. Suppose we can imagin-
atively construct perceptual devices that are superior (say, in reliability) to
our own. Does this mean that none of our perceptual beliefs are justified?
Suppose we can devise a memory system that is better (again, say, in relia-
bility) than ours. Does it follow that none of our memory beliefs are justi-
fied? Surely not. So why should the fact that we can design more refined in-
ference procedures than our native ones show that the latter cannot yield
justified beliefs? Another parallel may be instructive. Suppose that before
Russell presented his paradox, all existing set theories were paradoxical.
Does it follow that nobody before Russell had any justified beliefs about
sets? Surely not. Just because more refined conceptions of sets were later
generated is not enough to show that all earlier beliefs about sets were
unjustified.
All I have been arguing here is that the norms of ideal statistical method-
ology need not be imposed on basic cognitive processes. But it remains to be
seen just what the proper J-rules do require, and just what criterion of
rightness should be chosen to fix correct J-rules. Although I have here
fended off one specter of skepticism, I do not mean to preclude the possibil-
ity that skepticism might enter through another door. Once we choose a cri-
terion of rightness, it might still turn out that no J-rule system that satisfies
the criterion is humanly realizable.
One main result of this section is to emphasize the importance of the dis-
tinction between basic psychological processes, on the one hand, and vari-
ous sorts of algorithms, heuristics, or learnable methodologies, on the other.
Let us call the former simply processes and the latter methods. I shall now
argue that a good theory of justifiedness should systematically separate
these two classes. There is a family of different conceptions of justifiedness,
each having some pull on intuition, which can only be dissected with the
help of this separation.
5.3. Processes, Methods, and Levels of Justijiedness

I have just distinguished native psychological processes from acquired
methods, including algorithms,' heuristics, skills, and techniques of various
sorts. The former mark the province of primary epistemology, while the
latter mark that of secondary epistemology. I now wish to distinguish two
corresponding levels of justification: primary justijiedness (P-justifiedness)
and secondary justijiedness (S-justifiedness).Roughly, P-justifiedness results
from the use of approved processes, and S-justifiedness results from the use
of approved methods.
The natural theory of full justifiedness requires the use of both approved
processes and approved methods (if methods are used at all). However,
there are several reasons why proper use of processes is more fundamental.
First, as we have seen in the previous section, use of correct methods is not
sufficient for justifiedness. Methods must also be properly acquired; and this
ultimately rests on the use of right processes. Second, questions of full justi-
fiedness must not only deal with the matter of how methods are acquired
when they are acquired, but also why they fail to be acquired even when
they are available. If a cognizer has plenty of opportunity to acquire certain

correct methods but fails to do so, his subsequent failure to use such
methods may render his resultant beliefs unjustified. Here failure to use
proper methods is not a flaw at the level of methods so much as a flaw at the
level of processes. The processes are defective because they fail to add
things to the cognitive repertoire that ought to be added.
This sort of case reminds us that within the class of processes are not only
first-order processes but also second-order processes. Second-order pro-
cesses were originally defined as processes that produce new processes. But
with our distinction between 'processes' and 'methods', we should reinter-
pret second-order processes as ones that produce new methods, or are in-
volved in method acquisition. In the previous section I indicated that a
belief produced by right methods is not ips0 facto justified-is not fully jus-
tified, as we should now say. These methods must be acquired by suitable
processes. The best way to phrase this requirement is that the methods must
be acquired by permitted second-order processes. A complete account of P-
justifiedness, then, must provide not only a rightness criterion for first-order
processes, but also a rightness criterion for second-order processes.
The use of proper methods not only depends on suitable processes of
method acquisition, but also depends on suitable processes of method selec-
tion. In any cognitive problem-solving task, a cognizer must choose which
methods to apply. A mature cognizer has an enormous number of stored
methods-principles and heuristics of various sorts that he has either de-
vised by himself or been explicitly taught by others. The task is to deter-
mine which of these methods is relevant and applicable to the problem at
hand. The cognizer can misfire in this task in two ways. He may choose an
incorrect method, or he may fail to choose a correct method (which he pos-
sesses). Either of these mistakes can result from poor processes of two sorts:
(a) poor retrieval processes, or (b) poor reasoning processes. Poor retrieval
processes might result in failure to access the best, or most relevant,
methods stored in memory. Poor reasoning processes might result in a poor
choice among the methods that are retrieved. So inadequacy of method
may be a fault assignable to processes for either of two reasons: a flaw in
method-acquiring processes or a flaw in method-selecting processes.
These points having been made, it remains true that full justifiedness de-
pends partly on the propriety of m e t h o d ~ i t h e ones
r actually used or ones
omitted-not just on the suitability of processes. So a complete theory of
justifiedness must deal with rightness for methods as well as for processes.
Later in this chapter (in section 5.5) I will present an objective criterion
of rightness for process rules, more specifically a reliabilist criterion. (Ac-
tually I will present a family of three criteria, each with substantial appeal.)
Still later (in section 5.7) I will sketch a similar criterion (or family of cri-
teria) of rightness for second-order process rules. It would be reasonably ex-

pected that a parallel criterion for methods should be presented as well. In-
deed, I have strong sympathy for an objective, reliabilist conception of
rightness for methods. But some considerations favor a different conception,
which I shall now note.
Consider examples from early stages of science or mathematics. Some-
body invents methods that, in hindsight, are naive and unreliable. Relative
to the information available at the time, though, practitioners were P-justi-
fied in believing those methods to be sound. We may well be prepared to
credit those early practitioners with fully justified beliefs. If so, we must
be employing some nonobjective conception of S-justifiedness. Specifi-
cally, we may be employing a weaker, culturally relative, conception of
S-justifiedness.
How might we formulate a socially subjective, relativistic, criterion of
rightness of the sort I have in mind? It might run as follows:
A method M is right for culture C at time t just in case M is widely believed
by members of C to be the best method (or a good method) for the class of
problems at issue.
If a scientist belongs to scientific culture C, and if he uses methods widely
endorsed at the time by members of C, then the beliefs formed by these
methods are S-justified.
Let us assume that the rightness criterion at the primary level is some sort
of objective criterion. Then the pairing of that criterion with the socially
subjective rightness criterion at the secondary level yields a mixed criterion,
partly objective and partly (socially) subjective. I noted this possibility in
passing in Chapter 4. I do not mean to endorse this approach, as contrasted
with a fully objective approach. I only mean to air some of the possible
approaches.
Although S-justifiedness is clearly an important component of full justi-
fiedness, the remainder of my discussion of justifiedness in this chapter is
.
confined to P-justifiedness. Elsewhere in the book I occasionally return to
questions of methods, but the bulk of my attention is devoted to processes.
Methods will be addressed more fully in the future volume on social
epistemology.
5.4. Consequentialism
In earlier sections I endorsed a process format for J-rules, arguing that right
J-rules (at the primary level) can only permit basic cognitive processes. This
is really a partial specification of a rightness criterion for J-rules. An ac-
ceptable criterion must, at a minimum, require J-rules to be process rules,
96 Justificationand Reliability
not state-transition rules, for example. What further conditions should ap-
pear in an acceptable criterion remains to be seen. But let us pause and re-
flect on this constraint.
This constraint plays a critical role in vindicating the claim that psychol-
ogy has a role to play in the theory of justifiedness. If right J-rules must spec-
ify basic psychological processes, then cognitive psychology has an
important role to play in determining right J-rules. For cognitive psychol-
ogy (or, more broadly, cognitive science) is the field that explores the
human repertoire of basic cognitive processes. Unless and until these pro-
cesses are identified, no selection can be made of the ones that should be
licensed by a right rule system.
Of course, further conditions must be included in a rightness criterion, As
I develop my account of an acceptable criterion, the argument may con-
ceivably go astray at any point. But if my contention up to this point is cor-
rect-that right J-rules must be basic process rules-then I have already
substantiated one central theme: the importance of psychology for individ-
ual epistemology (at least at one level). I stress this point because the ratio-
nale for investigating psychological processes in Part I1 does not hinge on
all my epistemological theses of Part I. Although Part I1 does investigate
psychological processes with an eye to the truth-linked criteria I embrace in
Part I, a proper place for psychology is already established by the theses
defended thus far, independent of truth-linkedness. Thus even if the rest of
my account of justifiedness (and power) should fail, psychology's role is
secure.
One objection needs to be considered in this connection, though. The
proposed partnership with psychology in the epistemological enterprise-
specifically in the theory of justifiedness-assumes that part of epistemol-
ogy's task is the choice of specific J-rules. But some readers might take ex-
ception to this assumption, asking: Why shouldn't epistemology stop at the
selection of a rightness criterion? Why should epistemology include the job
of identifying the J-rules (or J-rule systems) that actually satisfy the
criterion?
There are three answers here. First, epistemology has traditionally been
concerned not simply with abstract criteria for good methodology or pro-
cedure, but with the endorsement of particular methods or procedures. This
is especially clear in the philosophy of science wing of epistemology. Since
psychological processes, I have argued, are the suitable class of 'procedures'
in primary justifiedness, it is appropriate for primary epistemology to ac-
tually specify the approved processes, not simply to lay down a criterion
they must meet.
Second, it is just part of a full theory of justifiedness to say which J-rules
(or J-rule systems) are right. To stop epistemology at stage 2, the level of the
criterion, is to abort it at an unnatural juncture. Admittedly, the job of arti-
Justificationand Reliability 97
culating a detailed J-rule system goes beyond the horizon of this book; none-
theless, that job is still an epistemological one.
Third, once a criterion of rightness is chosen, it may be an open question
whether any J-rules satisfy it. Perhaps no human psychological processes
meet the constraints of the criterion. In that case skepticism (in the justifi-
cation theme) is victorious: it is not humanly possible to have justified be-
liefs. But since the prospects for this sort of skepticism stand or fall with the
existence of nonexistence of J-rules satisfying the criterion, examining the
satisfiability of the criterion is certainly an assignment for epistemology.
With these points in mind, let us now work toward a more specific right-
ness criterion. The next step is to distinguish two possible approaches,
familiar from ethical theory: consequentialim and deontology. Consequen-
tialism holds that rightness is a function of consequences; in this case it
would hold that rule(-system) rightness is a function of the consequences of
conforming with, or realizing, the rules (or rule system). In opposition to
consequentialism is deontology, which makes rightness rest either wholly or
partly on nonconsequential factors. Criteria discussed earlier are examples
of justificational deontology, namely, the criteria appealing to truths of
logic--(Cl) and (CIh)-and evidence proportionalism-(C4). Evidence
proportionalism is a particularly good example. It says that justificational
rightness is a matter of proportioning strength of conviction to weight of
evidence, no matter what the consequences. This is analogous to the deon-
tological theory of punishment, which says that punishment should be pro-
portioned to the seriousness of the crime, no matter what the effects
may be.
No deontological criterion examined thus far has proved satisfactory.
This does not prove, of course, that none will. But I cannot construct other
good prospects, so I shall look elsewhere. I suspect that some form of conse-
quentialism is correct, and I will now explore varieties of epistemic conse-
quentialism to find the most promising. Of course I restrict my attention to
rule consequentialisms, for I am interested in rightness of rules (or rule sys-
tems). I ignore entirely the suggestion that the justificational status of each
belief is a function of that very beliefs consequences. I also assume that we
are only interested in the direct consequences of rule conformity, not in-
direct ~ o n s e ~ u e n c e s . ' ~
Consequentialisms can differ along two dimensions. First, they may differ
on the issue of which consequences have (positive or negative) ualue. Sec-
ond, they may differ on which meusure of valued and disvalued conse-
quences is relevant to rightness. One familiar measure is mxirnimtion of
net value; another possibility is some sort of satisficing measure. Let me
begin with the question of value. It should be emphasized that I am not
concerned with epistemic value in general, but only with justificationally
relevant value, for I am presently interested in rightness of J-rules.
98 ]ustification and Reliability
Several candidates of justificationally valuable consequences are readily

suggested.
(1) uerijic consequences: believing truths, not believing falsehoods
(2) coherence consequences: achieving coherence in one's belief (or credal
probability) corpus
(3) explanatory consequences: believing propositions that explain other
believed propositions
(4) pragmatic consequences: realizing one's practical, nonintellectual goals
(5) biological consequences: surviving, reproducing, propagating one's
genes
One feature of these candidates, which complicates the picture, is their
causal interdependence. Believing truths often helps a person achieve prac-
tical goals. Realizing practical goals, such as food, shelter, and mating, typi-
cally promotes biological ends. There is also plausibility in supposing that
many cognitive functions subserving the attainment of true belief, or goal
realization, were selected for in evolution because of their biological conse-
quences, that is, their contribution to genetic fitness.
Since consequentialisms often focus on 'ultimate' values, these causal in-
terdependencies and evolutionary possibilities might incline one toward a
biological theory of epistemic value. But I think this would be a mistake.
For one thing, causal interdependence is not identity or equivalence. Be-
lieving truths does not always promote pragmatic or biological ends; wit-
ness the threat to our survival posed by our belief in truths of nuclear
physics. Furthermore, while some cognitive processes promote happiness or
goal realization by way of producing true means-ends beliefs, others pro-
mote happiness by inhibiting true belief. They may shore up the ego with
false, but pragmatically convenient, self-ascriptions.
Second, even if the desire for truth acquisition is ultimately traceable to
biological fitness (curiosity about one's environment can promote survival),
it still appears in the organism as an 'autonomous' desire. People do not de-
sire true belief merely as a means to survival, or the achievement of practi-
cal ends. Truth acquisition is often desired and enjoyed for its own sake, not
for ulterior ends. It would hardly be surprising, then, that intellectual norms
should incorporate true belief as an autonomous value, quite apart from its
possible contribution to biological or practical ends.
These considerations, among many others, lead to the unremarkable con-
clusion that true belief is a prime determinant of intellectual value, and in
particular, a critical value for justifiedness. At any rate, they suggest that
verific consequences are a better candidate for justificational values than
pragmatic or biological consequences. But what about coherence and ex-
planatory consequences? These have figured prominently in many theories
of justification and science. Why choose verific value over these forms of
value in an account of justificational rightness?
Let us first examine coherence consequentialism. 'Coherence' admits of

many interpretations. In a weak sense a corpus of beliefs is coherent just in
case it is logically consistent. But surely this kind of coherence is not strong
enough to serve as a benchmark of justificational rightness. If the promotion
of logical consistency suffices for rightness, it is compatible with the promo-
tion of completely fanciful belief corpora. Mere logical consistency can be
satisfied by belief systems concocted out of thin air, fabricated by paranoia
or neurotic defense mechanisms, or otherwise generated by sheer fantasy.
These sets of beliefs certainly do not count, intuitively, as justified. But
under a consistency criterion of rightness, such beliefs would count as
justified.14
The notion of coherence is readily applied to systems of subjective proba-
bility. Credal, or personal, probabilities are coherent if and only if they
conform to the probability calculus. If the personal probability for hypoth-
esis H is .75, then the personal probability for not-H must be .25, and so on.
Subjective Bayesianism is a theory that makes such probabilistic coherence
necessary and sufficient for rationality. If it were presented as a sufficient
condition of iustificational rightness, it would be a species of justificational
coherentism. (Of course, Bayesians do not normally present their theory as a
form of consequentialism; but that construal, I take it, is innocuous.)
A somewhat stronger version of probabilistic coherentism is kinematic
probabilistic coherentism. This requires not only that credal systems be
probabilistically coherent at any given moment, but that they obey certain
requirements over time: that there be certain continuities between a per-
son's subjective probabilities at time tl and time t2.15
Both forms of probabilistic coherentism-static and kinematic-suffer
the same sorts of difficulties as logical consistency coherentism. Mere fanta-
sizers and deluded paranoids can satisfy their requirements as long as they
are excellent probabilists. Yet the subjective probabilities of wild fantasizers
and deluded paranoids certainly are not justified doxastic states.
Actually, some qualifications must be added to the sketched counterex- .
amples. We must not suppose that the cognizers in these examples are total
fantasizers. They cannot merely fantasize about the probability calculus, if
they are going to make their credal systems conform to it. Nor can they
merely fantasize about their past credal systems, if their present credal sys-
tems are going to display the requisite continuities. They may need genu-
inely accurate memories. Apart from these requirements, though, they may
fantasize as much as they wish. They need not pay attention to their per-
ceptions, for example, but niay discount them as thoroughly as they wish.
Surely, such cognizers provide strong enough counterexamples to these
forms of coherentism.
There are stronger senses of 'coherence', however, beyond logical consis-
tency and conformity with the probability calculus. Some epistemologists,
such as C. I. Lewis and Roderick Chisholm, formulate a conception of co-

herence which requires that a coherent set of propositions mutually support
and confirm one another.'' We might call the weaker forms of coherence
negative coherence. This only requires the absence of logical or probabilis-
tic conflict. The stronger form of coherence may be called positive coher-
ence. l7 It requires positive (mutual) support.
Even positive coherence, however, may suffer from the fantasizer diffi-
culty. For systems of belief that satisfy the mutual support requirement
could be constructed through idle (though clever) fantasy. Even the positive
coherence theory, then, is not strong enough. Yet in some respects it is too
strong. Why should right J-rules require promotion of a belief system that is
wholly mutually supportive? Couldn't there be little islands of belief un-
supported by the rest of the mainland corpus? I may now be justified in be-
lieving that I have an itch in my left thigh even though this belief neither
supports nor is supported by any other belief of mine. The theme of holism
in the theory of justification has, to be sure, many attractions. But it should
not be carried too far.
Any theory of justification that neglects coherence relations entirely is li-
able to be defective. But it does not follow that coherence requirements
should be built into the criterion of rightness. M y contention is that coher-
ence is not the highest standard of justificational rightness but a derivative,
subsidiary, standard. The fundamental standard concerns the formation of
true belief. Coherence enters the picture only because coherence considera-
tions are generally helpful in promoting true belief. On the assumption that
prior beliefs are largely true, new prospective beliefs should be examined
for consistency with prior beliefs, since a necessary condition of a beliefs
being true is consistency with all other true beliefs.
I turn next to the explanatory approach. First, I need to distinguish two
senses of 'explain', a truth-implying sense and a truth-independent sense. In
the former proposition q explains proposition r only if q is true; here 'ex-
plains' has a truth-condition, like 'know'. In the latter sense there is no such
condition: q can explain r in the sense that it would be a good explanation of
r if it were true. It has all other earmarks of explanatoriness, whatever ex-
actly these are.
In the truth-implying sense of 'explain' the explanatory approach is not a
distinct rival to the verific approach. How much the two compete depends
on whether the explanatory approach holds (a) that only explanatory true
beliefs are valuable, or (b) that explanatory true beliefs are just more valu-
able than others, though not the only valuable ones. I question the correct-
ness, though, of either view.
First, cognitive operations can confer justifiedness because they detect
truths even if the truths have no explanatory value. The truth that I now
have an itch in my thigh does not explain anything else I believe; nor, per-
Justijication and Reliability 101
haps, is it explained by any of my other beliefs. Still, it is justified. (On the

reliabilist view that is because introspection is a reliable process.) Similarly,
it is doubtful whether propositions of pure mathematics or logic explain
anything; but beliefs in these propositions can still be justified.
Second, there seems to be nothing distinctively justification-conferring
about the production of explanatory beliefs. Explanatory power does not
seem to increase, or contribute to, justifiedness. This is compatible with the
view that explanation is itself an important epistemic desideratum, cer-
tainly an important part of the aim of empirical science. It does not follow
from this view that explanatoriness is critical for justijiedness. The most ex-
planatory propositions, it is generally held, are highly theoretical proposi-
tions of empirical science. This would seem to imply, on the view under
scrutiny, that inference processes leading to such theoretical beliefs are the
best at conferring justification. But this seems quite wrong. Processes like
perception or even introspection seem to be better at conferring justifica-
tion, though their outputs may have much less explanatory power. Presum-
ably, this is because they have a higher truth propensity than theoretical
inference processes.
What about the truth-independent notion of explanation? Such an ap-
proach may be favored by someone who is attracted by consequentialism
but who questions the suitability of verific values. This approach, however,
falls prey to the fantasizer counterexamples. Clever imagination can con-
struct elaborate potential explanations, perhaps even simple and elegant ex-
planations, of what we antecedently believe. But if these fanciful
constructions were consistently false, I would feel little inclination to say
that this sort of process confers justifiedness. Certainly it would not yield
the kind of justifiedness that brings us close to 'knowledge'. As in the case of
coherence, it may turn out that explanatory factors have an appropriate
place at the level of J-rules, should explanatory processes indeed be truth-
conducive. But explanatoriness should not replace verific values at the level
of the criterion.
Finally, I want to examine an approach that is a consequentialism of
sorts, but a rather different style of consequentialism than others under dis-
cussion. The approach I have in mind is epistemic decision theory, as de-
veloped by Hempel, Hintikka and Pietarinen, Levi, Hilpinen, and ~ e h r e r . "
Epistemic decision theory views belief, or 'acceptance', as an action, and
lays down a principle of rational, or justified, belief that is very similar to
the Bayesian principle of rational decision making. The theory assumes that
the cognizer has epistemic 'utilities', specifically, the verific utilities of ac-
cepting truths and avoiding falsehoods. It also assumes that the cognizer has
subjective probabilities. It proceeds to identify a rational, or justified, dox-
astic act as one that maximizes expected epistemic utility, as determined by
the verific utilities and subjective probabilities.
Since epistemic decision theory makes reference to the prospective out-

comes of a doxastic choice, it is ostensibly a kind of consequentialism. But it
is a quite different form of consequentialism from the others I have been
discussing. It makes epistemic correctness not a function of the genuine
consequences of doxastic actions, but rather a function of their expected
consequences, This is a very fundamental difference.
All other forms of consequentialism under discussion have been 'objec-
tive' consequentialisms. They all deem a rule system to be right if confor-
mity with it would really achieve certain values. This parallels the standard
example of consequentialism in ethics, namely, utilitarianism, which makes
moral rightness of actions or rules a function of the happiness and unhappi-
ness that would really result from them (and their alternatives). By contrast,
epistemic decision theory makes rightness a function of the expected conse-
quences of a doxastic action. If this is a consequentialist view at all, it should
be classed as a different variety: a 'subjective7 consequentialism.
Matters are complicated somewhat, however, because epistemic decision
theory offers a rule of rationality, or justifiedness, not a criterion. The other
theories surveyed have all been criteria for rule rightness. This complica-
tion interjects an element of incomparability among the views. I find it dif-
ficult to say what sort of criterion tacitly underlies the epistemic decision
theoretic rule. But I will not press this point.
I now offer some criticisms of epistemic decision theory, considered as an
approach to justifiedness. The first and main criticism is that it allows beliefs
to be formed on the basis of subjective probabilities without placing any
constraints on how those subjective probabilities are formed. Apparently,
the latter may be formed in any fashion at all, including hunch, fancy, and
the like. But if a belief is based on subjective probabilities of that ilk, there
are no grounds for regarding the belief as justified. It certainly is not justi-
fied in any sense that links up closely with knowledge.
Of course, epistemic decision theory could be supplemented with condi-
tions for the proper formation of subjective probabilities. These might re-
quire attention to observed frequencies, for example. But now a new theory
is required, which may have little or nothing in common with the initial,
epistemic decision motif. One then wonders whether some entirely differ-
ent criterion isn't needed-for example, an objectively verific criterion-to
fill out the approach. At a minimum, this supplement would destroy any
semblance of a unijied approach to doxastic justifiedness.
A third critical comment pertains to an element of epistemic decision
theory hitherto neglected. In characterizing epistemic 'utilities7, I men-
tioned only acceptance of truths and avoidance of error. There is another
wrinkle in the approach, however: acceptance of 'stronger' truths-truths
with greater content measures-has more utility than acceptance of
'weaker' truths.lg Aside from the difficulties of content measurement, I do
not think this is an appropriate determinant of value for the theory of justi-
fication. It may be appropriate for other epistemic terms of appraisal, but
not for justifiedness. Consider two J-rule systems R, and R2. Conformity
with R, would yield true beliefs with greater content than conformity with
R2, but in other verific respects (such as truth ratio) the rule systems are
equivalent. It is unclear to me that the processes licensed by R2 are justifi-
cationally inferior to the processes licensed by R,. Rz7sprocesses may be in-
ferior on other dimensions of appraisal, but not on the dimension of
justifiedness.
Having excluded its principal competitors, I conclude that the verific ap-
proach to justificational value-in the objective consequentialist style-is
the most promising. True belief is the value that J-rules should promote-
really promote-if they are to qualify as right. But so far, this is just a theory
of value and a choice of objective over subjective consequentialism. We still
do not have a determinate criterion of rightness.
Thus far I have narrowed the class of acceptable criteria, first to process
criteria, then to consequentialist process criteria, and finally to uerific con-
sequentialist process criteria. But this still does not pick out a unique crite-
rion. The task of this section is to narrow the class of acceptable criteria still
further. I will advocate a particular criterion as best, but also acknowledge
several related criteria as legitimate conceptions of justifiedness. Thus, in-
stead of plumping for a uniquely correct criterion, I will endorse a small
family of criteria, all in the 'truth-ratio' category.
Within the class of verific consequentialisms, there are two important
subclasses to notice: criteria invoking number of true beliefs, and criteria
invoking ratio of true beliefs. It is clear that J-rule rightness is not solely a
function of the number of truths that would be produced. The 'process' of
believing everything one can think of, including each proposition and its .
negation, might produce as large a total of true beliefs as any other process.
A criterion that required (just) maximization of the number of true beliefs
could be satisfied by a J-rule that permits such a process. But, obviously,
beliefs so formed are not justified.
This strongly suggests that what is critical to J-rule rightness is ratio of
true beliefs, not total number. The trouble with the foregoing process (as-
suming it is psychologically feasible, which is doubtful) is that its ratio of
true beliefs to total beliefs would only be .SO.These considerations prompt a
reliability approach to justifiedness. A justificationally permitted process must
be one that yields a high truth ratio, higher-perhaps appreciably higher
-than .50. Further support for reliabilism comes from a survey of cases.
Among the prototypical cases of justification-conferring processes are: (1)
forming beliefs by standard perceptual processes, (2) forming or retaining

beliefs by memory, and (3) certain patterns of deductive and inductive rea-
soning. Among the prototypical cases of non-justification-conferring pro-
cesses are: (1) wishful thinking, (2) sheer hunch or guesswork, and (3) failure
to take account of all one's relevant evidence (or failure to take account of
obviously relevant evidence). What the former processes seem to share is a
reasonably high truth ratio. What the latter processes seem to share is a
low(er) truth ratio.
Support for reliabilism is bolstered by reflecting on degrees of justified-
ness. Talk of justifiedness commonly distinguishes different grades of justi-
fiedness: 'fully' justified, 'somewhat' justified, 'slightly' justified, and the
like." These distinctions appear to be neatly correlated with degrees of relia-
bility of belief-forming processes. Not all perceptual processes or memory
processes, for example, confer equal grades of justification. A belief based
on fleeting or careless perceptual exposure is generally less justified than a
belief based on ample and attentive perceptual examination. A belief based
on a hazy recollection is less justified than one involving vivid memory.
These intuitions of justificational strength are traceable to truth ratio. Per-
ceptual processes featuring more extensive, leisurely observation tend to
have higher truth ratios. Vivid recollections yield a higher truth ratio than
hazy ones. Thus, what seeins critical is the ratio of truths produced by a pro-
cess, not the total number produced.21
Even within the class of ratio criteria, there are choices to be made. Let
us first distinguish resource-independentand resource-relatiue criteria. A re-
source-independent criterion fixes an acceptable truth ratio without regard
to the resources of the (type of) cognitive system in question. A resource-rel-
ative criterion fixes an acceptable truth ratio as a function of the target
cognitive system's resources, in the present case, human cognitive
resources.
Resource-relative criteria are likely to be comparative. One such crite-
rion is a maximizing criterion. It would say that a rule system is right if and
only if the processes it permits would maximize truth ratio, that is, maxi-
mize relative to humanly available processes. This is what some writers call
a criterion of "doing the best one can."22 A different sort of resource-rela-
tive criterion is a satisficing criterion. For example, a criterion could say
that a rule system is right just in case the processes it permits have a truth
ratio of a level humans commonly attain, or that it is easy for them to attain.
A resource-independent criterion is liable to be absolute rather than
comparative. It would set an absolute truth ratio that must be met, whether
or not human resources are capable of meeting it. Such a criterion might
say that a rule system is right only if its realization would produce a truth
ratio of at least .90.Given such an absolute, resource-independent criterion,
it becomes an open question whether any system of J-rules is right for
human beings. Human beings might possess no set of basic cognitive pro-
cesses that allow the criterion to be satisfied. Less dramatically, such a crite-
rion might exclude belief permission rules for one class of processes, for
example, inductive processes, though not exclude belief permission rules al-
together. Human perceptual processes might be capable of generating be-
liefs with a truth ratio of .90, but perhaps no human inductive processes
could generate such a high truth ratio.
Notice that a maximizing, resource-relative criterion could have implica-
tions that are either stronger or weaker than some specific resource-inde-
pendent criterion. Take an (absolute) resource-independent criterion that
sets the acceptable truth ratio at .80. If it turns out that human resources
include processes capable of a truth ratio higher than .80, then a maximiz-
ing, resource-relative criterion would require (at least) that higher ratio.
But if human resources are weaker, a maximizing resource-relative criterion
might only require a truth ratio of, say, .70. In principle such a criterion
might even count a rule system as right with a truth ratio less than .50.
Is a criterion that permits a truth ratio of less than .50 a 'reliabilist' crite-
rion? Probably not. This means that not all the criteria that fall into the
truth-ratio category qualify as reliabilist criteria. Perhaps only the absolute,
resource-independent criteria should be so labeled. In that case among the
family of criteria that I am prepared to count as legitimate-that is, as cap-
turing s o n e intuitively satisfying conception of justifiednes-not all are re-
liabilist criteria. They are, however, all truth-ratio criteria.
The satisficing, resource-relative approach has its attractions. In many
domains we make evaluations based on what the 'common man' does, or
what it is reasonable to expect people to do given their capacities. Plausi-
bly, this is what we (sometimes) do in appraising belief formation too. We
expect people to use better, more reliable, processes as long as they are rela-
tively easy, or natural, processes. We do not expect people to use very diffi-
cult, or sophisticated, cognitive process sequences, even if these could
increase the truth ratio. Omission of such processes does not preclude justi-
fied belief, on one plausible conception of justifiedness. In Part I1 I will have
several occasions to distinguish the relative difficulty of cognitive feats and
to link judgments of justifiedness to these distinctions. The satisficing, re-
source-relative criterion would neatly accommodate such judgments.
All three of the aforementioned approaches have their merits. In most of
what follows, though, I shall opt for an absolute, resource-independent cri-
terion of justifiedness. One virtue of this approach is that it makes the chal-
lenge of skepticism both serious and credible. If some fixed level of truth
ratio is necessary for justifiedness, it is an open question whether human
processes ever yield justified belief, whether they are even capable of yield-
ing justified beliefs. By contrast, the maximizing, resource-dependent crite-
rion guarantees that some human processes can yield justified beliefs. And
the satisficing, resource-relative approach also makes it easy to have justi-

fied beliefs.
There is room for disagreement, of course, on the question of how
seriously global skepticism should be taken. Some epistemologists might
favor the resource-relative approaches precisely because they soften, or
eliminate, the threat of (global) skepticism. But I am inclined to take skep-
ticism as a serious challenge, so I would not back away from the absolute,
resource-independent approach on that account. Another virtue of this ap-
proach is that it effects an intimate thematic tie between accounts given
here of justification and knowledge. Recall that the global reliability re-
quirement was central to the theory of knowing (Chapter 3). Of course,
knowledge has other requirements not needed for justifiedness. Unlike a
piece of knowledge, a justified belief need not be true. And the local relia-
bility requirement for knowledge ('no relevant alternatives') is not imposed
for justifiedness. But otherwise, the two concepts turn out to be very close.
Let me now formulate the absolute, resource-independent criterion of
justifiedness:
(ARI) A J-rule system R is right if and only if
R permits certain (basic) psychological processes, and the instan-
tiation of these processes would result in a truth ratio of beliefs
that meets some specified high threshold (greater than 50).
Since (ARI) does not designate any particular threshold, it is really a crite-
rion-schema. A determinate criterion would be fixed by choice of a thresh-
old parameter. However, I am content to leave the theory with this degree
of vagueness, since the ordinary concept of justifiedness is similarly vague.
Before the theory could actually be applied, however, a specific threshold
value would have to be chosen. One cannot decide whether a particular J-
rule system is right--even if one knows its truth ratio-until a required
truth ratio is selected.23
I will not formulate here the two resource-relative criteria, but their for-
mulation should be fairly straightforward. Although I will have occasion to
refer to them from time to time, (ARI) is the main criterion(-schema) I shall
utilize, one which is obviously reliabilist in character.
At this point the following question arises: Is the rightness of a rule sys-
tem determined by its truth ratio in the actual world, and in that world
only? Or should the performance of the rule system also be judged by its
performance in other possible worlds? Or is a still different performance
measure appropriate? Obviously, a given rule system could perform well in
one possible world-say the actual world-and poorly in another. Which
possible worlds are relevant to the rightness of a rule system, and ultimately
to the justifiedness of a belief formed in compliance with the system?
Before addressing this issue, we should ponder the following related
]ustijication and Reliability 107
question. Is a J-rule system that is right in one possible world also right in all
possible worlds? In other words, is rightness a 'rigid' designator? Or can a
system vary in rightness across different possible worlds? Rightness is easily
relativized to a possible world by saying that, for any world W, a system is
right in W just in case it has a sufficiently high truth ratio in W. A different
theory would be chauvinistic toward the actual world. It might say that a
rule system is right in any world W just in case it is sufficiently reliable in
the actual world. This kind of actual-world chauvinism would of course
imply that each rule system has the same rightness status in all possible
worlds.
I think that neither of these answers is correct. At any rate, neither seems
to capture our intuitions about justifiedness in hard cases. The answer that
best accords with our intuitions-as will emerge in the next section-runs
as follows. We have a large set of common beliefs about the actual world:
general beliefs about the sorts of objects, events, and changes that occur in
it. We have beliefs about the kinds of things that, realistically, do and can
happen. Our beliefs on this score generate what I shall call the set of n o m l
worlds. These are worlds consistent with our general beliefs about the ac-
tual world. (I emphasize 'general', since I count worlds with different par-
ticular episodes and individuals as normal.) Our concept of justification is
constructed against the backdrop of such a set of normal worlds. My pro-
posal is that, according to our ordinary conception of justifiedness, a rule
system is right in any world W just in case it has a suffjcientlyhigh truth
ratio in n o m l worlds.24 Rightness is rigidified for all worlds; but it is
rigidified as a function of reliability in normal worlds, not reliability in
the actual world. Rightness of rules-and hence justifiedness-displays
normal-world chauvinism.
Obviously, the notion of normal worlds is quite vague. But I do not find
that objectionable. I think our ordinary notion of justifiedness is vague on
just this point. But let me expand on the approach by putting it in a more
general, nonepistemological, context.
The idea of actual-world chauvinism is related to the causal theory of
names and natural kind terms proposed by Saul Kripke and Hilary Put-
nam.25On this view the reference and/or meaning of a term is derived from
the actual sample of objects to which the term was applied in some original
baptismal acts, or meaning-giving contexts. Words like 'gold' or 'catY,for
example, acquire reference and/or meaning from the class of things to
which they were originally applied. Objects sharing the same nature or es-
sence as that sample, or class, are also gold pieces, or cats. Without review-
ing the many discussions of this theory, let me turn to a problem case raised
by Peter Unger, a case which also provides rationale for normal-world
~hauvinism.~
In an early paper Putnam introduced an example in which the things we
have always taken to be cats have really been inanimate robots, placed on
earth by Martian scientist^.^' Unger now introduces the case of feline robots
in the past and feline animals in the future. Starting now, the Martian scien-
tists (unbeknownst to us) replace all the feline robots with feline animals.
According to the causal theory, the feline animals would not correctly be
called 'cats'. The theory implies that 'cats' would refer to the nature or es-
sence of the things originally dubbed 'cats', namely, feline robots. Since the
feline animals do not share that nature or essence, they are not cats. But this
is intuitively incorrect. Surely the feline animals would be cats.
Furthermore, I find it intuitively wrong to say that the feline robots are
cats, even though they are what humans in the example have been calling
'cats' all along. This suggests that the reference and/or meaning of 'cats' is
fixed not by the actual nature of objects so called, but by their presumed
nature. What we believe about the things we call 'cats' determines the
meaning.
Similarly, I suggest that the meaning of the term 'justified' (in its episte-
mic use) is fixed by certain things we presume about the world, whether we
are right or not. Specifically, beliefs are deemed justified when (roughly)
they are caused by processes that are reliable in the world as it is presumed
to be. Justification-conferring processes are ones that would be reliable in
worlds like the presumptively actual world, that is, in normal worlds.
Let me trace an implication of this theory. Suppose it turned out that the
actual world is not as we think it is. Suppose it is a Cartesian demon-style
world, in which we are radically deceived about 'external' things (though
we are not deceived about the nature of our cognitive processes). This
means that our cognitive processes are not reliable in the actual world. At
least our perceptual processes are unreliable, and perhaps our inductive
processes as well. It does not follow, on my theory, that our actual-world
beliefs are uniustijied. The cognitive processes we use may still be reliable
in n o m l worlds, in worlds similar to the presumed actual world. This pos-
sibility follows from normal-world chauvinism vis-a-vis J-rule rightness.
I do not suppose that the fundamental world regularities that define the
class of normal worlds extend to properties of our own cognitive processes.
Thus, what we believe about our cognitive processes in the actual world
need not hold in (all) normal worlds. For this reason, my proposal does not
imply that the processes believed to be reliable (in the actual world) are re-
liable in normal worlds. Consequently, the processes we believe to be justi-
fication-conferring need not really be justification-conferring. We could be
mistaken in thinking that certain cognitive processes of ours are reliable in
normal worlds. For example, we may presume that our processes of esti-
mating the chances of future events are quite accurate, but we may be quite
mistaken about this. Hence, it is still a contingent and empirical question
whether the processes we take to be justification-conferring really are so;
indeed, it is left open whether there are any justification-conferring pro-

cesses available to humans at all. Thus, normal-world chauvinism does not
beg the question of skepticism.
Does the current stratagem turn my theory of justifiedness into a subjec-
tivist account rather than an objectivist one? I think not, for the reason just
indicated. Whether a given rule system is right does not simply depend on
what we believe about the world. It depends on whether the processes per-
mitted by the specified rule system really do have a high enough truth
ratio in normal worlds. This is something that hinges on the 'facts' about
these processes, not simply on our opinions. Hence, the proposal is objecti-
vist, not subjectivist, in contour.
One final point. My proposal to judge reliability by reference to normal
worlds is made in the spirit of trying to elucidate the ordinary conception of
justifiedness. I am prepared to be persuaded that this ordinary conception
can be improved upon. I would lend a receptive ear to proposals to 'regi-
ment' the concept of justifiedness so as to judge rightness by reliability in
the actual world, or by reliability in the possible world of the b,elief in
question. Either of these approaches might seem preferable from a system-
atic or theoretical point of view. Nonetheless, they do not seem to be what
is implied by the ordinary conception as it stands; and that is all I am cur-
rently trying to capture.
5.6. Counterexamples and Replies

Previous versions of reliabilism have been the objects of a number of coun-
terexamples and criticisms. Let us look at the most prominent of these, to
see how the present version of reliabilism fares.
Reliabilism may be inadequate for either of two reasons: it may not be
sufficient for justifiedness, or it may not be necessary. One charge of insuffi-
ciency comes from ~utnam." Suppose, Putnam says, that the Dalai Lama is
in fact infallible on matters of faith and morals. Then anyone who believes in
the Dalai Lama, and who invariably believes any statement the Dalai Lama
makes on a matter of faith or morals, uses a method which is 100 percent
reliable. Such a person's beliefs in matters of faith and morals should a11 be
justified, according to reliabilism, even if his argument for the belief that
the DaIai Lama is never wrong is simply "the Dalai Lama says so." But ob-
viously these beliefs are not justified.
A second set of arguments for insufficiency is presented by various au-
thors, including Laurence ~ o n ~ o u rBonJour
.~' presents four cases, in all of
which the cognizer has a completely reliable clairvoyant power. According
to earlier versions of reliability theories of justifiedness, the clairvoyance-
caused beliefs of these cognizers should be j~stified.~'BonJour makes it
plausible that his four cognizers' beliefs are not justified. In some cases the
11 0 Justijication and Reliability
cognizer has explicit counterevidence against his or her possession of a

clairvoyant power but ignores it. In another case the cognizer has no rea-
sons to believe that he or she is clairvoyant but has counterevidence against
the clairvoyance-caused belief, namely, that the President is in New York.
The cognizer persists in believing that the President is in New York in the
face of the counterevidence. And so on for the other cases.
A third sort of counterexample purports to show that reliability is not
necessary for justifiedness. Consider a possible world in which a Cartesian
demon systematically deceives a certain cognizer (or lots of cognizers). The
cognizer employs the very same psychological processes that you or P use,
but the result is a massively false set of beliefs. Since his processes are not
reliable, reliabilism implies that his beliefs are not justified. But intuitively
they are j ~ s t i f i e d . ~ ~
Let us start with Putnam's counterexample. Notice, first, that the method
of believing whatever the Dalai Lama says on matters of faith and morals is
just that, a method, in the slightly technical sense I have given that term.
But my theory does not say that being produced by a reliable method is suf-
ficient for justifiedness. For one thing, I have not really offered a theory of
secondary justifiedness, although I do lean in the direction of a reliability
theory thereof. But let us set that issue aside. I have committed myself to
the view that, however reliable or right a method may be, it does not confer
justifiedness unless it is acquired by suitable processes, for example, metare-
liable processes (see section 5.7). Now this condition is not clearly met in
Putnam's example. Putnam says that the person's only argument for the be-
lief that the Dalai Lama is never wrong is simply "the Dalai Lama says so."
But what are we to infer from that? The person's only defense of the
method seems to be another application of the method. But this still leaves
open the question of what processes have led him to acceptance of this
method. Pending specification of some legitimate (metareliable) psychologi-
cal process, one is left with the suspicion that no metareliable process has
been used, that the believer has simply stumbled upon a method which, by
chance, is infallible. In that case the belief is not justified, on my theory. If it
turns out that metareliable processes were used in acquiring the method,
then I think that the person is justified in believing the method to be reli-
able, and hence the beliefs that issue from it may be justified. (Even this re-
sult, though, will not suffice for justifiedness on my theory. The method
itself must be certified as right. Under a culturally relative criterion of
rightness, trusting the Dalai Lama may not be a right method, even if it is in
fact infallible; for the relevant culture might not approve of such a method.)
Before turning to BonJour's examples, which are rather like the Millicent
case introduced in Chapter 3, let us see how my theory handles the Milli-
cent case. Millicent has perfectly normal and (we assume) reliable visual
powers. She also has a false though justified belief that her visual powers are
Justijicatiorr and Reliability 11 1
seriously impaired. Despite this belief, Millicent forms other beliefs that
endorse the testimony of her visual appearances. Intuitively, these beliefs
are not justified. But doesn't our reliability theory imply that they are justi-
fied? No. To appreciate this, we must simply recall the nonundermining
clause of our framework principle (P3*).The nonundermining provision is
violated in the Millicent case; at least that is how the provision should be
understood.
Recall the initial discussion of permission undermining in section 4.1. A
sufficient condition for undermining was suggested there: the cognizer be-
lieves--justifiably or not-that he or she is not permitted to hold the target
belief. Now Millicent does not fulfill this condition, we may suppose. Let us
say she does not possess the concept of permission by a right J-rule system,
and therefore possesses no beliefs featuring this concept. What is true of
Millicent, though, is the following. She believes that her visual powers are
impaired, so that the belief-forming processes that utilize her visual powers
must not be reliable. What she believes, then, is such that if it were true,
the beliefs in question (her visually formed beliefs) would not be permitted
by a right rule system. Satisfaction of this condition, I now propose, is suffi-
cient to undermine permittedness. In other words, it is sufficient for under-
mining that a cognizer believes that certain conditions obtain which, if they
did obtain, would entail that the target beliefs are not permitted by a right
rule system. The cognizer need not actually have any beliefs about rule sys-
tems, rightness, or criteria of rightness.
Millicent's visually formed beliefs are permitted by a right J-rule system.
But according to my framework principle, the beliefs still are not justified,
since their being so permitted is undermined. This yields the result needed
for handling the knowledge question: Millicent's visually formed beliefs are
not pieces of knowledge because they are not justified. And they are not
justified despite the fact that the producing processes are indeed reliable.
Let us now turn to BonJour's examples. A difference between BonJour's
cases and the Millicent case is that the characters in BonJour's examples do
not believe they are not using reliable processes. However, undermining
should be so construed that the permittedness of their target beliefs is un-
dermined nonetheless. For although they do not in fact believe that the tar-
get beliefs are unreliably caused, they are justified in so believing (in a sense
to be specified below).
Let us look at these cases in detail. Maud has a completely reliable clair-
voyance power, and uses it in forming a belief N that the President is in
New York. Maud believes herself to have the power of clairvoyance. But
Maud also has been inundated with massive quantities of apparently cogent
scientific evidence that no such power is possible. Intuitively, then, Maud is
justified in believing that she has no reliable clairvoyance power, although
she does not believe this. And she is justified in believing that her belief in N
is not reliably caused, although again she does not believe this. Since she is
justified in believing the latter, however, that should suffice to undermine
the permission of her believing N. (Use of her clairvoyant powers would be
permitted by a right rule system because they are reliable.)
To make this proposed treatment stick, I need to go a bit beyond the the-
ory of justifiedness offered thus far. The account I have offered is a theory of
when an actual belief is justified. It is a theory of what I call (in "What Is
Justified Belief?") ex post justifiedness. But the present undermining deals
with being justified in believing a proposition one does not actually believe.
The idea is that it is a proposition one would (or could) be justified in be-
lieving, given one's present cognitive state, although one does not in fact
believe it. This is what I call ex ante justifiedness. A theory of ex ante justi-
fiedness can be constructed, I think, along lines very similar to my account
of ex post justifiedness, though I will not pursue all the details. One differ-
ence is that we may have to require right rule systems to feature obligation
rules as well as permission rules. Thus, we might suppose that a right rule
system would require Maud to utilize certain reasoning processes, processes
that would lead her from the scientific evidence she possesses to belief in
the proposition that she does not have any reliable clairvoyance power. And
similarly, she would be required to infer that her belief in N is not caused by
any reliable processes. On the proposed account she would be justified (ex
ante) in believing that her belief in N is not reliably caused. And this suffices
to undermine the permittedness of her belief in N.
BonJour's case of Casper would be handled similarly. Casper has at-
tempted to confirm his allegedly clairvoyant beliefs on numerous occasions,
but they have always turned out apparently to be false. Given this evidence,
he ought to use processes that would lead him to believe (permissibly) that
he lacks reliable clairvoyant processes. This undermines the actual permissi-
bility of his belief in N.
BonJour's next case, the case of Norman, is arguably a little tougher, and
indeed it is introduced to show that even a nonundermining condition still
does not produce an adequate reliabilist theory. BonJour describes this case
as one in which Norman possesses no evidence or reasons of any kind for or
against the general possibility of clairvoyance, or for or against the thesis
that he possesses it. But it is hard to envisage this description holding. Nor-
man ought to reason along the following lines: 'If I had a clairvoyant power,
I would surely find some evidence for this. I would find myself believing
things in otherwise inexplicable ways, and when these things were checked
by other reliable processes, they would usually check out positively. Since I
lack any such signs, I apparently do not possess reliable clairvoyant pro-
cesses.' Since Norman ought to reason this way, he is ex ante justified in be-
lieving that he does not possess reliable clairvoyant processes. This
undermines his belief in N. Thus, the nonundermining clause of (P3") han-
dles BonJour's cases.
Let us turn now to the Cartesian demon example. This case is intended to
show that reliability is not necessary for justifiedness. Intuitively, a cognizer
being massively deceived by an evil demon is still forming justified beliefs.
But in an evil demon possible world, the cognitive processes being utilized
just are not reliable. So reliability is not a necessary condition for
justifiedness.
The solution to this counterexample is straightforward. The counterex-
ample rests on the assumption that the justificational status of a belief in
possible world W is determined by the reliability of the beliefs processes in
W. But this is what I have denied (in the previous section). The justifica-
tional status of a W-world belief does not depend on the reliability of the
causing processes in W. Rather, it depends on the reliability of the processes
in normal worlds. Now an evil demon world is a paradigm case of a non-
normal world. So it does not matter that the processes in question are highly
unreliable in that world. It only matters whether they are reliable in normal
worlds, and that is apparently the case. (At any rate, that is what our intui-
tive judgment of justifiedness presumes, I think.)
The same treatment of the putative counterexample applies if the coun-
terexample is revised to make it an actual-world case. Imagine, the objec-
tion goes, that our actual world turns out to be an evil demon world. (Or
imagine we are actually brains in a vat being deceived by scheming scien-
tists.) Intuitively, our beliefs would still be justified; yet the belief-forming
processes being deployed are not reliable. Again the case is easy to handle.
Its apparent strength rests on the assumption that the justificational status
of the beliefs is determined by the reliability of their causal processes in the
actual world. But this does not accord with our theory. Reliability is mea-
sured in normal worIds; and in this case, the actual world is an abnormal
world! So reliability in the actual world just does not matter.
The reader can now see how the normal-worlds account of reliability fits
our intuitions very naturally in the evil demon class of cases. Indeed, reflec-
tion on these cases (partly) prompted the normal-worlds approach. Once
formulated, though, it seems to work naturally for other cases as well.32
5.7. Calibration and Second-Order Processes

This section picks up several loose ends in my theory of justification. I take
up, briefly, three topics. (1) I suggest a way to generalize the truth ratio ap-
proach from categorical to partial beliefs (subjective probabilities). (2) I in-
dicate why rightness has been defined in the first instance for rule systems
rather than individual rules. (3) I discuss rightness criteria for second-order
processes.
First, note that the theory presented thus far addresses only the justifica-
tional status of binary belief. It is generally assumed, however, that there
are degrees of belief, that people believe propositions with more or less
1 14 Justification and Reliability
confidence; and it is often claimed that confidence levels can be scaled on

the unit interval. Many theorists therefore seek an account of epistemic
norms for these different confidence degrees or subjective probabilities. Can
the reliabilist framework accommodate such a refinement?
It certainly can, as I shall outline below. Before proceeding, though, let
me register my reservations about the psychological reality of degrees of
belief. Although there is unquestionably some (introspective) basis for the
notion of degrees of confidence, I question whether the psychological facts
warrant a systematic treatment of belief states as subjective probability
states. Grounds for these doubts will be scouted in Chapter 15. Meanwhile,
I will make the working assumption that the psychological issues can be re-
solved. I will try to show how truth ratio consequentialism is a well-
equipped framework for doing the epistemology of partial beliefs. Doubts I
harbor about the psychological tenability of this construct will be deferred
until Chapter 15.
What I want to do, then, is generalize the truth-ratio (ARI) criterion of
rightness, so that it applies to processes that generate partial beliefs. I as-
sume that partial beliefs can take any real value in the interval from 0 to 1.
Very simply, the proposal is that a J-rule system is right if and only if it is
'well calibrated'.
The notion of calibration is presented and defined by Sarah Lichtenstein,
Baruch Fischhoff, and Lawrence Phillips as follows: "A judge is calibrated
if, over the long run, for all propositions assigned a given probability, the
proportion that is true equals the probability assigned."33 Application of
this definition to J-rule systems would run as follows. First, assume that in-
stantiations of (that is, compliance with) a given J-rule system R would pro-
duce some partial beliefs. (I am interested in instantiations of R in normal
worlds, but I will focus on the actual world as presumably one such world.)
Now among the partial beliefs so produced, let those of the same degree be
grouped together: all of degree .SO, all of degree .70, and so on. Then a sys-
tem R is well calibrated if and only if, for each degree D, the truth ratio of
propositions believed to degree D is approximately D. For example, ap-
proximately half of the propositions at .SO subjective probability are true,
approximately 70 percent of the propositions at .70 subjective probability
are true, and so on. Finally, a J-rule system is right if and only if it is well
calibrated.
Given this rightness criterion, the framework principle for justification,
(P3*),is directly applied. A token partial belief is justified if and only if (a)it
is the result of a sequence of (basic) processes permitted by a right rule sys-
tem, and (b) the permittedness of this sequence is not undermined (for the
cognizer in question).
Good calibration as an epistemic norm is not without its problems.34
These would have to be resolved before the good calibration approach
could be finally accepted. But I am only interested here in outlining a

good-calibration approach to justification, and showing how the truth-ratio
conception readily lends itself to a related treatment of justification for
partial beliefs.
The second loose end I wish to pick up in this section concerns the choice
of rule systems, as opposed to individual rules, as bearers of justificational
rightness. The rationale for this need not occupy us at length. The point is
that rules are interdependent with respect to their epistemically relevant
properties. In particular, they are interdependent with respect to truth-ratio
properties. This is especially clear for inferential rules. A rule of 'detach-
ment', for example, might not spell trouble if complemented by certain
rules, but will breed what is known as the lottery paradox if complemented
by other rules. A sound inferential rule will generate additional true beliefs
when applied to true input beliefs. But if other rules permit false beliefs to
be formed, then even a sound inferential rule may produce innumerable
errors. Truth-ratio propensities, then, only make sense as applied to rule
systems, not isolated rules. A similar point holds, I suspect, for different
epistemically relevant properties that other theories might advocate.
The third loose end I wish to pick up concerns second-order processes. In
section 5.3 I noted that a complete theory of primary justifiedness should
include a rightness criterion for second-order as well as first-order processes.
It is now time to adumbrate what this criterion should be.
Given my truth-ratio approach to first-order rightness, the natural expec-
tation is a truth-ratio approach to second-order rightness as well. This in-
deed strikes me as promising; but there are several permutations to
consider. Two of these were suggested in section 1.4, where I first intro-
duced the notion of second-order processes. A second-order process might
be considered metareliable if, among the methods (or first-order processes)
it outputs, the ratio of those that are reliable meets some specified level,
presumably greater than SO. For example, suppose a threshold of .80 is set
for both reliability and metareliability. Then a second-order process is me-
tareliable if and only if at least 80 percent of the processes (or methods) it
produces are at least 80 percent reliable.
A weaker conception of metareliability would merely require increases
in reliability, not an absolute ratio of reliability. On this conception a sec-
ond-order process would be metareliable if it modifies or replaces processes
(or methods) so as always to increase levels of reliability. Even more weakly,
a second-order process might be metareliable if the modifications or re-
placements it produces usually (rather than always) increase reliability.
In addition to these possibilities, there are possibilities obtainable by
adapting either of the resource-relative conceptions of J-rightness intro-
duced in section 5.5. Corresponding to the maximizing resource-relative
criterion would be a criterion of second-order rightness that requires use of
1 16 Justification and Reliability
second-order processes with maximum metareliability (for humans). Corre-

sponding to the satisficing resource-relative criterion would be a criterion of
second-order rightness that merely requires a level of metareliability that is
easily, or commonly, achieved. Each of these criteria can be combined with
either of the themes examined above: (a) the proportion of reliable pro-
cesses (or methods) that get produced, or (b) the tendency to increase the
reliability of one's repertoire.
It is worth emphasizing that second-order processes should be judged not
simply by their acquisition performance, but also by their retention per-
formance. Even if a set of second-order processes frequently acquires poor
methods, this will not be so bad if the poor methods are weeded out quickly.
But if poor methods are retained over considerable periods of time, it is a
much more serious deficiency. In assessing the quality of second-order pro-
cesses, retention (and modification) characteristics should be weighed as
heavily as acquisition characteristics.
I will not try to choose a unique criterion from among the candidates
sketched above. I will rest content with this family of criteria (though I will
return to the topic in Chapter 17, where further complications will be in-
troduced). I suspect that our ordinary conception of justifiedness is simply
quite vague about the constraints on method acquisition and retention, so
no unique criterion is saliently indicated. In any case these candidate cri-
teria all confirm the notion that truth ratios (or improvements in truth
ratios) are central to second-order rightness.
This notion is further confirmed if we turn our attention from second-
order processes for categorical belief-forming processes to second-order
processes for partial belief-forming processes. To complement our good-
calibration criterion of rightness for partial belief rules, the natural sugges-
tion is that second-order processes are permissible if they usually produce
reasonably well-calibrated processes (or methods); or at least improve cali-
bration by managing, over time, to discard poorly calibrated processes.
5.8. Truth-Linkedness
The rightness criteria I have proposed are all verific consequentialist cri-
teria, and these are a species of truth-linked criteria. Many people a.re suspi-
cious of such criteria. There are charges of circularity and worries about the
difficulty of applying a truth-linked criterion. This section is intended to an-
swer the charges and allay the worries.
Let us begin with a charge of definitional circularity. This charge (in
slightly different words) is leveled by Putnam:
Truth, in the only sense in which we have a vital and working notion of it, is
rational acceptability . . . But to substitute this characterization of truth into
the formula 'reason is a capacity for discovering truths' is to see the empti-
ness of that formula at once: 'reason is a capacity for discovering what is . . .
Justij5cation and Reliability 11 7
rationally acceptable' is not the most informative statement a philosopher

might utter. The . . . epistemologist must either presuppose a 'realist' . . . no-
tion of truth or see his formula collapse into vacuity,35
If truth is indeed rational acceptability, Putnam is right in claiming that a
reliabilist-or more generally, a truth-linked-account of justifiedness is
circular (or "vacuous"). I therefore need to show that truth is not rational
acceptability. In short, I need to argue for a realist theory of truth. This task
will be undertaken in Chapter 7. Until then, I simply presuppose a realist
understanding of truth.
A different charge of circularity is posed by Roderick Firth.
Each of us decides at any time t whether a belief is true, in precisely the
same way that we would decide at t whether we ourselves are, or would be,
warranted at t in having that belief. . . If we are rational we must assume . . .
a correlation between warrant-conferring rules and true beliefs . . . in order
to identify true beliefs. Our reasoning is obviously circular if we then use
beliefs so identified as data for inferring a correlation between warrant-con-
ferring rules and true beliefsM
The worry here (as indicated by the final sentence of the passage) seems to
be directed at what we have been calling a reliabilist 'criterion' of justified-
ness. Firth appears to be charging that in order to substantiate that criterion
we must presuppose it. This charge rests on a misplaced assumption. It is
apparently assumed that we can only substantiate the criterion by general-
ization from observed 'data'. We observe warranted beliefs, we observe
their truth, and we infer a correlation between the two. This picture of cri-
terion substantiation is misplaced. We do not have to observe warranted
beliefs and determine their actual truth-values to decide whether the crite-
rion is correct. As in other cases of analysis, one can proceed entirely by hy-
pothetical cases. We can simply ask whether an imagined belief caused by
unreliable processes would count as warranted. We use hypothetical cases
to elicit our intuitions. The procedure does not require us to determine the
actual truth-values of actual beliefs.
Another passage from Firth raises the specter of a different sort of circu-
larity. He writes:
[If by] 'warrant-conferring rules' . . . we mean statistically reliable rules . . . is
there any way to decide rationally whether or not a particular set of rules. . .
is or is not warrant-conferring?To decide this, which involves predicting the
future, would require an inductive argument based on the past failure or suc-
cess of those rules in generating true beliefs. But if we are rational, how can
we identify true beliefs except by assuming, explicitly or implicitly, that
some particular set of rules is warrant-conferringP7
This objection concerns not the choice of a criterion, but the determi-
nation of which rule-system (if any) satisfiesthe criterion. The objection is
that in the attempt to decide which rule system satisfies the criterion, one
must already assume some system as right. Some system must be presup-
posed in order to form beliefs at all, including beliefs about which system
satisfies the criterion's requirements. This is apparently viewed as objec-
tionable on grounds of circularity.
Does the suggestion that a rule-system must be assumed, or presupposed,
mean that one must already have beliefs about what rule system is right?
That is not so. One need only use cognitive processes (and perhaps methods)
to decide whether a target rule system satisfies the criterion. Using cogni-
tive processes does not in general-nor in this case--require beliefs in the
epistemic properties of rule systems.
Firth is also wrong in suggesting that application of a reliabilist criterion
requires prediction of the future. One can identify a truth propensity in
normal worlds-at least an approximate truth propensity-without worry-
ing about the exact course of events in the future segment of the actual
world. Consider, for example, assessment of the truth propensity of visual
mechanisms. By studying their underlying structure, one can determine
that they would breed illusory impressions in a certain class of stimulus
conditions. If one has a rough idea of the frequency of this class of condi-
tions in normal situations, one can assess the degree of unreliability this
property engenders. Some inductive reasoning is involved in the study of
any mechanism's properties, but no prediction of individual future events is
required.
The next point to notice-and emphasize-is that truth-linked criteria
are no worse off than any other criteria in the matter of application. To
apply any sort of criterion, some reasoning and other cognitive processes
are required. How else could one decide whether a particular rule-system
satisfies a given criterion? This fact does not reflect adwersely on truth-
linked criteria as compared with other criteria.
Perhaps the intended challenge is not whether a truth-linked criterion
can be applied at all, but whether it can be applied either (a) correctly, or
(b)rationally (or justifiably). Firth's formulation concerns (b), but let us start
with (a).
First, again note that the issue does not reflect adversely on truth-linked
criteria as compared with other criteria. There can always be a question of
whether antecedent cognitive processes or methods will work correctly in
trying to apply a criterion. Consider a criterion that says that a rule system
is right if and only if conformity with it would maximize 'positive coher-
ence'. An attempt to apply such a criterion would first involve an assess-
ment of just how much coherence would be produced by each of a number
of specified rule systems (given some measure of positive coherence). Next
one would have to determine that all possible rule systems had been sur-
veyed, at least all strong candidates. Clearly, all sorts of mistakes might be
made in this selection process. Even much weaker criteria would feature
Justijkation and Reliability 119
the possibility of mistaken application. Suppose a criterion says that a rule

system is right if and only if it guarantees a consistent belief corpus. Appli-
cation of this criterion is also tricky. Even famous logicians have made mis-
takes about the consistency of certain formal systems. It is no trivial matter
to determine whether any given rule system ensures consistency.
It might be argued that some criteria would be easier to apply correctly
than truth-linked criteria. Some criteria would not require empirical inves-
tigation of psychological mechanisms, and this would alleviate some of the
difficulties posed by (ARI), for example. But notice that the cited difficulty
does arise not from the truth-linked feature of (ARI), but rather from its ref-
erence to basic psychological processes. That feature of a satisfactory crite-
rion is independent of truth-linkedness. Second, notice that in a certain
sense, truth assessments are involved in application of all criteria. For ex-
ample, to decide whether a candidate system R satisfies a coherence maxi-
mization criterion one has to decide whether it is true that R maximizes
coherence.
Third, even if it is granted that truth-linked criteria are harder to apply
than others, what reason is there to suppose that relative ease of application
must be a property of a correct criterion? No argument has been given to
this effect. (This might be a plausible constraint if the criterion and tule
systems were intended to be regulative, or action-guiding. But they are not
so intended.)
I offer no guarantee that (ARI), or any of the other truth-ratio criteria
proposed here, can be correctly applied. But as we have seen, no such guar-
antee can be offered for other criteria either. Still, it may be replied, can't
we at least provide a guarantee that it is possible for truth-ratio criteria to
be applied correctly? Surely such a minimal guarantee ought to be forth-
coming for any adequate criterion.
At this point we need to distinguish logical possibility from factual possi-
bility. We can certainly guarantee that it is logically possible for truth-ratio
criteria to be correctly applied. It is logically possible that our native pro-
cesses, together with our antecedently accepted scientific methods, should
lead us to identify one or more rule systems that really do satisfy (ARI) or
my other truth-ratio criteria. When it comes to factual possibility, though,
the matter is open to doubt. It is conceivable that our actual cognitive re-
sources are so poor, and our currently accepted scientific methods so weak,
that they are incapable of correctly applying these criteria. I say that this is
conceivable, but I am not inclined to think it is true. I am much more in-
clined to think that our native processes and our current methods offer a se-
rious possibility of correct application, though the task would not be easy.
However, since I cannot prove that our actual processes and methods are
adequate to the task, I cannot offer a guarantee to this effect. Once again,
though, no invidious judgment is implied vis-a-vis truth-linked criteria. If
our native cognitive processes and currently accepted methods are too de-
fective, they will not allow non-truth-linked criteria to be applied correctly
either. Indeed, if our processes and accepted methods are so flawed, almost
any worthy intellectual endeavor is doomed to failure.
Some philosophers would insist that such a dire scenario can be excluded
a priori. Proper principles of 'interpretation', or belief ascription, they
would say, guarantee that most of a person's beliefs are true. So belief-
forming processes cannot be too defective. I can live with this suggestion,
but I do not wish to insist upon it. In particular, I want to stress a strong
contingent component in process reliability (see Chapter 8). But a modest a
priori component could be allowed as well, if the argument for it could be
made good.
Although I am leery of the thesis that our belief-forming processes could
not be too defective, I acknowledge that nobody could believe that his or
her own processes were defective in a wholesale way; at least nobody with a
modicum of logical acuity and reflectiveness. How could anyone think that
all his or her belief-forming processes are radically defective? A little re-
flection would indicate that this very belief is thereby undermined, since it
too must have been formed by putatively defective processes. Still, while
wholesale self-deprecation is excluded, a person can certainly believe that
some of his or her belief-forming processes are faulty.
Let us now turn to a slightly different argument, which states that there is
no point in seeking a right rule system under a reliabilist criterion, because
there is no possibility of improvement. If we unfortunately start out with an
unreliable set of processes, any attempt to identify proper processes and
methods (right rules) will fail. So we would be stuck in an epistemic quag-
mire. If, however, we are fortunate to start with a reliable set of processes,
any attempt to identify proper processes will be pointless, since we will al-
ready be using such processes.
It is evident, though, that this argument is a flop. First of all, even if we
natively use (mostly) reliable processes, it is another matter to identify what
those processes are. Second, no convincing argument has been given against
the possibility of improvement, of epistemic bootstrapping. The main sort
of bootstrapping scenario runs as follows. We start with a set of available
processes with varying degrees of reliability. We use the more reliable pro-
cesses to identify good methods. We then use the more reliable processes,
together with some of the good methods, to identify the various processes
and their respective degrees of reliability. The superior specimens are so
identified, and their use is said to be justification-conferring. The inferior
specimens are so identified, and their use is said to be non-justification-con-
ferring. If the faulty processes are subject to direct or indirect control, we
try to avoid those processes. Or we try to devise methods to minimize their
impact. In this manner epistemic melioration is possible.
One must now ask whether it is possible to apply a reliabilist criterion

rationally, or at least justifiably. This is (logically) possible. As long as we
have some sufficiently reliable processes (which are not undermined by any
of our beliefs), and they are the sorts of processes needed for the question at
hand, a right rule system would permit the use of those processes. So we
coukl arrive at a justified belief about which particular'rule system is right.
(Of course, even a justified belief could be false. So justified application of
the criterion does not ensure correct application of the criterion. But we
have already addressed the problem of correctness.)
My conclusion is simple. There is nothing wrong with truth-linked cri-
teria in general, nor with a reliabilist criterion in particular. Given the
strong arguments I have marshaled on behalf of such a criterion (or family
of criteria), we have every reason to accept it.
chapter 6
Problem Solving, Power, and Speed
6.1. Intelligence and Goal-Directedness

In chapter 1 I called attention to a variety of terms of epistemic appraisal
and standards attached thereto. But thus far all discussion has centered on
knowledge and justification, which invoke one principal standard of episte-
mic evaluation: reliability. It is time to broaden our perspective, to expand
the range of terms and standards within our purview. No apology is needed
for the lengthy treatment of knowledge and justification. These are critical
to skepticism, and skepticism is historically critical to epistemology. 'But
skepticism, knowledge, and justification should not be the whole province
of epistemology. Nor should reliability be the only pertinent standard.
There are other terms, dimensions, and standards with an equal right to oc-
cupy the epistemological limelight.
For example, 'intelligence' is a more prevalent term of cognitive evalua-
tion in extraphilosophical contexts. Though doubtless very vague, it de-
serves more attention in epistemology than it has hitherto received. I shall
not undertake a full scale analysis of intelligence. However, I wish to ex-
plore some standards of cognitive appraisal that appear to be among the
criteria for intelligence, and defend the importance of these criteria for
epistemic purposes.
The main standards I will emphasize are power and speed. Both are dis-
tinct from reliability and comparable in importance. As noted in Chapter 1,
reliable cognitive mechanisms are an antidote to error, but not necessarily
an antidote to ignorance. Reliable processes guarantee a high truth ratio
among the beliefs they generate. But they may generate very few beliefs
indeed; they may leave the cognizer with very little information, even on
the issues that interest him or her most. The antidote to ignorance is not re-
liability, but (intellectual) power. Powerful cognitive mechanisms are
(roughly) mechanisms capable of getting a relatively large number of truths.
An intelligent cognitive system, it seems clear, is not simply a reliable
system. Intelligence is at least partly a matter of power: a capacity to com-
Problem Soluing, Power, and Speed 123
bat and relieve ignorance, not simply error.' To be sure, an intelligent crea-
ture may be lazy, and may therefore remain mired in ignorance. But an in-
telligent creature is at least capable of relieving its ignorance.
It would be a mistake to equate intelligence with power, at least if power
is equated with number of truths that can be obtained. First, it is not en-
tirely clear that number of truths, rather than, say, complexity, is the appro-
priate standard of intelligence. Suppose system A is capable of getting more
true beliefs than B, but B can solve more complex problems. B might be
credited with more intelligence. This could be accommodated by building a
complexity parameter into a measure of power. But I have no proposals for
a metric of complexity.
Second, a system may be capable of acquiring numerous truths, but
mostly incidental ones, unrelated to its needs or interests. For example, a
system may be capable of indefinitely expanding its true-belief set by add-
ing disjuncts to truths it already believes. But a system that just cranks out
more and more disjunctive true beliefs, oblivious to other pressing questions
left unanswered, might well be considered very stupid.
The last point is really twofold. First, intelligence probably has some-
thing to do with breadth of answerable questions, not simply their number.
Second, a system oblivious to the questions it wants to answer, which
blindly computes answers to irrelevant questions, is not very intelligent. In-
telligence is partly a matter of goal-responsiueness.
If we seek to crystallize a standard of (intellectual) power that is reason-
ably close to intelligence, we need to distinguish several conceptions of
power. One simply makes power a function of the number of true beliefs a
cognitive system can generate. A second makes it a function of number plus
breadth. A third makes power a function of the proportion of questions it
wants to answer that it can answer (correctly), or the proportion of prob-
lems undertaken that it can solve correctly. Only the third conception
makes power a goal-responsive notion, so it is the best conception in this
group. Notice, though, that we should restrict attention not to questions or
problems actually posed by the cognitive system, but rather to those it
could pose.
Further comments on goal-responsiveness are in order. Biological cogni-
tive systems are typically deployed to help the organism satisfy its practical
needs or ends. This involves the choice of actions, designed to achieve spe-
cified goals. Good choices of action-ones that lead to a high level of goal
satisfaction-typically depend on good information, that is, true beliefs or
relatively high confidence in truths. Thus, cognitive mechanisms are used to
try to get true beliefs on pretargeted questions, questions relevant to the
practical tasks at hand. Even when no practical task is at stake, cognitive
activity involves prespecified questions. Purely intellectual questions com-
monly engender a host of subsidiary problems, solutions to which become
124 Problem Soluing, Power, and Speed
targets of inquiry. Intelligence is the ability to succeed in such inquiries, to

find the answers, or solve the problems, antecedently posed. Getting true
beliefs on incidental matters does not count (as much). It is like a game of
billiards, where a player calls his shots, saying which ball he will get in
which hole. If he sinks a different ball, or the right ball in a wrong hole, it
does not count. Success in cognitive matters is also measured mainly by
solving the problems one is tying to solve, not forming a lot of incidental
true beliefs along the way. (Of course, incidental truths may be useful later,
when one's goal structure alters, so incidental true beliefs are not valueless.
The issue is not as sharply defined as in billiards.)
Unlike power, reliability has nothing to do with goal-responsiveness.
Cognitive processes could be wonderfully reliable but wholly unresponsive
to current goals. Thus, the notion of power I am placing on our agenda in-
vokes considerations of a different order than those invoked by reliability. I
suggest that intelligence involves some such goal-responsive parameter, and
any total, well-balanced evaluation of a cognitive system should do so as
well.
It would still be a mistake to suppose that intelligence can be equated
with power. In addition to complexity, at least one other variable must be
included: speed. Other things equal, a more intelligent person is one who
solves problems relatively quickly. It is possible that speed can be partly
subsumed under power. First, the faster one solves problems, the more
problems one can solve in a given time period. Second, many cognitive
goals, or tasks, can be conceptualized as finding an answer to a given ques-
tion by a certain time. When the time limit is short, slower problem-solving
processes will not satisfy the constraint, hence will not be able to execute
the specified task at all. But time constraints are not always built into intel-
lectual tasks. So it is best to retain power and speed as two distinct stan-
dards of evaluation.
The human cognitive architecture clearly has design features sensitive to
power and speed. Its attentional control mechanisms respond to current
goals, and its memory seems structured for retrieval of relevant informa-
tion. The system also appears to have a number of devices that shortcut
belief formation: heuristics that enable a belief to be formed by nonal-
gorithmic and nonexhaustive procedures. These are obviously conducive to
speed. I shall explore a variety of such mechanisms in Part 11.
It may be asked, however, whether such features of the human system are
normatively acceptable. Don't they conflict, in some cases, with reliability?
Clearly there are going to be trade-offs between speed and reliability, as
well as trade-offs between goal-responsiveness and reliability. How does
possession of the indicated mechanisms stack up against a proper balancing,
or weighting, of the several standards I have introduced?
If there were an attractive way to fuse these variables into a single mea-
sure, I would cheerfully endorse it. But I know of no such formula. We must
live for now with an irreducible multiplicity of standards. This may be a bit
disappointing, but not surprising. Consolation may be taken from ethics and
the law, where analogous complications arise. It is hard to balance the
competing claims of justice and mercy, of public good and individual rights.
We should not expect an easy formula for epistemic evaluation either. Dis-
tinct standards need to be treated separately. That is why I qualified all talk
of rules in Chapters 4 and 5 as iustijicational rules: they were not rules for
overall epistemic status. Still, all appropriate standards need to be kept in
mind. Thus, my assessments of cognitive mechanisms in Part I1 will heed
power and speed as much as reliability.
6.2. Problem Solving and Truth

Implicit in the preceding discussion is a truth-linked conception of problem
solving: getting a true answer to a question. But recent epistemological dis-
cussions of problem-solving contexts have not all so interpreted 'problem
solving'. Indeed, the most extended use of problem-solving terminology,
namely, that of Larry Laudan in Progress and Its Problems, expressly di-
vorces truth from problem solving.2 A more precise account and defense of
my truth-linked usage is therefore in order. This will also pave the way for a
view of the cognitive activities involved in problem solving and unveil a
fuller vision of the cognitive processes that are proper objects of epistemic
evaluation.
The language of problem solving has cropped up intermittently in episte-
mology and the philosophy of science. John Dewey's theory of inquiry, for
example, rested on a notion of a "problematic ~ituation."~ More recently,
philosophers of science like Karl Popper, Thomas Kuhn, Imre Lakatos, and
Laudan have all talked of problem solving, or puzzle solving, as an impor-
tant facet of ~ c i e n c eSome
. ~ of this usage, however, has only selected points
of contact with mine. For example, Kuhn, Lakatos, and Laudan view prob-
lem solving as a property of theories, not an activity of cognizers, which is
what my usage will indicate.
However, a number of recent philosophers of science have linked the
language of problem solving with that of questions and question answering,
as I shall do. Jaakko Hintikka, for example, proposes an 'interrogative'
model of scientific inquiry.5 And both Sylvain Bromberger and Bas van
Fraassen make question answering central to their theories of scientific ex-
planation.6 My own approach is intended to be even more general. It is in-
tended to subsume every sort of intellectual question: scientific or
nonscientific, theoretical or nontheoretical. This is appropriate for a general
epistemological framework.
Why interpret problems as questions in the first place? For the purposes
126 P r o b h Soluing, Power, and Speed
of epistemology, a problem is an intellectual problem. This makes it natural

to view problems as questions, to which cognizers (or groups of cognizers)
either need or want answers. Doubtless there are other senses of 'problem'
not covered by this approach. Emotional problems are sources of trouble, or
distress, but are not (typically) questions (though they may give rise to a
question: How can the person be rid of the distress?). But for epistemologi-
cal purposes, identification of problems with questions should prove
congenial.
To analyze problems and problem solving, let us begin with the issue of
what it is to have a problem, in the intellectual sense that concerns us.
S has problem Q if and only if:
(1) Q is a question,
(2) S wants to have a (true) answer to Q,
(3) S does not believe that he has a (true) answer to Q, and
(4) S does not have a (true) answer to Q.
This account is somewhat stronger than it needs to be. There might be a

sense of having a problem that is compatible with believing that one has an
answer, contrary to (3).A cognizer could have a problem without realizing
it, because he thinks he's got the answer. In this sense, though, having
problems will not necessarily motivate one to try to find answers. Since I am
ultimately interested in activities directed at finding answers, I am inclined
to neglect this possible sense. Another definition weaker than the above one
would delete condition (4). This would make having a problem compatible
with having an answer (but not realizing that one has it). This can occur if
you believe a proposition that does answer Q, without believing it answers
Q. (Perhaps you just haven't noticed that it answers Q.) However, this sense
can be neglected for simplicity. The central cases are those captured by the
definition provided.
An important feature of this definition is the desire element. Clause (2)
implies that S has Q as a problem only if S wants to have an answer to Q.
(Compare Bromberger's notion of a "p-predicament.")7 Thus, someone ig-
norant of a certain question's answer, but wholly uninterested in it, does not
have that question as a problem; he does not make it one of his questions.
Notice that wanting to have an answer to Q is not the same as wanting to
be told an answer to Q. In general, my discussion focuses on intrapersonal
questions and answers, not interpersonal question asking and answering. A
teacher, who already knows the answer to Q, may want to be told the an-
swer by a student. But this is incidental to the having of a problem.
Another central feature of the definition is that having a problem in-
volves wanting a true answer. This seems natural enough. When you ask a
question, you don't just want a random candidate answer; you want the
real, correct, answer. Sometimes people do want merely candidate answers.
Problem Soloing, Power, and Speed 127
But then their question is: What are some candidate answers to that ques-
tion? and there again a correct answer is desired, namely, things that are
candidate answers.
Given this definition of having a problem, the natural definition of a
problem solution is:
Proposition A is a solution (answer) to problem (question)Q if and only if
(1) A is a potential answer to Q, and
(2) A is true.
Erotetic logic studies the logical relations between questions and answers.
This includes the study of which propositions are potential (relevant) an-
swers to which questions. By a potential answer, I mean roughly what Nuel
Belnap and Thomas Steel call a "direct" answer: a proposition that answers
the question completely, but just completely? 1 will not survey or endorse
any theories of erotetic logic. I will just presuppose our ability to recognize
potential answers.'
Given the controversy over the truth-condition, it is worth underlining its
appropriateness further. Its suitability is especially clear for the terms
'problem' and 'solution'. In common usage one does not solve a problem
unless one gets a correct solution: a real solution is a correct solution. No-
body is credited with solving a mathematical problem unless he or she finds
(is credited with finding) a true solution. Nobody is credited with solving a
game problem or puzzle, such as Rubik's cube or cannibals-and-mission-
aries, without getting a correct solution. A detective does not solve a mys-
tery without figuring out who really committed the crime. In crediting
someone with a solution, we mean that he or she has a true solution-al-
though we can, of course, be wrong about this.
Is a solution to a problem necessarily a proposition, as our definition im-
plies? In the Rubik's cube puzzle, or other such puzzles, a solution might
appear to be a specified end state, such as a configuration of the cube; and
this is not a proposition. However, the propositional formulation can read-
ily handle this and similar cases. A solution may be construed as a proposi- '
tional description of an appropriate end state. Or, when the problem is to

find a series of moves that will transform an initial state (such as a certain
cube configuration) into a specified end state, then a solution consists in a
proposition that describes a series of moves from the initial state to the end
state.''
There are two senses in which someone may be said to have a solution to
a problem, a weak and a strong sense. The weak sense is this:
S has a solution to problem Q if and only if
there is a proposition, A, such that
(1) A is a true answer to Q, and
(2) S believes that A is a true answer to Q.
128 Problem Soloing, Power, and Speed
The strong sense is this:

S has a solution to problem Q if and only if
there is a proposition, A, such that
(1) A is a true answer to Q, and
(2) S knows (or believes iustifiably) that A is a true answer to Q.
Both of these senses are legitimate. If a student gets a correct solution to a
problem by mere guessing, he still has a solution, even if his belief in it is
neither a justified belief nor a piece of knowledge. However, in saying that
someone has, or has found, a problem solution, we often imply some sound
method of belief formation, sufficient (along with the satisfaction of other
conditions) for knowledge or justified belief.
This concludes my analysis of problem solving. It is extremely simple,
which is a virtue, not a defect. However, there is much more to say, even in
an 'analytical' vein, about problems and their solutions.
As writers on erotetic logic stress, there are many varieties of questions:
whether questions (yes-or-no questions), why questions, where questions,
how questions, and so on. Some examples are: Is it possible to find true hap-
piness? Who is the mayor of Tucson? What element has atomic number 94?
What is a proof of such-and-such a theorem? How can I fix this leaky
faucet?
There are other ways we might categorize questions. One taxonomy
might be: (a) theoretical, (b) practical, and (c) productive. Theoretical ques-
tions would include those of science, mathematics, and philosophy. Practi-
cal questions would include how-to questions like, How can I fix this faucet?
and others directly linked to practical needs or desires, such as, Where can I
get immediate medical attention? Notice that most practical desires or
needs have associated questions, namely, What actions can I perform to sat-
isfy this desire or need? This is the sense in which practical problems nor-
mally have associated intellectual problems. (However, solving a practical
problem consists in eflecting a desired result, not merely getting a true
proposition that specifies a way of effecting that result. This is another rea-
son why the intellectual conception of problem solving does not exhaust all
senses of 'problem solving',) Productive questions are really a species of
practical questions. But it is worth highlighting the fact that creative, artis-
tic, and technical activities generate problems and solution-seeking activi-
ties. A painter may ask, What pattern or color in this portion of my canvas
will achieve such-and-such an aesthetic effect? and a composer might ask an
analogous question while composing a quartet.
Another possible taxonomy would classify by number of correct answers.
Questions can have zero, one, or many correct answers. Normally a person
wants one correct answer, even if he believes there are more. It is also pos-
sible, of course, to want an answer to a question when there aren't any: one
can mistakenly believe that there are. Philosophers used to study a special
class of such questions, 'pseudoquestions'. Pseudoquestions were said to be
meaningless, or irresoluble, philosophical questions that have a semblance
of meaningfulness and answerability. This status was invested with some
philosophical importance. Of course, questions can also lack true answers
for philosophically uninteresting reasons. What is a proof of this formula?
has no true answer if the formula is not a theorem. Where is a good restau-
rant in town? has no true answer if the town has no good restaurants. ('No-
where' is not a true answer so much as a denial of the question's
presupposition. Typically, questions have presuppositions, the falsity of
which can imply the absence of any true answer.)"
Some questions delimit, either explicitly or implicitly, a fixed set of po-
tential answers. Yes-or-no questions are a special case, where the two can-
didate answers are 'yes' and 'no' (or corresponding full sentences). Other
questions enumerate a list of possible answers, for example, Who is the old-
est: Sally, Mary, or Beverly? Many questions, however, delimit no finite or
fixed set of possible answers. These include, What explains the moon illu-
sion? and What is a cure for cancer? For epistemological purposes, this is
the most interesting class. To find an answer to such questions, a cognizer
must construct possible answers. He cannot simply assess truth-values of the
specified candidates. This cognitive task is clearly more complex. Admit-
tedly, truth-value assessment is often difficult enough, but not always. It
may be easy to determine whether a proposed proof of a putative mathe-
matical theorem really is a proof. So it is easy to answer the question, Is this
putative proof a genuine proof? But constructing or inventing a genuine
proof may be extremely difficult, requiring lots of creativity. So it is much
harder to answer, What is a proof of this statement?
Science characteristically asks questions for which it is difficult both to
devise plausible answers and to verify their truth-value. Because of verifi-
cational difficulties, there are often deep methodological disputes over
principles of verification. Does this mean that tnlth should be dispensed
with as a scientific desideratum? Surely not. Just because something is hard
to get, even hard to recognize if and when you've got it, does not imply that
it should not be sought. There is ample reason, moreover, for understanding
scientific inquiry as a quest for truth, like other inquiries. One significant
scientific activity is the devising of new kinds of equipment, technologies,
and methodologies that can facilitate theory testing. The natural explana-
tion of such activities is that better devices, better statistical methodologies,
and so on, will enable researchers to identify, or get closer to, the truths
they seek.
I will say more about truth seeking in science in Chapter 7. But let me
add a few remarks here about some authors' explicit rejection of truth in the
conception of scientific problem solving. Laudan, in particular, offers a
130 Problem Soloing, Power, and Speed
conception of problem solving, applied to theories, in which a theory can

solve many problems without being true. Is this usage legitimate? If so, how
can it be reconciled with my truth-requiring analysis of problem solving?
The predicate 'solves a problem' sometimes is applied to admittedly false
theories. As Laudan indicates, we can say that Ptolemy's theory of epicycles
solved the problem of retrograde motions of the planets, though Ptolemy's
theory was admittedly false.12 One possible use of 'solves' is merely in the
quoted sense, analogous to the occurrence of 'knew' in 'The ancients knew
that the sun goes around the earth'. A speaker who utters this sentence does
not really mean to assert that the ancients knew this; he would just be put-
ting the matter as the ancients themselves would have put it (language dif-
ferences aside).
To understand a different usage of 'solves', suppose we grant a non-truth-
implying sense of 'problem solving', which is applicable to theories. Let it
be what Laudan suggests, roughly, that a theory solves an (empirical) prob-
lem if it (together with suitable auxiliary assumptions) implies a statement
describing the problematic phenomenon, while competing theories fail to
imply it. The question remains: Is the scientist's main theoretical aim to
find theories that solve problems in this sense? Or does the scientist have a
deeper, truth-linked aim? I suggest there is a deeper aim. The scientist typi-
cally wants to find a true theory, wants to answer the question (solve the
problem), What theory really explains the phenomena in the specified do-
main? where no theory can 'really explain' a phenomenon unless it is true.
Now a theory's ability to solve problems in Laudan's sense is readily under-
standable on this truth-linked aim. One piece of evidence in favor of a the-
ory's truth is that it implies actual phenomena that rival theories do not
imply. Thus, I propose that scientists are interested in this non-truth-linked
'problem-solving' property of theories as a means to finding a true theory.
Of course, the fact that scientists often settle for mere approximations to
truth must be reconciled with the claim that they are really after truth it-
self. This large topic cannot be fully addressed here. But the main lines of an
answer, I think, have to do with the idea that in settling for approximations,
scientists live with a temporary stage of what they view as a longer enter-
prise. Since scientific theory deals with the most complex scientific ques-
tions, they do not expect to find fully accurate theories soon, perhaps even
in their lifetime. Still, progress has been made (toward the truth) if outlines,
or general features, of a true theory are identified. They settle for such an
outline, or model, as a momentary step toward a more distant goal. (Of
course, scientists may err in thinking that what they have is even an outline,
or an approximation, of a true theory. But such fallibility is granted by all.
The present question is why scientists settle for a theory sketch, which is
admittedly not correct in all details. That is what the 'means toward a dis-
tant end' explanation is supposed to explain.)
What I am endorsing here is 'aim realism'. This should be distinguished

from other brands of scientific realism, such as so-called convergent realism,
the doctrine that science has in fact been getting closer to the truth; or the
doctrine that scientists try to make new theories continuous with their
predecessors. More will be said about these varieties of realism in Chapter
7.
The proposed truth requirement for problem solving is precisely analo-
gous to the truth requirement for knowledge, and equally appropriate.
Some people might object that it makes problem solving harder than a
weaker sense would allow. Certainly. But it does not make it unnecessarily
harder. In its central sense, problem solving does require truth, just as
knowledge requires truth. So our analysis does not impose anything ex-
cessive on the concept of problem solving. If it turns out to be hard to solve
certain problems, so be it.
6.3. Cognitive Components of Problem Solving

Whether a cognizer actively seeks an answer to a given question depends on
the relative strength of the desire for an answer, as compared with other
desires. It also depends on his beliefs about the comparative ease or diffi-
culty of (his) finding an answer. If he thinks it would be hard, and if he has
better uses of his time and cognitive resources, he may skip the quest. Sup-
pose, however, that he devotes at least some cognitive resources to the
question. What specific activities, or operations, will be utilized? This raises
the more general question: What are the components of problem-solving
activity?
Activity aimed at problem solving frequently involves overt actions, such
as opening books, asking experts, and physically intervening in the world
to get relevant data and responses. But I shall focus on purely mental opera-
tions. This follows from my central interest (here) in primary epistemology. ,
Furthermore, physical activities aimed at problem solving are always con-

trolled by psychological processes. So the latter have a natural priority.
What are the cognitive components, then, of problem solving (often
called 'inquiry')? The positivist tradition tended to focus on (a) hypothesis
generation, (b) test, and (c) doxastic decision, where (c) consists in a 'deci-
sion' (though not necessarily a voluntary decision) to accept, reject, or as-
sign some probability to a hypothesis.
This list of components is easily expanded. In addition to hypothesis gen-
eration, there is also concept formation (also discussed by the positivists).
Recent philosophers of science have stressed a further component, namely,
pursuit.13In trying to find an answer to a theoretical problem, one decision
the scientist makes is which of various research programs to pursue, that is,
132 Problem Solving, Power, and Speed
which general approach to a problem is worth refining or amending, in the

hopes of getting an optimal theory.14
Cognitive scientists interested in problem solving are normally con-
cerned with more detailed, microcomponents of the foregoing categories.
Their studies are not fundamentally at odds with those of philosophers.
Sometimes, though, the categories do not correspond to one another in ob-
vious ways. Let us look at a few aspects of problem solving, as often dis-
cussed by cognitive scientists.
One such aspect is the way a cognizer understands, formulates, or 'repre-
sents' the problem to himself-how he conceptualizes it. One conceptual-
ization of a problem can make it hard to solve, while another makes it easy.
An example from Michael Posner is the following:
Two train stations are 50 miles apart. At 2:00 P.M.two trains start toward
each other, one from each station. Just as the trains pull out of the stations, a
bird springs into the air in front of the first train and flies ahead to the front
of the second. When the bird reaches the second, it turns back and flies to-
ward the first. It continues to do this until the trains meet. If both trains
travel at 25 miles per hour and the bird flies at 100 miles per hour, how many
miles will the bird have flown before the trains meet?''
If the problem is conceptualized in terms of the bird's trajectory, one

must calculate how far the bird goes on each trip between the trains. The
problem is then difficult. But if it is conceptualized in terms of flight time, it
becomes trivial. The trains must travel one hour before meeting (50 miles at
a closing speed of 50 miles per hour), and at 100 miles per hour the bird will
fly 100 miles.
The importance of how a problem is 'represented', or 'framed', cannot be
underestimated.'' Researchers have therefore explored the factors that in-
fluence representation. What stimuli prompt the cognizer's representation,
and how easy or hard is it to revise initial representations? Initial represen-
tations, it has been found, often tend to structure subsequent thinking,
sometimes to confine thinking to rigid 'loops', in which the cognizer keeps
recycling the same themes. The mathematician G. Polya, noting the threat
of rigidity, advised problem solvers to avoid commitment to an initial repre-
sentation.17 Successful problem solving sometimes depends on holding hy-
potheses in check.
The impact of initial representations is illustrated by Posner from studies
of impression formation." These studies indicate that the items first en-
countered on a list structure one's mental organization of the list. When
subjects are asked to determine which of the following four words is inap-
propriate (to the rest)-'skyscraper', 'cathedral', 'temple', and 'prayer'-
they tend to respond 'prayer'. Since 'skyscraper' occurs first, the list has ap-
parently been organized in terms of buildings, so 'prayer' is the odd entry.
Problvm Soluing, Power, and Speed 133
When the same list is presented in a different order-'prayer', 'cathedral',

'temple', and 'skyscraper'--one tends to organize the items by the theme
religion, and 'skyscraper' is the odd item.
The significance of representation is explored in other contexts too. One
such area is the differences in representation between experts and novices.
What enables an expert to solve problems better than a novice-solve more
of them, and more quickly? Part of the answer appears to consist in how ex-
perts conceptualize the problem domain. For example, expert chess players
seem to group or chunk configurations on a chessboard into larger units
than do novices. They have more global, or abstract, categorizations of the
problem space.lg
The importance of abstraction is cited by many researchers in cognition.
Walter Kintsch and T. A. van Dijk's theory of story understanding empha-
sizes increasingly abstract 'macrostructure' representations of a prose pas-
sage, representations that correspond to summaries of the passage at high
levels of generality.20Such abstraction is also held to be essential in devising
analogies deployed in problem solving. Breakthroughs in problem solving
often occur when the problem solver discerns an analogy between the tar-
get problem and previously encountered problems, possibly from different
domains. The psychologists Mary Gick and Keith Holyoak have explored
the function of analogy, indicating the critical value of an appropriate level
of abstraction before analogies in different domains can be d i ~ c e r n e d . ~ ~
I suggested earlier that cognitive scientists are interested in many of the
same categories of inquiry as philosophers. But sometimes the cognitive sci-
ence perspective can transform those categories. Consider the category of
evidence gathering, for example. As usually construed by philosophers of
science, gathering evidence would consist in making observations of the
world. But the perspective of cognitive science suggests a further sense of
'evidence gathering'. It is normally assumed that the main work of problem
solving occurs in the active portion of the mind-sometimes called 'work-
ing memory', or 'short-term memory'. But most of a cognizer's information
is stored in long-term memory. Whatever has previously been learned and
might be useful for the problem at hand will not be utilized unless it is re-
trieved from long-term memory (LTM). Indeed, it may often happen that
relevant information-relevant 'evidence'-fails to be retrieved from LTM,
and the target question goes unanswered as a result. This means that a cog-
nizer's evidence-gathering task includes not merely making suitable obser-
vations, but also gathering (retrieving) information from LTM. Success or
failure in answering a question depends on retrieval effectiveness. Hence,
the problem-solving power of a cognitive system-at least any system with
a dual memory structure---depends on retrieval properties of that system.
This aspect of intellectual power has been largely ignored.
The importance of information retrieval should be obvious in the case of
inductive inference, where any piece of evidence might override, or under-

mine, other evidence. But even in deductive tasks, retrieval from LTM is
essential. Suppose that the innumerable beliefs in your memory bank in-
clude the following:
Barney sent Betty a valentine.
Betty loves Barney.
Mary can't stand Mike.
All lovers love any lover.
Sally is sad.
Billy is Sally's brother.
Sally secretly admires Betty.
Mary loves Bill.
Bill didn't send Betty a valentine.
Suppose you now pose the following question to yourself: Does Mary love
Betty? Relevant to this question is whether anything you know (or believe)
entails this proposition. To determine this, you must probe your memory for
such items. Under my hypothesis, there are propositions you believe (stored
in LTM) which jointly entail that Mary loves Betty, namely, (2), (4), and (8).
But will a probe of LTM access these items? Will memory 'report' them to
you, that is, 'activate' these items so that you can perform deductions from
them? Whether LTM does this depends on how these items are organized in
your LTM, and how retrieval from LTM works in general. These factors are
critical in determining whether you will correctly answer the question at
hand. More generally, they are critical in determining a human cognizer's
problem-solving power (as will be discussed more intensively in section
10.3).
Now if memory retrieval is one of the determinants of problem-solving
power, and if power is one of the standards of epistemological interest, then
not all operations of epistemic interest are matters of rationality or irratio-
nality. For memory retrieval is not naturally classed as rational or irrational.
Traditionally, epistemic evaluation has tended to confine itself to questions
of rationality, or similar questions. But under the approach advocated here,
other dimensions of epistemic evaluation are also of interest. Specifically,
the efficiency or inefficiency of various cognitive operations can be sub-
jected to epistemic scrutiny, where by 'efficiency' I mean effectiveness in
getting problems solved.
Another critical aspect of problem-solving activities is the sequence of
problem-solving steps. This aspect is emphasized by studies of problem-
solving 'strategies', especially in artificial intelligence. These studies explore
various principles for searching a problem space for a solution, by succes-
sively examining various 'moves' or 'states' in the problem space. One stock
example of a search strategy is hill climbing. Assume that a cognizer can
specify a respect in which a move from one state to another brings him or
her closer to the goal state of the problem. Then the hill-climbing strategy
tells the cognizer always to seek and choose a move that goes up the present
hill, brings the cognizer closer to the goal state. The notion of closeness is
frequently fairly natural. In the cannibals-and-missionariesproblem, for ex-
ample, a move that results in two cannibals and two missionaries being on
the other side of the river gets one closer to the goal than the initial state.
(The initial state is no cannibals or missionaries on the other side; the goal
state is all three cannibals and all three missionaries on the other side.) Hill
climbing has its defects as a general strategy. For example, it can get one
stuck on a local hill (a local maximum), and give no method for ultimately
climbing the mountain (getting to the real goal state). But it illustrates the
idea of a strategy that results in a series of search steps2'
A cognitive system must operate sequentially, by a series of steps. (I do
not mean to exclude parallel processing.) In trying to solve a problem, it al-
ways faces the implicit question: What should be done next? Furthermore,
it must perform its operations in real time, not in some idealized situation
where time is suspended. Since cognitive tasks often have time con-
straints-especially when linked to practical exigencies-the exact tem-
poral sequence of operations is critical. The epistemic adequacy of
cognitive 'heuristics', or short cuts, must be assessed in light of these
constraints.=
Traditional epistemology and philosophy of science have largely ne-
glected the temporal and operational constraints on the intellect. Discus-
sions of justification and rationality have typically proceeded in terms of
logical and probabilistic relations between theory and evidence; but these
sorts of relations abstract from temporal and operational considerations.
The so-called historical wing of philosophy of science, to be sure, empha-
sizes diachronic factors. But this tradition has not attended to cognitive op-
erations, in the sense intended here. It seems clear, though, that an
adequate primary epistemology must be centrally concerned with the real-
time situation of a cognitive system and the sequences of operations it exe-
cutes. Cognitive science, of course, studies cognition in precisely this .
fashion.
The importance of sequential operations does not just result from our
concern with the epistemic standard of speed. To reason effectively even
without deadlines, a cognitive system must ensure that relevant premises
are accessed at the time they are needed. Suppose that I periodically think
of the fact that the university library opens on Sundays at noon. That will
not be of much help if I neglect to access this fact on Sunday mornings,
when I actually contemplate going to the library. If I fail to retrieve this
fact on Sunday mornings, I may make some futile trips. Thus, the timeliness
of mental operations is an important facet of cognitive life, which must not
be neglected in a comprehensive primary epistemology. But it is readily ne-
glected without detailed study of cognitive operations.
I have focused predominantly on stages of inquiry prior to the doxastic
decision stage, primarily because past epistemology has riveted its attention
on the latter stage. Past emphasis on the doxastic decision stage may partly
be rooted in the dominant interest in justification, which seems mainly con-
cemed with this stage. But if intelligence, or power, is of equal importance,
there is no good rationale for this concentration. Arguably, a cognizer can
have justified doxastic attitudes just by using suitable doxastic decision pro-
c e ~ s e sBut
. ~ ~one cannot solve problems effectively just by using good dox-
astic processes. Doxastic decision processes only serve to select from the
available candidate solutions. At best, selection can reject false candidates
and accept a true candidate if one is generated. But if no true solution is
generated, even the best doxastic decision process cannot produce belief in
a true solution. To be a good problem solver, then, one needs good hypoth-
esis-constructing processes. Similar remarks apply to processes governing
sequential search.
6.4. Social Problem Solving and Product Epistemology

I have concentrated on individual problem solving. But problem solving
can also be a group enterprise. All sorts of groups have shared interests in
finding answers to specified questions. Salient examples are scientific and
scholarly groups, but there are also committees, juries, government agen-
cies, and whole societies. So the theory of problem solving, both descriptive
and normative, extends into the social arena. Problem-solving power is an
appropriate standard for social epistemology. For any number of group
practices and structures we may ask whether they promote or inhibit prob-
lem-solving power. Although detailed reflection on this issue will not be un-
dertaken here, a few general comments may be made, as a preview of future
research.
How might collective efforts at problem solving outstrip those of an indi-
vidual? Some obvious possibilities are these: (1)People can pool their fac-
tual information, yielding more data for each individual. (2) Ideas and
hypotheses can be multiplied more easily, each cognizer profiting from the
larger menu of candidate solutions. (3)Critical assessment of candidate so-
lutions can be facilitated, since an isolated cognizer is easily seduced by the
allure of his own ideas, and defects are easier to detect, and weed out, in the
concoctions of others. (4) Because complex solutions often need many skills
to refine and test, division of labor can facilitate execution of the several
tasks. (5) Incentives can be offered that make intellectual specialization
possible, which may be needed for the long-term investment required for
solving complex problems. (6) Sanctions can be imposed for behavior inimi-
cal to t&th (lies, fabrications, and the like).
All these are commonplace reasons why scientists and scholars form pro-
fessional and academic communities. There are equally commonplace fea-
tures of collectivities that may work against problem solution. Social groups
P r o b h Solving, Power, and Speed 137
generate systems of status, authority, and (social) power. When those ac-
corded a high status get things wrong, their false doctrines may gain wide
currency and may be difficult to eradicate, even in the face of striking
counterevidence. The influence of authority figures may congeal ideational
frameworks and inhibit novelty and invention.
Setting aside these issues, which are too large for present scrutiny, notice
that there are questions to be raised about what it even means for a group to
solve a prob1e.m. Does it mean that all members of the group believe a true
answer? Presumably that is much too strong. Exactly which members, then,
or how many, need to believe a true answer? Once this issue is raised, it be-
comes obvious that a finer-grained treatment of the topic is in order. Take
the case of governments. Certain experts in the State Department, or the
Defense Department, may know the truth about a certain question, but
cannot persuade their higher-ups of this truth. Perhaps they cannot even
have timely communication with decision makers to get the truth heard. If
the decision makers lack (belief in) the truth, it is arguable that the govern-
ment does not have the solution, even though some of its members do.
In any case, the distribution of information (or belief) within an organiza-
tion is certainly of paramount importance. This topic should be placed on
the agenda of social epistemology. But the point is also of interest because
of analogies with (primary) individual epistemology. As noted in section 6.3,
the mind also has functional locations, such as working memory and long-
term memory. Just as it matters to an organization where information is lo-
cated within it, so it matters to the mind where information is located. If
information in an organization resides only in some lower-echelon official,
and never gets persuasively communicated in a timely fashion to higher of-
ficials, its value to the organization may be nil. Similarly, information only
stored in LTM,and never accessed at all or in a timely fashion, has little
value to a cognitive system, either for purposes of behavioral decision or
further problem solving.
Evidently, a critical facet of social problem solving is the process of com- .
munication. This brings into prominence the units of communication by
which individuals hope to shape the thinking of others. These units include
oral discourses, books, articles, research reports, theorems (or proofs), legal
opinions and briefs, and various forms of fiction. Let us call all such works
intellectual products.
I have said that epistemology's aim is intellectual evaluation. Thus far,
our attention has been engaged by the evaluation of mental events and pro-
cesses, such as, beliefs and belief-forming processes. In fact, intellectual
evaluation far more frequently centers on what I am calling products than
on internal states and operations. In everyday intellectual and cultural
commerce, it is not mental acts that comprise the chief objects of appraisal;
it is discourses and published works. It would be a lean epistemology indeed
that did not feature such objects in its purview. Social epistemology, at a
minimum, must be concerned with these objects, since they are the life-
blood of cooperative intellectual endeavor. So let us momentarily shift our
focus from belief epistemology to product epistemology.
However, I doubt that product epistemology can be divorced entirely
from belief. A principal aim of most intellectual products is to inform an au-
dience, to induce beliefs in putative truths. Several dimensions of evalua-
tion flow naturally from this aim. First, a premium is placed on products
that present new truths, truths not already circulating in the public forum.
If some true contents have previously been aired, people have had the op-
portunity to learn and believe those contents. So their significance is re-
duced. But this needs some qualification. Even if a given truth has
previously been presented, a product that presents it more cogently, per-
suasively, or comprehensibly is still highly valued. Thus, second, the mode
or manner of presentation counts substantially in a product's appraisal.
Third, value is placed not simply on truth per se, but on truths that interest
the audience. Where individuals or groups place high priority on answering
certain questions, products that succeed in answering those questions-or
are judged to answer them successfully-receive special plaudits.
Still another determinant of product evaluation may also be traced to
mental factors. The grading of a product typically involves the question of
originality, or creativity. By how much does this product go beyond previ-
ous work in the field? How difficult was it to produce this work, given the
common knowledge and tools of the discipline? A certain mathematical
lemma, for example, may never have been stated or proved before, but it
might still be an obvious consequence of a well-known theorem. The mag-
nitude of an intellectual achievement is partly a function of the difficulty of
generating it. Clearly, such judgments of difficulty presuppose some metric
of originality or creativity, and the question arises of the basis of such a met-
ric. It can only be based, I believe, on some conception of normal, human
cognitive capacities, a matter on which cognitive psychology may be able
to shed some light. I shall return to this theme in Chapter 11. Meanwhile, I
content myself with the remark that even the evaluation of products in the
social arena may be illuminated by reference to human psychology.
6.5. Multiple Values: Adjunct and Specialized

A recurring theme of the past several chapters is the epistemological impor-
tance of true belief. My accounts of knowledge, justifiedness, and problem
solving were all variations on this theme. This may convey the impression
that I favor a monistic approach to intellectual value. But I wish to correct
any such impression. Value monism is an inappropriate posture for a com-
plete epistemology.
Nevertheless, I do wish to preserve the centrality of true belief in the net-
work of intellectual values. The envisaged network is partly expressed in
the following conjecture. We have various cognitive faculties and opera-

tions that seem to promote true belief (or accurate representation), and this
is their principal utility. Still, we do not value these faculties and operations
solely as means to true beliefs. The exemplary exercise of these cognitive
faculties is a pleasure in itself. We enjoy and admire the free exercise of per-
ceptual, calculational, imaginative, and recollective powers quite apart
from our interest in the truths they reveal. Thus, although the fundamental
value of these cognitive powers consists in their promotion of truth acquisi-
tion, they also have udjunct value. The cognitive activities themselves, and
the external products they generate, may be cherished for their own sake.
Indeed, these activities and products may often conflict with the pursuit of
truth, and they may be valued in spite of this. Hence, at least from a de-
scriptive vantage point, a theory of intellectual activity must embrace value
pluralism.
Let me briefly illustrate and concretize this idea. A pervasive aspect of
cognition is sensuous experience, that is, mental events associated with the
sense modalities. By sensuous experience, I mean not only perception itself
but the use of visual, auditory, and other 'codes' in imagery and imagina-
tion. The existence and nature of such processes is controversial, a matter I
will explore further in Chapter 12. Meanwhile, appealing to both introspec-
tion and psychological experimentation, I will assume that modality-spe-
cific codes are a facet of our mental operation. My conjecture is that the
underlying function of such equipment is to help form accurate representa-
tions or true beliefs. This is realized primarily in perception of the environ-
ment, but secondarily in imaginative construction of possible patterns and
scenarios.
While the evolutionary, or adaptive, value of modality-specific codes
may be linked to accurate representation of reality, we delight in our sen-
suous powers for their own sake. The visual and musical arts clearly draw
their pleasures from the creation and apprehension of sensuous forms and
patterns. Virtually every art form has a close link with modes of sensuous .
representation. Values embedded in these art forms arise from deep-seated
features of our inner sensuous life.
A second pervasive aspect of cognition is imagination itself, by which I
mean the mental construction of new patterns, combinations, sequences,
structures, and so on, out of old elements or parts. The elements can be
either sensuous or purely conceptual. Thus, imagination can consist of the
mental construction of a visual scene or the construction of a hypothesis out
of antecedently existing concepts. The child who imagines a battlefield, the
composer of a sonata, and the mathematician framing a new conjecture, all
construct new mental combinations out of old mental materials.
The primary use of imagination is in the formation of accurate represen-
tations about reality. To answer questions about the world, one typically
has to generate candidate answers, and such generation typically involves
new permutations of familiar conceptual ingredients. But while imagina-

tion is therefore critical to problem solving, or question answering, it is
clear that free exercise of the imagination is a joy in itself. Children's play
largely consists of fashioning combinations of actions or plots out of pre-
viously encountered materials. Such play has no intended effect on true be-
lief; the spontaneous pleasure attached to it is manifest. (At the same time,
spontaneous pleasure in multiple generation of new combinations itself
promotes problem-solving ability. Perhaps the young learn faster than the
old because they are quicker and more indefatigable in generating new
hypotheses.)
Since virtually all cognition is laden with imaginative constructions, most
of this activity is taken for granted. A baseline of such constructivity is as-
sumed for all humans. One could not be human without generating combi-
nations out of antecedently present elements. Furthermore, the general
lines of construction may be largely species-specific. This would account for
the commonality of the concepts we generate, the predicates we project,
the possibilities we envisage. But certain constructed patterns we recognize
as relatively unusual or exemplary. These are what we dub 'imaginative'
creations. Imaginativeness is not a characteristic of all products of the imag-
ination: only those surpassing the routine outputs thereof. This is the theme
sketched at the end of the previous section, and one I will develop further in
Chapter 11.
In addition to the adjunct values attached to our cognitive powers, the
plurality of cognitive values has another source: the institutionalizationof
specialized branches of science, scholarship, arts, and culture. First of all,
these fields often develop a specialized interest in a selected class of truths.
Second, traditions of pursuing these truths may draw disproportionatelyon
selected cognitive skills. So habits of criticism in the field may place rela-
tively more emphasis on those sorts of skills.
Consider first scientific and scholarly fields that (pretty) clearly aim at
truth. Some of these are specially interested in certain sorts of truths. The
nomothetic sciences are interested in lawlike, causal, or theoretical truths.
They prefer truths that are comprehensive, full of explanatory power, and
simple. Historical subjects, by contrast, may not aim at lawlikeness or
theoreticity. Simplicity of a historical account may be regarded as a defect.
History arguably disdains the abstract in favor of richness of detail, individ-
uality of the personages or economic circumstances of the episode in
question.
In this context recall my rejection in Chapter 5 of explanatory power as a
value to be linked with justification. That point can now be clarified. It is
undeniable that science aims at explanation. But this is a specialized aim,
which is not obviously present (or present to the same degree) in all cogni-
tive undertakings.
Other specialized aims and vehicles account for specialized values at-
tached to diverse traditions and genres. Although language and symbolism
are the medium of the sciences, humanities, law, and the arts, the aims and
uses of language and symbolism differ widely, and the values emphasized
differ accordingly. In law and philosophy, disputation and argumentation
play a central role. The skills of the debater and the controversialist are
emphasized. In other fields, such as the fictive arts, different kinds of rhe-
torical styles are emphasized; disputation plays little role. While the law-
yer, the philosopher, and the playwright may all aim at some sort of truth,
the formats and techniques for expressing their respective ideas are very di-
vergent. Similarly, canons of science and mathematics require precision and
rigor, whereas the literary fields call for expressions that are evocative and
rich in connotation, possibly precluding precision and rigor.
It may indeed be suggested that different disciplines and genres have
fundamentally different modes of conveying truths. Science, scholarship,
and philosophy concentrate on the assertive mode of expression, whereas
literary works may try to convey truths by showing, or depicting, them. The
novelist shows what it is like to be a certain kind of person, to live in a cer-
tain era or culture. This is a commonly valued way of getting understanding
of human and cultural facts; important information is somehow conveyed.
Epistemology must be concerned with such diverse modes of communica-
tion, and it must countenance multiple values specifically keyed to these
diverse modes. Even among purely assertive communications, differences in
style are crucial to the mental or doxastic upshots in an audience. Style af-
fects perspicuity, which affects how well a message is understood, compre-
hended, or represented by the audience. This too is reflected in critical
evaluations of intellectual products.
Epistemology has focused heavily on statemental units of communication
and on propositional units of mentation. But this may be an inadequate ap-
proach if epistemology seeks to deal with understanding or comprehension.
How does a cognizer comprehend a large body of data or a highly complex
theory? Can such comprehension be well represented by a list of believed
propositions? Does incomprehension simply consist in failure to grasp the
relevant propositions (whatever 'grasping' is)? Recent work in psychology
and artificial intelligence posits different sorts of mental structures-sche-
mata, frame-systems, scripts-that encode complex informational struc-
tures. An adequate belief epistemology may want to make use of such
mentalistic constructs. At the same time, a product epistemology may want
to appraise, or elucidate appraisals of, discourses in terms of how well they
succeed in effecting suitable sorts of mental structures. Thus, true 'belief
may be too restrictive a conception of epistemic aims, even when other ad-
junct and speciaIized values are set aside.
chapter 7
Truth and Realism
7.1. Varieties of Realism

The notion of truth is important for almost every epistemology, but it is
central to my epistemology for special reasons. I have invoked truth not
only, as is commonly done, as a condition for knowing, but also as a critical
element in two dimensions of epistemic appraisal, namely, 'justification'
and 'intelligence'. The roles to which truth is assigned in my epistemology
may be a source of queasiness to some. As we saw in section 5.8, some phi-
losophers think they discern a circularity in using 'true' in an account of the
epistemic term 'justified', because the notion of truth is itself epistemic.
Other philosophers might hold that it is misguided to assign truth or even
the pursuit of truth, a central role in problem solving, because at least one
central domain of problem solving-the domain of scientific theory-in-
volves no pursuit of truth at all, but only predictive power, calculational
convenience, or the like. For these and other reasons, a discussion of truth is
an important item on our agenda. It must be shown that the truth concept
can perform the functions to which my epistemology assigns it, and that it
has the 'realistic' properties that I have attributed to it throughout.
As I have indicated, I do wish to defend some kind of realistic view of
truth. But not everything called 'realism', not even every variety of realism
about truth, is a doctrine I mean to endorse. So I need to disentangle differ-
ent conceptions of realism, in order to keep our bearings.
In recent years the most influential work on truth and realism has been
done by Hilary Putnam and Michael ~umrnett.'Much of my discussion will
therefore relate to theses that they have advanced or explored. Dummett
conceives of realism as a family of doctrines about different subject matters;
you can be a realist about one subject matter and not about another. He
prefers to say that realism is a view about a certain class of statements, for
instance, physical object statements, mathematical statements; ethical
statements, or statements in the future tense. Here is a recent formulation
of realism he offers: "The very minimum that realism can be held to in-
volve is that statements in the given class relate to some reality that exists
independently of our knowledge of it, in such a way that that reality renders
each statement in the class determinately true or false, again independently
of whether we know, or are even able to discover, its truth-value."'
There are two components in this characterization of realism. First, there
is the principle of biualence: every statement in the class is determinately
either true or false. Second, there is the principle of uerijication-transcen-
dent truth: a statement is true or false independently of our knowledge, or
verification, of it (or even our ability to verify it). The second of these theses
will figure prominently in my discussion. But the first, the bivalence princi-
ple, is tangential.
It is not clear just how firm a criterion of realism bivalence should be.
Dummett regards it as necessary for realism, though not ~ufficient.~ But it is
not even clear that it should be accepted as necessary. Why can't a philoso-
pher be a realist about a given class even if he takes the Frege-Strawson line
that a singular statement containing an empty proper name-such as, 'The
present king of France is bald'-is neither true nor false? Why can't he be a
realist even if he acknowledges vague statements within the given class,
which are neither true nor false?4
Setting these points aside, suppose we grant Dummett the label 'antireal-
ist' for philosophers who deny the principle of bivalence for a given class of
statements. Is it important for my purposes to defend the corresponding
kind of realism, as a general claim? Not at all. There is nothing in my episte-
mology that requires defense of the principle of bivalence for all classes of
statements. I could readily admit the failure of this principle for any num-
ber of classes of statements: statements about the future, subjunctive condi-
tional statements, even statements of mathematics. What is critical is that
when any such statement i s true (or false), what makes it true (or false) is
independent of our knowledge or verification. Thus, the second of the two
realist theses explained above, verification-transcendence, is critical. To put
it another way, truth must not be an epistemic matter. But the issue of bi- .
valence does not affect my epistemology.
Admittedly, the two theses are interconnected in some philosophies.
Some denials of bivalence are consequences of a denial of verification-tran-
scendence. The constructivist in mathematics denies verification-transcen-
dence, saying that a mathematical statement is true only if there exists a
proof of it, and is false only if there exists a proof of its negation. The denial
of bivalence readily follows. However, people may also deny bivalence for
other reasons. Nothing in my epistemology requires any quarrel with biva-
Ience per se.
My concern with realism, then, is a concern with truth; with what makes
a statement, or belief, true, if it is true. So I shall primarily be concerned-
in much of the chapter, at any rate-with the theory of truth.
144 Truth and Realism
In calling my principal topic the theory of truth, two clarifications should

be made. First, I am interested (roughly) in the 'meaning' of truth, not in
procedures or marks for telling which propositions are true. Classical de-
fenders of the coherence theory of truth probably failed to observe this dis-
tinction. They ran together coherence as a test of truth and coherence as a
definition of truth. As a test of truth, coherence has some attractions; as a
definition, it has no plausibility at all. I, of course, am only interested (here)
in the definition of truth. A theory of tests for truth is better understood as a
theory of justified belief.
The second clarification is that I am not concerned here with the formal
theory of truth, with an account of truth that resolves the liar paradox, for
example. Rather, I am only concerned with the metaphysical questions sur-
rounding the notion of truth. There are many philosophers who doubt
whether there are any interesting metaphysical questions, whether there is
anything informative that can be said about these questions. They maintain
that Alfred Tarski's work on the semantic conception of truth, or the idea of
truth as disquotation, resolves all the legitimate questions that can be raised
about truth. I believe that Putnam has given decisive arguments for reject-
ing this claim: the Tarskian, or disquotational, theory does not give us an
adequate analysis, or understanding, of the notion of truth.' Substantial,
classical questions remain, questions that divide realists from antirealists.
These are the ones that will engage my attention.
I have isolated one central dispute about truth that I shall address:
whether truth is verification-transcendent or not, whether it is a purely
nonepistemic notion or in some sense epistemic. This will be the topic of
the next section. But there is another important dispute about truth that
standardly divides realists from antirealists; that is, whether truth consists in
'correspondence' with reality. This is another way of thinking about realism
versus antirealism. Whether a realist account of truth needs to be commit-
ted to correspondence, and if so, what kind of correspondence, will have to
be explored. That will be the topic of section 7.3.
7.2. Truth as Nonepisternic

Putnam writes of two conceptions of truth: an externalist versus an intemal-
ist conception,' which he earlier dubbed "metaphysical realism" and "in-
ternal realism," respectively.7 Metaphysical realism sees truth as "radically
non-epistemic".8 Internal realism, by contrast, finds an epistemic compo-
nent in truth. I do not find any of Putnam's terminology optimal. In particu-
lar, his."internal realism" is best regarded as a form of antirealism. So I shall
count the nonepistemic approach to truth as realism sans phrase, and the
epistemic approach to truth will be counted as (a species of) antirealism.
How, exactly, would an epistemic account of truth go? As we will see, the
proponent of an epistemic approach is caught on the horns of a dilemma. If
a strong epistemic component is introduced, the resulting account is liable

to be unnecessay for truth. If he weakens the epistemic component, the
resulting conditions are liable to be i m . i e n t for truth. Before considering
some sample accounts, let us consider some dimensions along which episte-
mic theories of truth might vary.
First, there is a choice between actual verification and merely possible
verification. For example, constructivism is depicted as maintaining that a
mathematical statement is true only if we are (actually) in possession of a
proof of it, and false only if we are (actually) in possession of a refutation?
This can be contrasted with a weaker possible view, namely, that a mathe-
matical statement is true only if it is possible that we should possess a proof
of it.
Second, there is a choice between present verification and verification at
some time or other, for example, at some future time. C. S. Peirce defined
truth as "the opinion which is fated to be ultimately agreed to by all who
investigate."10 This is clearly a weaker requirement for truth than one
which requires, for p to be true, that p is agreed to now.
Third, there is a choice between conclusive verification and inconclusive,
or prima facie, verification. A very strong requirement for the truth of p is
that it be conclusively verified. A weaker requirement is that, for p to be
true, it must be supported to some degree. Of course, either of these re-
quirements can be weakened by taking the possibility option in the first
choice: p is true if it is merely possible for p to be conclusively verified.
The fourth choice concerns the person or persons for whom the proposi-
tion must be verified. Dummett regularly speaks with an indefinite first-
person plural: "we." In the constructivist's approach, cited above, a mathe-
matical statement is true only if we possess a proof of it. But who, exactly,
are "we"? All human beings? One or more human beings? Or should the re-
striction to human beings be lifted entirely?
Clearly, many permutations of the epistemic approach to truth can be
generated from these options. Let's quickly look at some sample accounts .
and the difficulties they encounter.
(a) All of us now have conclusive evidence for p.
(b) Somebody now has conclusive evidence for p.
(c) Somebody will some day have conclusive evidence for p.
(d) Somebody now has prima facie evidence for p.
(8) Somebody will some day have prima facie evidence for p.
( It is possible for somebody to have prima facie evidence for p.
(g) It is possible for somebody to have conclusive evidence for p.
Considered as necessary and sufficient conditions for 'p is true', each of
these 'suffers from palpable defects. Each is either too weak or too strong;
some of them are both.
Let us examine them as conditions for the truth of ordinary physical ob-
ject propositions. Clearly, (a) is much too strong. It could certainly be true
that Julius Caesar had a mole on the nape of his neck although it is false that
all of us now have conclusive evidence for this. Similarly for (b):this could
be true even though nobody now has conclusive evidence for p. Even (c) is
still too strong. This minor historical fact about Julius Caesar may be one for
which nobody will ever have conclusive evidence. Even (d) is too stringent.
Perhaps nobody now has even prima facie evidence for this fact'. This would
not keep the proposition from being true.
Not only is (d) too strong, it is also too weak. Suppose some historian does
have prima facie evidence for this proposition. That surely does not entail
that the proposition is true. For example, suppose a Roman text of the pe-
riod attributes a remark to Brutus that seems to indicate the presence of a
mole on Caesar's neck. That evidence is not sufficient for the truth of the
proposition. The same twin problems arise for (e), for similar reasons. It is
both too strong and too weak.
Turning to the 'possibility' formulations, (f3 is too weak for the same rea-
sons as (d) and (e). The mere possibility of prima facie evidence for p does
not guarantee p's truth. The difficulties for (g) are slightly harder to assess.
They depend on how 'conclusive evidence' is interpreted. If it is interpreted
very strictly-so that conclusive evidence entails truth-then it is doubtful
whether it is possible for somebody to have conclusive evidence for any
proposition, at least any physical object proposition. So (g) would be too
strong. If 'conclusive evidence' is interpreted less strictly, so that it does not
entail truth, then (g) is too weak.
There is another plausible objection against (g), if it is interpreted strictly.
On the strict interpretation we are to understand 'p is conclusively verified
(for some person S)' to entail (by virtue of its meaning) 'p is true'. But if this
is right, truth is covertly contained in the analysis of 'p is conclusively veri-
fied'. And then it would be circular to employ the latter locution in the
analysis of 'p is true'!
This point is worth underlining because it should also figure in the assess-
ment of the epistemic account of truth endorsed by Putnam. Putnam
writes:
Truth is an idealization of rational acceptability. We speak as if there were
such things as epistemically ideal conditions, and we call a statement 'true' if
it would be justified under such conditions."
To clarify this proposal, let me reformulate it as follows.

(h) -- If someone were in epistemically ideal conditions vis-A-vis the question
of p versus not-p, then he would be justified in believing p.
The correctness of this proposal depends on how 'epistemically ideal

conditions' is defined. Is there a way to define it so that the equivalence is
correct? There is a definition, I think, in which the equivalence could suc-

ceed. Suppose we define 'S is in epistemically ideal conditions vis-A-vis p' as:
S is so situated and so constructed that he would believe p if and only if p is
true. Being in epistemically ideal conditions vis-a-vis p might consist, for
example, in (1)being in a perfect position for observing p, (2) having perfect
detector organs vis-a-vis states of affairs like p, and (3)having belief-forming
processes perfectly coordinated with these organs' outputs. On a suitable
theory of justifiedness, then, proposal (h) might be right.
But this stratagem obviously incurs the sort of problem we posed for (g;)
under the strict interpretation. It makes the equivalence work only by
making it circular. The problem, of course, is that the proposed definition of
'epistemically ideal conditions' reintroduces the notion of truth!
Putnam obviously cannot use this way of defining 'epistemically ideal
conditions'. He hints at this himself when he restricts 'verified' to an "opera-
tional" sense.12 Is there an alternative way of defining 'epistemically ideal
conditions' so that (1)no nonepistemic notion of truth is presupposed, and
yet (2)(h) turns out to be equivalent to 'p is true'? That has yet to be shown.
I doubt that there is any such way. In short, there is no reason to think that
Putnam's epistemic equivalence can succeed unless its crucial epistemic
notion is cashed out circularly, in terms of the (realist) notion of truth.
The difficulty facing Putnam's proposal is a specimen of a difficulty facing
any epistemic approach to truth. How are the epistemic notions appearing
in any such account themselves to be understood? How are 'justified', 'veri-
fied', 'rationally acceptable', or other such terms to be analyzed? Putnam
and Dummett seem to take these terms as primitives. But I find this highly
counterintuitive. Epistemic notions strike me as far more in need of expli-
cation than truth. Furthermore, given the plausibility of my sort of theory
of justification-a truth-linked theory-the onus is on Putnam, Dummett,
or like-minded theorists to show that these notions have non-truth-linked
explications.
Let me not rest this contention on the details of my theory of justified- .
ness. I can argue more broadly for the necessity of having recourse to truth
in explicating epistemic standards. First, as I have noted, many epistemic
notions countenance degrees of strength, for example, degrees of confirma-
tion, weight of evidence, and the like. What accounts for these degrees?
What is the continuum on which these gradations occur? The only plausible
answers, I think, would appeal to truth. My own theory of justifiedness ap-
peals to true-belief-conducive psychological processes (and methods). But
other theories might appeal to an evidential link between evidence and hy-
pothesis. One has stronger evidence for p to the extent that the evidence
makes p more likely. But 'more likely' must mean 'more likely to be true'. Of
course, we could define a technical notion of confirmation that has no com-
mitment to truth, where the truth of the evidential propositions has no link-
age-logical, probabilistic, subjunctive, or what have you-with the truth
of the hypothesis. But in that case who would suppose that this technical
notion has anything to do with the intuitive notion of evidential support?
Second, take the notion of epistemic 'access' to a state of affairs, such as
perceptual access or memory access. Again we have the notion of better
and worse epistemic situations. And again this can only be spelled out in
terms of truth. Perceptual access to a situation is better just in case it is eas-
ier, in that situation, to form true beliefs about the objects or events in the
scene. The optimal distance and viewing conditions are those that promote
the greatest chance of truth and accuracy.
Third, the history of science and philosophy is filled with controversies
over methodology. Contending schools have vied over the right procedures
for fixing belief. It is hard to make sense of these debates save as debates
over which methodology is most conducive to true convictions. However
difficult it may be to settle such questions, they seem to presuppose the no-
tion of truth. So truth is conceptually prior to any operational specification
of epistemic methods or standards. Antirealists, who seek to define truth in
epistemic terms, are proceeding precisely backwards.
We must not ignore, however, an underlying impetus behind many
moves toward an epistemic account of truth. Antirealists are doubtless
worried about the epistemological consequences of a realist conception of
truth. If truth is definitionally detached from evidence, either actual or
possible, is there really any prospect that we can have epistemic access to
the truth? Isn't there a threat of skepticism? Some antirealists might be
prepared to concede that the ordinary, naive conception of truth is none-
pistemic. But they may feel that replacement of this naive conception is re-
quired to avert epistemic catastrophe.
Given this problem, we must determine whether candidate epistemic ac-
counts of truth really help, whether antirealist proposals really improve our
epistemological opportunities. The answer is: not necessarily. Putnam's
proposal, (h), is a case in point. Truth is there equated with what people
would be justified in believing if they were in epistemically ideal conditions.
But since none of us ever is in epistemically ideal conditions with respect to
any proposition, how are we supposed to tell-from our actual stand-
point-what a person would be justified in believing in that hypothetical
situation? To assess the truth of p according to this proposed equivalence,
we would have to assess the truth-value of the indicated subjunctive condi-
tional. But that is no mean task! If anything, it imposes an even nastier
epistemic predicament than the one feared from realism.
The nonepistemic nature of truth is well described by Brian ~ 0 a r .Our l~
ordinary (realist) conception of truth can best be understood, he says, in
modal terms. The idea of a proposition's being true is the idea of a state of
affairs such that it could happen (or could have happened) that it be true
even though we are not in a position to verify it. Furthermore, he asserts, it
is part of our theory of nature that this is generally so. Whether p concerns
objects in the garden, in Antarctica, or in the Andromeda galaxy, p's veri-
fiability by us depends upon various contingent circumstances not entailed
by laws of nature. For us to verify such propositions, light, sound, or some
such energy must reach our sense organs. But the regions of space through
which such energy must travel might have (or might have had) distorting
properties. There might be black holes in the relevant regions that 'soak up'
the light. Also, for us to verify such propositions (indeed, any propositions),
our brains must be constructed in certain ways. But they might not have
been suitably constructed. So it might have happened that these proposi-
tions are true although we are not in a position to verify them.
As this way of explaining truth indicates, a sharp distinction must be
drawn being a proposition's being true and the proposition's being verified.
The latter, but not the former, involves processes by which the truth is de-
tected or apprehended. (Here I have in mind both extracognitive pro-
cesses-for example, the traveling of light from object to retina-as well as
intracognitive processes.) Indeed, only such a sharp distinction-character-
istic of realism--can make good sense of certain of our verifying
procedures.
In seeing how realism makes sense of verification, I follow the suggestions
of William Wimsatt, who in turn credits Donald Campbell and R. ~evins.'"
Our conception of reality is the conception of something robust, an object
or property that is invariant under multiple modes of detection. The use of
multiple procedures, methods, or assumptions to get at the same putative
object, trait, or regularity is commonplace in ordinary cognition and in sci-
ence. Wimsatt lists many examples, including the following: (a) we use dif-
ferent perceptual modalities to observe the same object, or the same sense
to observe the object from different perspectives and under different obser- .
vational circumstances; (b) we use different experimental procedures to (try
to) verify the same alleged empirical relationships, that is, alternative pro-
cedures are used to 'validate' one another; (c) different assumptions, models,
or axiomatizations are used to derive the same result; (d) we use agreement
of different tests, scales, or indices as a sign of validity of a trait-construct;
and (e) we seek matches of theoretical description of a given system at dif-
ferent levels of organization. In short, our verification procedures, when
they are careful, seek to 'triangulate' on the objects or relationships under
study.
We can best make sense of a need for triangulation on the assumption
that the truths, or facts, about the object or system under study are sharply
distinguished from the processes of verification or detection of them. Any
particular method might yield putative information partly because of its
own peculiarities-its biases or distortions. This is a potential danger as
long as the properties of the studied object are distinct from the process of
detection. Triangulation hopes to correct for such biases on the assumption

that independent methods will not have the same distorting characteristics
(if any at all). Obviously, we cannot use no method at all to make observa-
tions, to 'get at' the system or relationship under investigation. So we run
the risk of getting mere artifacts of a measurement or observation process,
rather than the real character of the object. Using multiple methods of de-
tection minimizes this risk. In general, we try to get away from the idiosyn-
cracies of the verification procedure in order (better) to get at the
verification-independent properties of the target object. All this makes
sense on the assumption that there are truths about the object independent
of this or that verification procedure.
Admittedly, this realist stance is not the only possible diagnosis of the
need for triangulation. Triangulation may be viewed as a variant of coher-
entist methodology; and coherentism is an epistemology fully available to
the antirealist. However, the antirealist must take coherentism simply as a
brute fact about our epistemic procedures. He cannot offer any principled
rationale for triangulation, of the sort available to the realist. The realist, by
contrast, can explain the fallibility of our detection procedures, and hence
the need for triangulation, in terms of the distinction between the real facts
and the verification processes used to get at them.
As an aside, it is noteworthy that our perceptual systems seem 'designed'
to mark properties of the distal stimulus as opposed to those of the organic
process of detection. So-called object constancies are a case in point. A
moving observer does not see the stationary stimulus as moving or varying
in size, though such displacements do occur at the retinal level. The percep-
tual system suppresses these displacements, apparently on the presumption
that they are observer-induced changes. What the system seeks are proper-
ties of the observation-independent object. When input is diagnosed as re-
sulting from observer characteristics-in the present case, observer
movement-it simply is not communicated to higher cognitive levels.
Readers of Dummett may feel that I have thus far neglected a principal
reason for a verificationist account of truth. Dummett approaches the topic
of realism and antirealism from the context of a theory of meaning or un-
derstanding. In that context the question arises whether it is possible for us
to understand, or grasp, any verification-transcendent conception of truth.
Do we have a capacity to understand statements whose truth transcends
their verifiability? If a proper theory of understanding, or meaning, pre-
cludes such a capacity, then no sense can be made of a realist, verification-
transcendent, account of truth.'' And none of the foregoing arguments can
carry any weight.
I cannot enter in any depth into the theory of meaning or understanding.
That clearly exceeds the scope of this work. Nor can I comment in detail on
Dummett's proposals concerning a "use" theory of meaning, and the re-
quirement that uses must be "manifestable" in behavior. I am inclined to

regard these proposals with some skepticism, but I will not venture into this
tricky terrain.16 Instead, 1 put forth the following argument.
The claim that we cannot mean or understand verification-transcendent
statements can only refer to verification by us, that is, by human beings. But
as I have already argued, in the case of many if not most physical object
statements, their truth certainly appears to be possible independently of
human verification. For example, it might be true that such-and-such hap-
pened in the Andromeda galaxy although no human beings were (or are) in
a position to verify it. (To hold otherwise would involve an untenable form
of epistemic speciesism.) Moreover, this modal fact seems far more certain
than any (interesting) doctrine in the theory of meaning. So if we are con-
fronted with a choice between this modal fact and a meaning doctrine that
excludes verification-transcendent meaning, we are better advised to doubt,
or reject, the meaning doctrine.
7.3. Truth and Correspondence

I have now completed the discussion of whether truth is an epistemic no-
tion. I have defended the realist position that it is wholly nonepistemic. In
this section I turn to another facet of the debate over truth: whether truth
consists in 'correspondence'. Antirealists are often anxious to pin a corre-
spondence doctrine on their foes, because they find grave difficulties with
that doctrine. But it is not clear that a realist must be committed to corre-
spondence; that depends on what, exactly, is meant by 'correspondence'.
Realism is in trouble on this score only if there is an untenable version of the
correspondence doctrine to which realism must adhere.
One version of a correspondence doctrine, the Tractarian version, says
that the world is a totality of facts and that a proposition is true just in case
it corresponds with a fact." A familiar objection to this is that the world
does not contain factlike entities, the kinds of entities that would exactly
correspond to sentences or propositions. It is implausible to suppose that
there are disjunctive facts in the world, existential facts, conditional facts,
universal facts, and so on. Language or thought constructs disjunctive state-
ments, existential statements, conditional statements, and the like. But it is
misguided to suppose that structures of this sort constitute the world. If the
correspondence doctrine is committed to this picture, it is an untenable
doctrine.
The foregoing correspondence doctrine portrays the world as being pre-
structured into truthlike entities. This is one thing the antirealist opposes. A
closely related view, also opposed by (some) antirealists, is that the world
comes prefabricated in terms of categories or kinds. This is another facet of
the correspondence theory to which Putnam objects.
For an internal realist like myself . . . 'objects' do not exist independently of

conceptual schemes. We cut up the world into objects when we introduce
one or another scheme of description . . . [For an externalist the world con-
tains] Self-IdentifyingObjects, . . . the world, and not thinkers, sorts things
into kinds.18
The point here is essentially a Kantian point, and one also stressed by Nel-
son ~ 0 o d m a n .The
l ~ creation of categories, kinds, or 'versions' is an activity
of the mind or of language. The world itself does not come precategorized,
presorted, or presliced. Rather, it is the mind's 'noetic' activity, or the es-
tablishment of linguistic convention, that produces categories and cate-
gorial systems. When truth is portrayed as correspondence, as thought or
language mirroring the world, it is implied that the world comes precate-
gorized. But that, says the antirealist, is a fiction.
Let us use 'correspondence,' for the constellation of views just canvassed,
namely, that the world is prestructured into truthlike entities (facts), and
that truth consists in language or thought mirroring a precategorized world.
Although this is one conception of correspondence, a realist need not be
wedded to it. Weaker variants of correspondence are still compatible with
realism. Let me try to sketch--admittedly metaphorically-a less objec-
tionable style of correspondence theory, to be called correspondence^.
The mirror metaphor is only one possible metaphor for correspondence.
A different and preferable metaphor for correspondence is fittingness: the
sense in which clothes fit a body. The chief advantage of this metaphor is its
possession of an ingredient analogous to the categorizing and statement-
creating activity of the cognizer-speaker. At the same time, it captures the
basic realist intuition that what makes a proposition, or statement, true is
the way the world is.
There are indefinitely many sorts of apparel that might be designed for
the human body, just as there are indefinitely many categories, principles of
classification, and propositional forms that might be used to describe the
world. Although the body certainly has parts, it is not presorted into units
that must each be covered by a distinct garment. It is up to human custom
and sartorial inventiveness to decide not only what parts to cover, but what
types of garments should cover which expanses of the body, and whether
those garments should be snug or loose. For many bodily parts (or groups of
contiguous parts), there is a wide assortment of garment-types used to
clothe them. In outer footwear, for example, there are sandals, slippers,
shoes, tennis shoes, snowshoes, fishing boots, and hiking boots. In inner foot-
wear there are anklets, calf-length, and over the knee stockings. Among
trousers there are shorts, knickers, and full-length pants. For the torso there
are short-sleeve and long-sleeve shirts, jackets, vests, tunics, robes, saris,
capes, and huipils.
Despite all this variety-humanly invented variety-there is still the
question, for any specified type of apparel, whether a specific token of that
type fits a particular customer's body. This question of fittingness is not just
a question of the style of garment. It depends specifically on that customer's
body. Similarly, although the forms of mental and linguistic representation
are human products, not products of the world per se, whether any given
sentence, thought sign, or proposition is true depends on something extra-
human, namely, the actual world itself. This is the point on which realists
properly insist.
In inventing or evolving sartorial styles, people devise standards for
proper fittingness. These may vary from garment to garment and from fash-
ion to fashion. Styles specify which portions of selected bodily parts should
be covered or uncovered, and whether the clothing should hug snugly or
hang loosely. This is all a matter of style, or convention, which determines
the conditions of fittingness for a given type of garment. Once such fitting-
ness-conditions are specified, however, there remains a question of whether
a given garment token of that type satisfies these conditions with respect to
a particular wearer's body. Whether it fits or not does not depend solely on
the fittingness-conditions; it depends on the contours of the prospective
wearer as well.
The case of truth is quite analogous. Which things a cognizer-speaker
chooses to think or say about the world is not determined by the world it-
self. That is a matter of human noetic activity, lexical resources in the
speaker's language, and the like. A sentence or thought sign, in order to
have any truth-value, must have an associated set of conditions of truth. Ex-
actly what determines truth-conditions for a sentence or thought sign is a
complex and controversial matter. But let us assume that a given utterance
or thought, supplemented perhaps with certain contextual factors, deter-
mines a set of truth-conditions. The question then arises whether these con-
ditions are satisfied or not. The satisfaction or nonsatisfaction of these
conditions depends upon the world. Truth and falsity, then, consists in the
world's 'answering' or 'not answering' to whatever truth-conditions are in
'
question. This kind of answering is what I think of as 'correspondence,'. No-
tice that which truth-conditions must be satisfied is not determined by the
world. Conditions of truth are laid down not by the world, but only by
thinkers or speakers. This is the sense in which the world is not precate-
gorized, and in which truth does not consist in mirroring of a precate-
gorized world. Unlike correspondencel, correspondencez is compatible
with what is correct in the constructivist themes of Kant, Goodman, and
Putnam.
While these philosophers are right to emphasize constructivism, they
carry it too far. There is a strong suggestion in Goodman's writing that
"versions" are all there is; there is no world in itself. Although Goodman
speaks of truth as being a matter of "fit," it is not fitness of a version to the
world, but fitness of a version to other versions.* This, of course, a realist

cannot accept. To return to our sartorial metaphor, the realist insists on the
existence of a wearer: clothes don't make the world.
A chief motivation for the Goodman-Putnam view is epistemological.
One strand of their position holds that we can never compare a version with
"unconceptualized reality," so as to determine whether the world answers
to a thought or statement. Comparison of a theory with perceptual experi-
ence is not comparison with unconceptualized reality because perceptual
experience is itself the product of a sorting, structuring, or categorizing pro-
cess of the brain. So all we can ever do, cognitively, is compare versions to
versions.'l
Conceding this point, however, does not undermine realism. Since corre-
spondence' does not embrace the mirroring idea-nor the idea that true
thoughts must resemble the world-the epistemology of getting or deter-
mining the truth need not involve comparison.
Can the same problem still be posed for my theory? It may be reex-
pressed by saying that, on my theory, we can never grasp or encounter the
world, so as to determine whether some thought or sentence of ours fits it.
And if the realist's world is unconceptualized,how can it be grasped or en-
countered in a manner to determine fittingness?
I see no insuperable difficulty here. Perception is a causal transaction
from the world to the perceiver, so perception does involve an encounter
with the world (at least in nonhallucinatory cases). To be sure, the event at
the terminus of the transaction does not resemble the event at the starting
point. The terminus of perception is a perceptual representation, which in-
volves figure-ground organization and other sorts of grouping and structur-
ing. The initiating event does not have these properties. Still, the
transaction as a whole does constitute an encounter with something uncon-
ceptualized. We are not cut off from the world as long as this sort of en-
counter takes place.
But is this sort of encounter s d c i e n t for knowledge or other forms of
epistemic access? As far as I can see, realism about truth does not preclude
such knowledge. Suppose that the (unconceptualized) world is such that the
proposition 'There is a tree before me' fits it, that is, is true. And suppose
that the perceptual process is a reliable one, both locally and globally.
Then, according to my account of knowledge, I may indeed know that there
is a tree before me. The world that I learn czbout is an unconceptualized
world. But what I learn about this world is that some conceptualization (of
mine) fits it. How I learn this is by a process that begins with the unconcep-
tualized world but terminates in a conceptualization.
Does this (realist) theory make the world into a nownenal object, an o b
ject that cannot be known or correctly described? Not at all. On the pro-
posed version of realism we can know of the world that particular
representations fit it. So the world is not a noumenal object.
I cannot leave Putnam's critique of realism without commenting on a

central component in his antirealist argument on which I have thus far been
silent. In arguing against a correspondence theory of truth, Putnam has
pointed out that there are too many correspondences. Correspondence has
traditionally been construed as a word-world relationship. Putnam con-
tends that there are too many word-world relationships of the requisite sort.
Briefly, while there may be one satisfaction relation under which a given
sentence turns out true, there will also be other satisfaction relations under
which it turns out not to be true. So for any word-world relation purporting
to be the 'intended' truth relation, there are other, equally good candidates.
Since no unique word-world relation can be identified with truth, the cor-
respondence notion of truth is untenable."
Actually, the multiplicity of relations Putnam discusses are often candi-
dates for the reference relation rather than the truth relation. But since the
principal topic of interest is truth, the significance of his attack on a deter-
minate reference relation lies in its implications for truth. What we have,
then, is an indeterminacy argument against a word-world relation of truth,
which is taken by Putnam to be a general critique of ('metaphysical')
realism.
I believe there are two sorts of replies to this critique. First, one can
challenge the indeterminacy argument by denying that Putnam has taken
all relevant parameters into account. This point has been argued persua-
sively by Alvin ~ l a n t i n ~ a .Putnam
" seems to suppose that the terms of our
language get their meaning or extensions, somehow, by virtue (solely) of the
set theoretical models of first-order formalizations of the body of our beliefs.
In other words, roughly, our terms get their meaning and extensions by vir-
tue of a vast network of implicit definitions. But this theory is extremely
dubious. There must be further constraints at work in fixing our meanings
and extensions. So the multiplicity that Putnam identifies of possible assign-
ment functions to our terms does not have the implications that he alleges.
Second, the problem posed by Putnam, if it is a serious problem, is not
really a problem about truth. Rather, it is a problem about interpretation,
or the establishment of truth-conditions. These must be distinguished.
Questions of truth cannot arise until there is a suitable bearer of truth-value,
with an established set of truth-conditions, about which it can be queried
which truth-value it has. Sentences or thought events construed as mean-
ingless marks or nerve impulses are not bearers of truth-values. Only when a
sentence or thought event is interpreted-when it has suitable semantic
properties (including reference of singular terms and sense or reference of
general t e r m s k i s it even a candidate for being true or false.
Now Putnam has worries, as do other theorists in the field, about how,
exactly, words and thought signs get their meaning and reference. How, in
other words, do they get truth-conditions attached to them? These are im-
portant issues, which I will not in any way try to resolve. But unless and
until sentences and thought signs are conceded to have interpretations, or

truth-conditions, the question of truth cannot even arise. However, once we
concede a determinate interpretation, or set of truth-conditions, for sen-
tences and thought signs, we have assumed that Putnam's problems can be
resolved. We can then turn to the question of truth, to the possession of this
or that truth-value. I see no insurmountable obstacle here to the realist the-
ory I have sketched. Given truth-conditions for a sentence, or thought, what
makes it true or false is surely the way the world is, or whether it fits the
world.
7.4. Scientijk Realism

The doctrine of realism discussed thus far is a doctrine about truth. It is the
doctrine commonly called 'metaphysical realism'. But another epistemolo-
gical issue concerning realism is that of scientific theories and theorizing.
Whereas scientific theories are ostensibly designed to describe the world,
frequently the unobserved 'fine structure' of the world, some philosophers
of science contend that this is an improper or injudicious portrayal of scien-
tific theories. These philosophers-variously known as instrumentalists, fic-
tionalists, and e m p i r i c i s t d e n y that theories purport to give literally true
stories of what the world is like. Rather, they are just calculational instru-
ments for predicting the course of experience, devices to predict (and per-
haps explain) what is empirically observed. Denying the face-value
meaning of theoretical statements, philosophers of this ilk try to convey the
meaning of these statements in purely observational terms, or other con-
structions out of empirically acceptable materials. Against such reduction-
ist maneuvers, scientific realists hold out for a literal interpretation of
theoretical statements. Acceptance of such statements is seen as an ontolo-
gical commitment to entities ostensibly designated by their theoretical
terms. And such ontological commitment is viewed as epistemologically
conscionable.
I construe scientific realism (henceforth 'SR') as having a core consisting
of three interconnected theses. First, there is a semantic thesis: SR says that
theoretical statements in science are to be interpreted literally and referen-
tially. Their meaning should not be purged of its (ostensible) extra-empirical
content. Second, there is a thesis concerning the aim of scientists: SR says
that what theorists are (generally) trying to do is to find and believe true
descriptions of the world. Third, SR advances the epistemological claim
that this aim is legitimate, since scientists sometimes are justified in believ-
ing scientific theories so construed. Or, in a slightly weaker variant, it is a
legitimate aim because it is (humanly) possible for scientists to have justified
belief in scientific theories realistically construed. (I will address both vari-
ants of the third thesis, but will not always announce switches from one to
the other.)24
Although I have supported metaphysical realism as a theory of truth, the

question of SR is quite independent. One could be a realist about truth
while opposing SR. One could be, say, a reductionist about theoretical
statements, but go on to say that the truth of any theory, given such a re-
duction, still depends on how the world is, nonepistemically. For example,
the theory's truth would depend on whether the world really has a propen-
sity to obey the specified observational regularities.25I am inclined to be a
scientific realist as well as a metaphysical reahst, as intimated in Chapter 6.
But these doctrines are distinct. Moreover, I am more committed to meta-
physical realism--especially its view of the nonepistemic nature of truth.
My epistemology does not hinge so much on the SR issue. But to clarify my
general stance, a brief excursus into SR is appropriate.
In the writings of (early) Putnam, Richard Boyd and W. H. Newton-
Smith, SR has been given a new twist, rendering it a stronger doctrine than
the core view I presented above.26SR is depicted as holding not merely that
theoretical terms purport to refer, but that they (often) succeed in referring.
Not only do scientists aim to describe the world with their theories, but
they succeed in doing so. Or, more cautiously, their mature theories gradu-
ally get closer to the truth. In short, SR is viewed not merely as a doctrine
about scientific aims, but as one about scientific success. (By 'success' I
mean not merely predictive success, but success in delineating the fine
structure of the world.)
This brand of SR is sometimes called 'convergent realism', or 'historical
progressivism'. What I wish to emphasize about this doctrine is that it is an
addition to the core of SR; one can therefore embrace SR's core without
embracing convergent realism. This is the posture I mean to assume. More
precisely, although I feel some attraction to convergent realism, at least in a
modest, scaled-down version, I recognize that the historical record, viewed
from our present vantage point, is not entirely clear." Morever, the thesis
of convergent realism is at a different epistemological level than the core of
SR. And in the present context I want to remain neutral about that level. ,
Am I really entitled to accept SR's core while remaining neutral about

convergent realism? As I shall indicate shortly, there is reason to hesitate on
this point. To assess this matter, we need to look back at the three core the-
ses of SR. The semantic thesis of SR certainly does not depend on historical
progressivism. The meaning of theoretical sentences may exceed their em-
pirical content however scientists have been faring at limning the structure
of reality. The second thesis of core SR, the aim component, also does not
imply convergent realism. It may be true that scientists typically intend
their theories to get at fundamental, transobservational truths; but they
may nonetheless be falling flat, or achieving only modest success, in this
endeavor.
It is only when we move to the third thesis of core SR that convergent
realism may be implicated. If scientific theorizing has indeed fallen flat, his-
torically speaking, how can it be epistemically legitimate for scientists to

aim at getting theoretical truths? How is it possible for them to be justified
in believing scientific theories so construed? This seems to be a particular
problem given my theory of justifiedness. After all, if beliefs are justified
only when caused by reliable processes, and if scientific methodology has a
poor historical track record, the processes in question must not be reliable.
But then belief in scientific theories, at least construed realistically, is not
and cannot be justified. So the justifiedness of accepting scientific theories
appears to depend on the truth of convergent realism, or at least a good au-
wage record of true theoretical beliefs, even if there is no continual forward
progress.
The problem is crystallized in what Putnam calls the "disastrous meta-
induction" of scientific t h e o r i ~ i n Suppose
~.~ scientists decide that no term
in the science of more than fifty years ago referred, and no theory of that
vintage was true. If this keeps happening, shouldn't we be led to the meta-
induction that no theoretical term now or in the future will refer, and no
present or future theory will be true? From this, conjoined with my theory
of justification, won't it follow that we are not, and could not be, justified in
believing any theoretical statement?
No, there are several lacunae in this line of reasoning. Let us look first at
the "disastrous meta-induction" itself. Upon inspection, the meta-induction
is really self-undermining. We judge past theories to be false only by the
lights of our present theory. If we abandon our present theory, we are no
longer in a position to judge past theories false. So if we use the meta-in-
duction to conclude that no present scientific theory is true, we thereby
eliminate all grounds for believing that past theories were false. But then
we are no longer entitled to believe the premises of the meta-induction.
But, it might be replied, couldn't those premises still be true? And if so,
wouldn't this reinstate the difficulty for any prospect of justified theoretical
belief? Let us see, then, whether the meta-induction might be vulnerable on
other grounds. I think it is. In deciding whether we can inductively infer the
falsity of present and future theory from the falsity of believed theories of
the past, we have to assess the bases of those past theoretical beliefs. Scien-
tists in previous eras may have used different-and poorer-methodologies
than we now have available. After all, there have been dramatic and con-
tinued developments in statistics and other evidential procedures. Further-
more, we now have vastly more powerful instruments of discovery and
detection in many domains. Perhaps these differences in instrumental and
evidential tools can outweigh past failures. Thus, our present scientific situ-
ation rnay not be relevantly similar to past science. Suppose you are going
on a long automobile trip, and on all your previous long trips you had a flat
tire. Should you expect a flat on this trip too? Not necessarily. If you began
all previous trips with bald tires, but this time you start with a freshly pur-
chased set of tires, the induction would be weak. (This response may not be
conclusive, but it should give us pause.)
Setting aside the force of the meta-induction, we need to look more care-
fully at the implications of my theory of justification for the present ques-
t i ~ nThere
. ~ ~ are several reasons why a poor track record of positive belief
in scientific theory need not ruin scientists' chances for justified theoretical
belief according to my account. By 'positive' theoretical belief let us mean a
belief that a theory is true; by 'negative' theoretical belief let us mean a be-
lief that a theory is false, that is, a belief in the denial of a theory. Since the
same methods are used to arrive at negative theoretical beliefs as positive
theoretical beliefs, even if scientific methodology (assuming, inaccurately,
that a single methodology has continuously been in use) has consistently
generated false positive theoretical beliefs, it may still have generated lots
of true negative theoretical beliefs. Moreover, the latter class is probably
much larger than the former. Think of all the scientific theories we now re-
ject and of the innumerable theories that past scientists rejected. When the
entire class of outputs of scientific methodology is included, both negative
and positive, the overall truth ratio may be rather impressive.
The next point concerns the distinction between primary justifiedness
(P-justifiedness) and secondary justifiedness (S-justifiedness). Recall that
P-justifiedness turns on the truth ratio of basic psychological processes; and
S-justifiedness turns on properties of methods. Now the basic psychological
processes used in assessing scientific theories probably are no different from
those used for many other tasks; only the methods, at most, are distinctive.
So even if these processes have a poor track record in this very special do-
main, it does not follow that they have poor reliability in general.
When we turn to S-justifiedness the matter is up in the air, because I did
not fully commit myself to a criterion of J-rightness at the secondary level.
While I am certainly attracted to an objectivist, truth-ratio criterion here
too, I feel a tug in the direction of a culturally relative criterion (see section
5.3). Perhaps a belief is S-justified if it results from methods that are ap- ,
proved by a relevant community, say a community of experts. This might

suffice, then, to make present positive scientific beliefs S-justified as well as
P-justified.
Furthermore, we should not forget the wider family of truth-ratio cri-
teria presented in section 5.5. This family included not only the absolute
resource-independent criterion, (ARI), but also two resource-relative cri-
teria. There was the maximizing resource-relative criterion and the satisfic-
ing resource-relative criterion. Either of these approaches could be
extended to S-justifiedness. For example, a method could be counted as
right if it has the highest truth ratio among all available methods. This
comparative criterion might deem some of our present scientific methods to
be right even if their absolute truth ratio is not terribly high.
These points indicate that, within my full theory of justification, a poor

track record of scientific beliefs--especially positive theoretical belief*
need not preclude the possibility of justified belief in scientific theories.
That sfices to show that core SR is independent of convergent realism.
However, a few further comments are in order to guarantee a fair-minded
assessment of just how good or bad the track record of theoretical beliefs
may be.
One point is that philosophers and historians of science have a tendency
to concentrate on the most encompassing theories in science, those at the
highest level of generality, which also typically postulate entities or forces
that are least amenable to observation. Philosophers and historians of sci-
ence are preoccupied with quantum theory, or problems of evolutionary
theory. They spend little time citing more prosaic theories, such as the the-
ory that the liver's function is to detoxify the body. It may well be true that
once-accepted theories at the highest level of generality are frequently
found in hindsight to be mistaken. But it is unfair to restrict consideration to
these kinds of theories. Lower-level theories also occupy scientists, and
these typically have much greater staying power.
We should also notice that any large-scale theory normally has numerous
components, and scientists who believe the whole theory also believe its
several components. Even when the theory as a whole is subsequently over-
thrown, the several components are not necessarily all overthrown. The
general theory of the electron has certainly changed over time. But the
view of the electron as a basic unit of electrical charge, not resolvable into
smaller units, has been retained. That component has stayed constant. If, in
this fashion, we take account of beliefs in all components of theories, not just
entire theories, the specter of a "disastrous meta-induction" is much less
forbidding. It does not look, from our present vantage point, as if science
has been doing so poorly at
One final point. I have been trying to show how, given my account of
justification, it could well turn out that positive belief in scientific theories is
justified. But it is not critical to my epistemology that this be so. Maybe a
large-scale scientific theory is such a risky intellectual commitment, and the
alternative possibilities are so numerous, that scientists are not entitled to
believe such theories. This admission would not lead to general skepticism,
nor to any general problem for my account of justification. Even if global
skepticism were fatal to a theory of justification (which is not obvious), it is
not so objectionable for a theory to imply local skepticism, that beliefs in
one special arena are unjustified. There is nothing in my account of justifi-
cation, then, that forces me to endorse SR, especially its third thesis. Scien-
tific realism plays a less essential role in my view than metaphysical realism.
Still, for the reasons indicated, I am inclined to accept it, and to believe in
its compatibility with other tenets of my epistemology.
I draw this discussion to a close, having shown why realism about truth is
a pIausible and defensible doctrine. Truth is not an epistemic notion; and in
a properly softened form, even a correspondence theory of truth is tenable.
We have seen how core scientific realism, while distinct from realism about
truth, is an attractive view, and ultimately compatible with my general
epistemology.
chapter 8
The Problem of Content
8.1. Epistemology and Content

Having defended a realist, nonepistemic conception of truth, I am almost
ready to proceed to Part 11, which elaborates the projected relation be-
tween primary epistemology and cognitive science. However, there is one
last 'foundational' issue that needs to be broached before turning to Part 11,
namely, the problem of content.
My epistemology, like most others, assumes the existence of beliefs and
other propositional attitudes. A propositional attitude, of course, is a mental
state with (propositional) content. But recent literature abounds with wor-
ries and doubts about contents, especially the possession of contents by
mental states. These worries are directly relevant to this book, partly be-
cause they challenge one of my basic presuppositions, and partly for other
reasons to be explained below. Unfortunately, many of these issues lie at the
heart of philosophy of mind and philosophy of language, and I can only
hope to paint the landscape with very broad brushstrokes. Still, I must re-
connoiter the territory to inspect the possible pitfalls. I must consider, if
only broadly and provisionally, whether the alleged threats are genuine,
and if so, whether and how they can be met.
To say that something has content is to say that it has semantic proper-
ties: meaning, reference or truth-conditions, for example. Given my episte-
mological perspective, truth-conditions are especially important. Unless
mental states have truth-conditions, there can be no true or false beliefs; a
fortiori there can be no mental processes with the epistemological proper-
ties that interest us, such as, power and reliability. My investigation of such
properties of mental processes could not get started: it would be devoid of
relevant subject matter.
It should be emphasized that the absence of content-specifically, truth-
conditions-would be a problem for most epistemologies. Any epistemol-
ogy that takes knowledge seriously must take true belief seriously. Even an
epistemology that ignores knowledge and truth-values, but instead stresses
interpersonal agreement and disagreement, must take semantic content

seriously. For how can there be agreement or disagreement without the
sharing of content?
It would be hasty to assert that every epistemology needs the notion of
content, especially mentalistic content. Quine's formulation of the mission
of naturalistic epistemology suggests a concern with what people say, with
their concrete noises, rather than with their internal ideas and beliefs.' The
main aim is to explain how it comes about that people say what they do,
that they utter the conjectures they utter. For this purpose, there need be
no contentful mental states. So the untenability of mentalistic content need
not be a doomsday scenario for all brands of epistemology.
Setting aside utterance (or speech-act) epistemologies, however, your
standard, run-of-the-mill epistemology would be in a sorry state without
mentalistic contents. Certainly my epistemology invokes such contents
throughout. So any challenge to such contents, or any ground for thinking
that empirical psychology can have no truck with such contents, needs to
be confronted.
In this chapter I shall consider three theses concerning belief content that
(if true) might imperil my epistemological project. The first thesis is that
there are no contentful mental states. If true, this would imperil my argu-
ment for obvious reasons. This thesis will be addressed in the next section.
The second thesis concedes contentful mental states but asserts that scien-
tific psychology has nothing to say about them. If this were true, there
would seem to be no way in which epistemology could profit from scientific
psychology. Hence my proposed alliance between primary epistemology
and scientific psychology would crumble. This line of argument will be ex-
amined in section 8.3. The third thesis seeks to undermine the notion that
the reliability or rationality of our belief-forming processes could be ascer-
tained by empirical means. On the contrary, it is claimed, the very ascrip-
tion of content presupposes the reliability and rationality of believers.
Reliability and rationality are guaranteed a priori, so questions of rationality .
or reliability find no place on the agenda of empirical science, as my con-
ception of epistemology proposes. This challenge will be confronted in sec-
tion 8.4.
My consideration of all these challenges must be briefer than they merit.
They raise fundamental questions in the philosophy of mind and language
that require a full length book in themselves. I can only afford a cursory
look at selected lines of argument and rebuttal. In particular, there are two
large problems in the field that I shall not try to address.
First is the question of the objects of belief. Assuming that beliefs are re-
lational, what are their relata? Are they propositions (in some guise or
other)? Are they sentences of mentalese? Are they both propositions and in-
ternal sentence^?^ This admittedly significant question will receive no sus-
164 The Problem of Content
tained attention here. It is true that I shall give sympathetic treatment to

the notion that cognition operates with some sort of 'representations', or
'data structures' (whether or not these are viewed as sentential). But even if
there is an internal code, or a multiplicity of codes, it does not follow that
belief is best analyzed, or paraphrased, as a relation between a cognizer and
a formula in the language of thought. Nor is it clear how (or whether) the
semantic properties of belief are inherited from properties of mental sen-
tences. As Putnarn suggests, even if cognitive psychology properly seeks a
functionalist account of the mind in terms of representational systems, it
does not follow that such a system of representation yields the semantic
contents that a theory of belief requirese3
This brings me to the second main problem I shall not address systemati-
cally: the topic of interpretation. If mental states do have contents, how do
they acquire these contents? What confers semantic properties on mental
states; what makes them have the truth-conditions they have? The issue of
interpretation is perhaps the central problem in the philosophy of mind.
But although I will say a few relevant things about this problem, I make no
pretense of having a positive theory.
8.2. Challenges to Content

Let me begin, then, with challenges to the very existence of mentalistic
contents. The strongest statement of such a challenge has been presented by
Paul churchlandS4He accepts the popular idea that propositional attitude
notions are embedded in folk psychology, and that folk psychology is a the-
ory. Taking the theory-theory seriously, Churchland insists that, like any
theory, folk psychology runs the risk of being false. If it is false, he adds, its
principles and ontology should be replaced. Churchland apparently holds
that folk psychology probably is false. In Lakatos's terms it is part of a stag-
nant or degenerating research program.5 While philosophers have touted
the explanatory and predictive utility of folk psychology, Churchland
stresses its explanatory failures. He argues that it cannot explain mental ill-
ness, creative imagination, intelligence differences between individuals, the
ability to catch a fly ball, or the nature and functions of sleep. He also ex-
presses skepticism about the "inner language" model of the mind that folk
psychology embraces. Finally, he maintains that folk psychology is "irre-
ducible" to neuroscience. In this fashion Churchland argues for the falsity
of folk psychology, and for the ultimate abandonment of the propositional
attitude terms it employs.
To reply to Churchland, let me start with two smaller points. First, why
does the fact that folk psychology cannot explain mental illness, creative
imagination, or intelligence differences support its falsity? Perhaps the the-
ory just has a limited domain and does not purport to explain such phenom-
ena. Narrowness of scope needn't in itself disconfirm a theory. Second, folk

psychology need not be committed to mentalistic sententialism. There are
possible construals of propositional attitudes that do not invoke a lingua
mentis. So even if mental sententialism is indeed misguided, that is not
grounds for the elimination of propositional attitudes.
Let us now turn to larger points that render Churchland's conclusions
debatable. To begin with, is it really so clear that folk psychology is a the-
ory? Philosophers have been at pains to extract a theory of mind from ordi-
nary mentalistic discourse, but it is questionable whether these attempts
have succeeded. Efforts to elaborate the functionalist program in the philos-
ophy of mind, which sees mentalistic concepts as embedded in a theory,
have run into serious problems.6 But if propositional attitude concepts are
not theoretical concepts, it is an open question whether they should be
abandoned for the kinds of reasons that scientists abandon theoretical
posits.
However, assuming that folk psychology is a theory, and that proposi-
tional attitudes are posits of that theory, the issue that then comes to the
fore is: Under what circumstances should one theory in science be over-
thrown by another (successor) theory? And under what circumstances is it
proper to abandon a theory's ontological commitments?
As already mentioned, Churchland holds that folk psychology is irreduc-
ible to neuroscience. He apparently thinks that the only alternative to re-
ducibility is replacement. He therefore expects folk psychology ultimately
to be replaced by neuroscience. But this expectation might be forestalled
for one of two reasons. First, the comparative plausibility of reduction ver-
sus replacement is partly dependent on how 'reduction' is construed. Strong
forms of reduction are indeed unlikely to be realized. But some weaker
forms of reduction might be more tenable. Second, reduction and replace-
ment may not be an exhaustive set of alternatives. Some third alternative
might be the most reasonable.
It is indeed implausible that folk psychology should be reduced to neuro- .
science in the strong sense of 'reduction', formulated by Ernest ~ a ~ e lThis
.'
kind of reduction requires 'bridge laws' that establish a one-to-one correla-
tion between the predicates of the reduced theory and the predicates of the
reducing theory; and that is unlikely to be available. In general, it is un-
likely that folk psychology is sufficiently precise for this kind of correlation.
But there are alternative conceptions of reduction in science that might
be more favorable to the present case. In Kenneth Schaffner's account of
reduction, three theories are distinguished: (1)the original theory, (2) the
corrected version of the original theory, and (3) the reducing t h e ~ r y . ~
SchafFner points out that predecessor theories can seldom be derived from
successor theories in their original form, but corrected versions of them may
be derivable. For example, classical Mendelian genetics may not be deriv-
able from modern molecular genetics; but it is conceivable that a corrected

version of Mendelian genetics-modern transmission genetics-is derivable
from modem molecular genetics? Current formulations of folk psychology
may not be derivable from any future neuroscientific theories, but some
corrected or refined version of folk psychology may be so derivable. In this
imagined version folk psychology would retain its commitment to proposi-
tional attitudes-states with contents-but might considerably refine its
laws and perhaps its terminology as well.
It is doubtful, however, that even a corrected version of folk psychology
would satisfy Nagel's requirements for reduction. The requirement of a
one-to-one correlation between folk psychology's predicates and those of
neuroscience is probably unsatisfiable. Different species of higher organisms
might achieve the same psychological states and regularities by way of a
wide variety of neurological means. Thus, it is unlikely there can be a type-
type correspondence between folk psychological predicates and neurolog-
ical predicates. This point is stressed by Jeny Fodor, inter aliu, who points
out that, in general, such type-type correspondences do not obtain between
the "special sciences" and more basic science^.'^
However, the absence of such correspondences should not preclude mul-
tiple schemes of taxonomy and theory for the same type of system. Whether
a biological system, an economic system, an information-processingsystem,
or what have you, units and states can be delineated at different levels of
analysis. Each of these levels can be scientifically respectable, even if they
are not aligned by way of type-type correspondences. Furthermore, it is
unclear that they must compete with one another for predictive power in
order to maintain respectability. Even if a higher-level set of units is inferior
in predictive potential to some new, lower-level grain of analysis, it is not
obvious that the higher-level units must be abandoned. Replacement is not
the only alternative to reduction.
In applying this theme to propositional attitudes, it is especially impor-
tant to emphasize the distinction between a theory's laws and its ontologi-
cal commitments. The laws of current folk psychology may be imprecise,
and hence relatively weak for predictive purposes. But this does not show
that the favored terms of folk psychology have no reference. There may in-
deed be belief states and desire states, even if our commonsense dealings
with such mental states have not evolved very precise or powerful raws to
describe their relationships.
This point needs elaboration. The putative linkage between a term's pre-
dictive power and its ontological respectability should not be accepted un-
critically. There are many terms used in language that seem ontologically
unproblematic, despite the fact that their predictive utility is compara-
tively weak. Consider ordinary expressions like 'tree', 'automobile', 'alive',
'wealthy', and 'holds a Ph.D.'. These terms appear in no strict laws of con-
temporary science. Any generalizations made with their help are very
rough indeed. Nonetheless, there is little cause to deny the referential status
of these predicates. There are trees, automobiles, living things, wealthy
people, and holders of Ph.D.'s. Similarly, even if words like 'belief and 'de-
sire' have modest predictive utility---compared with projected predicates
of a maturing cognitive science-this should not be taken to establish their
referential bankruptcy. The thesis that the genuinely referring terms are
only those that occur in generalizations with maximal predictive power
(within a given domain) is very dubious.
Another challenge to mentalistic content is presented by Stephen Stich,
who sometimes seems to make a more cautious claim, namely, that the be-
lief construct is unsuitable for cognitive science. At other times, however, he
seems to call the very existence of beliefs into question. Stich claims that
cognitive science must pay a very heavy price if it adheres to the program
of a "strong representational theory of mind (RTM)," that is, a program that
couches generalizations in terms of content. Predicates of the form 'believes
that p' are both vague and context sensitive. The folk language of belief
characterizes a person's cognitive state by comparing it to our own. So peo-
ple who differ fairly radically from us in doctrine, or ideology, will fall out-
side the reach of belief description altogether. But some of these subjects
may have minds that work very much as ours do. Thus, Stich argues, if there
are important generalizations that cover both us and such exotic folk, a
cognitive science embodying strong RTM is bound to miss them."
One of Stich's examples is that of aged Mrs. T, who suffers from severe
memory loss. Whenever she is asked 'What happened to McKinley?' she
says, 'McKinley was assassinated', but she is quite unable to say what an as-
sassination is, or what its consequences are. It seems inappropriate, says
Stich, to describe Mrs. T as believing that McKinley was assassinated. In-
deed, no definite content ascription seems available to us. Thus, Mrs. T's
states are beyond the reach of a cognitive theory that cleaves to descrip-
tions in terms of content.
Is Stich's conclusion warranted? It is doubtless true that we often have
inadequate resources for describing the content of a person's belief. In par-
ticular, simple sentences of English (and other natural languages) cannot be
counted on to get things right. Our language marks certain distinctions but
not others. This means that it is difficult to phrase a set of distinctions drawn
by a given cognizer but not well marked in English. This holds, for example,
of Daniel Dennett's case of a six-year-old child who says that his daddy is a
doctor.12 The child does not yet have a full understanding of 'doctor'; he
may not know that this precludes being a quack or an unlicensed practi-
tioner. So perhaps the English word 'doctor' does not exactly capture the
content of the child's belief. Still, this should not imply that the child has no
belief at all.
This point is especially clear if we individuate propositions in terms of a

possible-worlds analysis, as does Robert stalnaker.13 The content of what
the child believes can be captured in terms of the range of possibilities, or
possible worlds, that the child envisages. This set of alternative possibilities
will doubtless grow as the child matures, and then he will have another be-
lief content associated with the same English sentence. This will be cap-
tured by a different demarcation in terms of possible worlds. There are
difficulties with pinpointing belief content with English sentences, but this
does not undermine the existence of contentful states.
As previous cases illustrate, the fact that a cognizer assents to a given
sentence does not guarantee that his belief content is well captured by that
sentence. This conclusion applies to other examples in the literature. One is
drawn from Saul Kripke's paper "A Puzzle about ~elief."'~ Kripke assumes
that the language user in the following example is sincere, reflective, and
not conceptually confused. Two further assumptions he takes as noncontro-
versial principles are: (Pl) If a speaker of a language L assents to 'p', and 'p'
is a sentence of L, then he believes that p; and (P2) if a sentence of one lan-
guage expresses a truth in that language, then any translation of it into an-
other language also expresses a truth in that other language. A problem case
Kripke then describes is as follows. Pierre, a native, initially monolingual
Frenchman, has heard or read about 'Londres' being beautiful. He has per-
haps seen pictures. So he assents to the sentence 'Londres est jolie'. Later he
emigrates to England and takes up residence in London. However, he does
not learn that 'London' refers to the same city as the French name
'Londres', which he had learned many years ago. He learns English by ex-
posure, and, given his observation of his surroundings, assents to 'London is
not pretty'. Given that 'London' and 'Londres' have the same b e d semantic
value, it follows from the disquotation principle, (Pl), and the translation
principle, (PZ), that Pierre believes that London is pretty and he believes
that London is not pretty. Yet he has not made any logical blunders.
The way to deal with this puzzle is to reject the disquotation principle.
As we have seen, the relation between belief content and what a person
says (even sincerely) is not at all simple and straightforward.15A person's
thought content may be either richer or narrower than a given sentence he
utters suggests, depending on his grasp of the words used and his back-
ground beliefs or ideology. This point also applies to some cases presented
by Tyler Burge. In one of Burge's best-known examples, a man says: 'I have
arthritis in the thigh'. This man is unaware that the term 'arthritis' is prop-
erly reserved for inflammation of the joints.16 Should we say that this man
believes that he has arthritis in the thigh? This may seem innocuous enough,
and certainly would do for most practical purposes; but it does not capture
his cognitive content in an ideal fashion.
There are additional reasons for doubting that belief content can be cap-
tured with simple sentences of a natural language, even one the believer
speaks. A person's perceptual representation of an object should probably
be included in the content of a belief concerning that object. But there may
be no satisfactory way of identifying this perceptual content with any sim-
ple sentence of a natural language. Of course, since we often want to com-
municate about the content of perceptual beliefs, we need some vehicle of
communication, especially a handy, efficient one. So we employ simple
sentences of a natural language. It is a mistake to suppose, however, that
such sentences can provide anything but a very rough approximution of be-
lief contents. Even in expressing one's own beliefs, it is doubtful that one's
utterances convey them very fully or accurately. None of this, however,
constitutes negative evidence against the existence of contentful beliefs.
Let us return now to Stich. Stich's worries are not exhausted by the diffi-
culty of pinpointing belief contents in puzzle cases. He presents two other
problems for the viability of the belief construct in cognitive science. First,
he claims that folk psychology's belief construct requires that beliefs be the
cognitive causes of both verbal and nonverbal behavioral effects. But cer-
tain psychological experiments show that verbal and nonverbal behavior
have distinct cognitive causes. Hence, he claims, the right thing to say is
that there are no such things as beliefs.
Second, Stich contends that beliefs should be identical with "naturally
isolable" parts of the cognitive system, but there are possible network
models of the mind within cognitive science that would violate this isolabi-
lity (or modularity) requirement.
These arguments are effectively countered by Terence Horgan and James
woodward.'' I briefly summarize their principal responses. First, even if
the dual-control thesis is partly right, it is still plausible to posit beliefs as
the two sets of causes. (Notice that Stich himself concedes that there would
be "belief-like" states.)18 Second, in m n y cases people display integrated
verbal and nonverbal behavior: for example, lecturing while doing a logic
problem on the blackboard. Here it seems inescapable that the same states
must be causing both verbal and nonverbal behavior. The control systems
must not be wholly independent. So belief states causing both verbal and
nonverbal behavior must be posited at least for this class of cases.
In reply to Stich's second argument, Horgan and Woodward reveal the
implausibility of the natural isolability, or modularity, requirement, point-
ing out that it is an enormously unattractive intertheoretic compatibility
requirement. There is no reason why beliefs could not be identical with
highly complex, even 'gerrymandered', conglomerations of lower-level
events. Such complex events could still play the causal role that folk psy-
chology imputes to them.
I have surveyed a variety of recent arguments against the existence of

beliefs and found none of them conclusive.19 So let us proceed to the two
other challenges to my epistemological undertaking.
8.3. Can Psychology Address Content?

I seek to enlist psychology's aid in determining the truth-conducive proper-
ties of mental processes. But can psychology do this? Only, it seems, if it can
address cognitive states in terms of their content. But there are arguments
in the literature that psychology does not, or should not, do this. While
mental states may h o e content, psychology is allegedly silent about it. If
that is right, how can psychology shed light on truth-conducive properties
of mental mechanisms? Truth-conduciveness is obviously fixed by proper-
ties related to content.
Let me introduce this argument with some familiar background. Accord-
ing to most recent theories, the truth-conditional content of a belief is a
function not just of what is in the cognizer's head, but of external things as
well. This sort of point has been made most influentially by Putnam with his
Twin-Earth examples.20 But it is equally a consequence of many kinds of
functionalism, which would link the identity conditions of propositional at-
titudes to their causal and counterfactual relations with external stimuli and
bodily movements.
Now consider the additional claim that psychology deals exclusively with
what's in the head. In particular, a proper cognitive science is only con-
cerned with the 'formal', 'syntactic', or computational properties of mental
processes. This thesis has been advanced in several quarters, chiefly by Jerry
Fodor. Fodor calls it the "formality ~ o n d i t i o n . "There
~ ~ is some question as
to whether Fodor really intends the thesis in question; I will return to this in
a moment. But let us examine the passages that suggest it.
Fodor distinguishes the formal properties of a mental representation
from its semantic properties. "Formal operations are the ones that are spe-
cified without reference to such semantic properties of representations as,
for example, truth, reference, and meaning".22 The former are purely inter-
nal characteristics of a representation, while the latter relate the represen-
tation to the external world. Fodor goes on to assert that cognitive
psychology proceeds (or should proceed) purely in terms of formal proper-
ties, that is, computationally. This apparently means that truth-conditional
content is to be ignored.
The formality condition, viewed in this context, is tantamount to a sort of
methodological solipsism. If mental processes are formal, then they have ac-
cess only to the formal properties of such representations of the environment
as the senses provide. Hence, they have no access to the semantic properties
The Problem of Content 1 71
of such representations, including the property of being true, of having refer-

ents, or, indeed, the property of being representations of the en~ironrnent.~~
However, there are other passages in Fodoryswriting that appear incon-
sistent with the foregoing. In another article he writes: "Propositional atti-
tudes interact causally and do so in virtue of their content."24 And in the
introduction to the volume collecting the papers from which I have quoted,
he writes: "In order to specify the generalizations that mentalistic etiologies
instantiate, we need to advert to the contents of mental states."= It appears
from these passages that psychology does need to deal with contents, a con-
clusion that apparently conflicts with the other cited passages. But Fodor
may here intend 'content' to refer to narrow content, which is not a seman-
tic notion.26 Since we are restricting attention to wide, truth-conditional
content, Fodor can indeed be interpreted as holding that cognitive psychol-
ogy ignores (or should ignore) that kind of content (wide content). In any
case, there are other exponents of this view. Stich, for example, forthrightly
endorses the view that cognitive science should proceed purely
"syntactically."27
How, then, would the presently envisaged challenge to my enterprise go?
It might be formulated in the following argument:
(1) Cognitive psychology proceeds purely syntactically, formally, or
computationally.
(2) Truth-conditional content is not (just) a function of syntactic processes
or relations. It is partly a function of external objects, of mind-world
relations.
Therefore,
(3) Cognitive psychology does not deal with truth-conditional content.
Hence,
(4) Cognitive psychology cannot shed any light on truth-conditional prop-
erties of mental processes, including truth-conducive properties of be-
lief-forming processes.
Once laid out in this fashion, though, the argument is clearly unsound. Most '
importantly, (3)does not folIow from (1)and (2). Any appearance of validity
trades on an ambiguity between opaque and transparent readings of 'deals
with', in (3).If psychology were confined to purely formal descriptions, it
would not deal with contentful states on an opaque reading of 'deal with'; it
would not talk about contentful states in terms of their content. But assum-
ing that mental states are contentful, psychology would still deal with con-
tentful states on a transparent reading. The states it addresses in contentless
terminology would in fact be contentful states.
So even if cognitive psychology proceeds purely nonsemantically, it can
still tell us things about beliefs and about various causal processes that gen-
erate beliefs. Even if it describes beliefs and belief-causing processes purely
formally, once the contents of these states are assigned, one would be in a
position to assess the truth-conduciveness of the processes, In other words,
even if cognitive psychology does not participate in the assignment of
truth-conditional contents, it can still play the role in primary epistemology
that I have assigned it, that of identifying-in some descriptive terms or
other!-the various belief-forming processes available to the organism. This
completely undercuts (4), which comprises the challenge to my project
being considered in this section.
Although this completes my main response to this challenge, a few other
comments are in order, apropos of premise (1). First, one might reject prem-
ise (1)as descriptively false. Many parts of psychology in fact deal with the
world-mind interface. Perceptual psychologists study the relationship be-
tween physical inputs and perceptual representations, in order to deter-
mine when and why perceptual illusions occur. And even when cognitive
psychologists do not study environment-organism interactions, they com-
monly do employ content descriptions to characterize internal events.28
But suppose (1)is construed as a prescription, rather than a description of
current practice. Perhaps it is a prediction of psychology's practice in some
ideal, future incarnation. On what would such a prescription be based?
One consideration, of course, is the computational model, which has in-
spired many of the themes of cognitive psychology. But notice that even the
computational model makes use of extrinsic descriptions. When we de-
scribe states of a corpputer in terms of numbers and the arithmetic function
addition, we are not giving intrinsic descriptions of those states; we are in-
terpreting the states in terms of some external, abstract objects. So the
computational model itself makes use of semantic interpretation.
A second possible reason for a purely formal psychology is that the true
laws of the system can be formulated only in syntactic, intrinsic terms. But
is this correct? It is admittedly plausible to suppose that if there are any
laws of the system statable in extrinsic terms, there must also be laws stat-
able in intrinsic terms.29But this would not show that no laws are statable in
extrinsic terms. Furthermore, some writers argue that one cannot state all
laws of the system without including laws formulatable in semantic termi-
n o ~ o ~Laws
~ . ~stated
' in intrinsic terminology would be different laws. So if
psychology wants to include all laws of the system, it must work at the se-
mantic level as well as the intrinsic.
A different kind of defense of psychology's use of content descriptions
would reject the heavy emphasis on laws. According to a recent model of
psychological explanation offered by Robert Cummins, the main style of
psychological explanation is not subsumption under laws.31 Rather, it is
"functional analysis": explaining a system's possession of one property by
appeal to properties of its components. If this model of psychological expla-
nation is correct, psychology would not be forced to abandon content de-
The Problem of Content L 73
scriptions even if it turned out that the best (or only) laws of the system
were statable syntactically. Psychology might still want semantic descrip-
tions of the organism at certain levels of description, which would ulti-
mately, of course, be explainable in terms of lower-level organization.
8.4. Content, Rationality, and Reliability

I turn now to the final challenge concerning content attribution. This chal-
lenge is based on the claim that certain questions I regard as questions for
psychology are already tacitly answered by the very ascription of content.
These answers are given, as it were, a priori, or transcendentally; they are
not questions for empirical psychology. Hence, psychology need not, and
cannot, play the role I ascribe to it.
This line of argument comes in two forms. First, it is claimed that the
ascription of content presupposes rationality. We automatically attribute
rationality to a being just by imputing propositional attitudes to him. But
rationality involves certain kinds of inferential processes. Hence, what in-
ferences people make is (largely) an a priotj matter, not one to be settled by
empirical psychology. Second, it is claimed that the ascription of content
presupposes reliability. In imputing propositional attitudes, we must so in-
terpret a creature's beliefs that most of them come out true. So questions of
reliability cannot be empirical questions. Since we are already committed
to (a large amount of) reliability by the very ascription of beliefs, the matter
of reliability is (largely) settled before psychology even enters the picture.
Strands of these theses have been endorsed by a number of authors. But they
are emphasized most heavily by Donald Davidson; so I shall pay closest at-
tention to his formulations.
The principal source of these themes, though, is Quine. In discussing in-
terpretation of a speaker's utterances, Quine advocates a "charity princi-
ple'': "Fair translation preserves logical laws . . . Assertions startlingly false
on the face of them are likely to turn on hidden differences of language . . .
One's interlocutor's silliness, beyond a certain point, is less likely than bad
tran~lation."~~
Davidson has enunciated similar themes in several papers. Regarding
rationality, he writes: "If we are intelligently to attribute attitudes and be-
liefs, or usefully to describe motions as behavior, then we are committed to
finding, in the pattern of behavior, belief, and desire, a large degree of
rationality and consistency."33 In the same vein (but apropos of preferences
rather than beliefs), he writes: "I do not think we can clearly say what
should convince us that a man at a given time (without change of mind) pre-
ferred a to b, b to c, and c to a. The reason for our difficulty is that we can-
not make good sense of an attribution of preference except against a
background of coherent a t t i t ~ d e s . " ~ ~
Concerning the imputation of reliability (to use my term), Davidson

writes: "Charity is forced on us;-whether we like it or not, if we want to
understand others, we must count them right in most matter^."^' "Since
knowledge of beliefs comes only with the ability to interpret words, the
only possibility at the start is to assume general agreement on beliefs. We
get a first approximation to a finished theory by assigning to sentences of a
speaker conditions of truth that actually obtain (in our opinion) just when
the speaker holds those sentences true."3B"What makes interpretation pos-
sible, then, is the fact that we can dismiss a priori the chance of massive
error. A theory of interpretation cannot be correct that makes a man assent
to very many false sentences: it must generally be the case that a sentence is
true when a speaker holds it to be."37
A similar stance on rationality is taken by Dennett. In a recent paper he
says that propositional attitudes are premised on the following principles:
(1)a system's beliefs are those it ought to haue, given its perceptual capaci-
ties, its epistemic needs, and its biography; (2) a system's desires are those it
ought to haue, given its biological needs and the most practicable means of
satisfying them; and (3)a system's behavior will consist of those acts that it
would be rational for an agent with those beliefs and desires to perform.38
Returning to the reliability theme, several others have sketched ap-
proaches to interpretation guided by the idea that representation derives
from the capacity of a system to be a reliable indicator of its context. These
writers include Stephen Schiffer, Hartry Field, Brian Loar, Fred Dretske,
and Jerry odor?' However, I will not undertake to detail their proposals.
To turn to critical discussion, let us look first at the thesis about rational-
ity, restricting attention to that segment of rationality concerned with be-
liefs and inferences (rather than, say, preferences). So restricted, rationality
is often equated with 'logicality'. An initial problem with this equation is
that logic does not entail any inference principles, an observation I stressed
in Chapter 5. So it just is not clear what it means for a belief system to be
logical.
At least two interpretations, though, readily come to mind. First, a belief
system may be regarded as logical (and therefore rational) only if it is con-
sistent. Second, a belief system may be regarded as logical only if it is char-
acterized by deductive closure: the agent believes all logical consequences
of his beliefs. Neither of these principles is very plausible as standards for
belief attribution. Deductive closure is especially unattractive. Surely no
human being infers all the logical consequences of his beliefs, yet we do not
hesitate to ascribe beliefs to people.
On r.eflection, the consistency principle is not plallsible either. Accep-
tance of this principle would imply that nobody could possibly have an in-
consistent set of beliefs, since any apparent inconsistency would disqualify
the person's mental states from being beliefs at all. But surely we do not
withhold belief ascriptions for this reason. Russell showed that the axiom of
comprehension leads to inconsistency; but this does not mean that people
who previously (ostensibly) accepted the axiom of comprehension did not
really have any beliefs.
Of course, Davidson does not claim that believers must always be rational
or consistent. He only says that propositional attitude attributions commit
us to finding a "large degree" of rationality and consistency in the subjects
of our attributions. But then it emerges that there is no a priori determina-
tion of the specific logical operations that believers must always, or even
regularly, employ. So, as I wish to contend, it presumably rests with em-
pirical science to determine just what these deductive operations are.
I may put the point this way. Since perfect logicality is not a presupposi-
tion of belief attribution, belief attribution at best presupposes minimal log-
icality. This point has been emphasized by Christopher cherniake4' But
what, exactly, does 'minimal logicality' imply? Precisely which deductive
inferences must a cognizer make, as a prerequisite of having beliefs at all? I
do not see how any plausible conditions can be laid down a priori. Formal
logic catalogs all sorts of argument patterns. How should we pick from
these the ones that a believer, qua believer, must (always?) execute. Even if
we laid down a disjunction of such patterns as requisite for believing, no
conclusion could be drawn about which of these disjuncts human beings
regularly, or typically, or easily, execute. As far as human inference capaci-
ties are concerned, the matter is surely an empirical one.
The point can be brought home another way. Before teaching a course in
logic, it is not obvious to the trained Ph.D. exactly which inferences fresh-
men will find easy or hard, natural or unnatural. This is manifestly an em-
pirical matter, and quite appropriately studied by the psychology of
reasoning. (See Chapter 13 for further discussion.)
I have failed to identify any plausible a priori logicality constraints, even
in the domain of deductive logic. But a priori constraints of rationality, if
there are such, should also include inductive logic. Here one hardly knows
where to begin. Principles of inductive logic are very controversial, in the
first place. Assuming we could settle on some such principles, are they
really to be required as conditions for belief ascription? What about princi-
ples of probability and statistics? Are creatures who violate such principles
necessarily devoid of propositional attitudes? That is extremely dubious.41
Empirical studies strongly indicate that people have little native grasp of
probability (as I will discuss in Chapter 14). But this hardly keeps them
from qualifying as believers.
It is time to turn to the second form of the present challenge, the one
concerning reliability. The first comment I need to make is that, contrary to
what Davidson avers, our procedures of content attribution do allow the
ascription of belief systems that are 'massively' false (in the ascriber's opin-
ion). Cartesian demon scenarios are apt for illustrating this suggestion. But
let me give a slightly more realistic variant of Descartes's demon.
Imagine a child who grows up normally and develops a mastery of
English. Several cognitive catastrophes then befall him. He is stricken by
diseases that distort his visual powers and his memory-except his memory
for language. As a result, he now describes his visual environment in ways
we regard as massively false. And when he recalls things from his personal
past, his accounts are systematically erroneous. To ensure his health and
well-being, he is placed in the hands of a guardian, who feeds and treats him
well except for instruction in cognitive matters. The diabolical guardian
reads him specially concocted newspapers and encyclopedias full of false-
hoods. So the child gradually acquires a radically false picture of the world.
The guardian even teaches him fallacious inference patterns, which take
him even further from the truth. His misconceptions are expressed in avow-
als of belief, and since there is nothing amiss with his linguistic memory, we
have every reason to interpret his utterances homophonically. We would
then be strongly disposed to ascribe to him a system of belief that is mas-
sively false.42
Suppose, nonetheless, that we concede a kernel of truth in the notion that
some sort of truth constraint is presupposed by the ascription of content,
though of a more modest nature than those Davidson endorses. What would
be the repercussions of this concession? A theory of interpretation still
would not fix a belief set's reliability, so the determination of just what that
global reliability is would still be an empirical matter.
Furthermore, even if Davidson were right in claiming that massive error
is excluded a priori, this would not damage my theory. Suppose belief
ascription commits us to an accuracy rate of 50 percent. The criterion of
justifiedness presumably requires a higher truth ratio, say 80,90, or 95 per-
cent. Can people's belief-forming processes attain such higher levels? This
would be an empirical question, not foreclosed by the theory of
interpretation.
Notice next that even if the theory of interpretation requires a high truth
ratio across the corpus taken as a whole, there might be small slums of de-
fective belief-forming processes, perhaps infrequently employed. The task
of identifying a right system of J-rules would include the identification and
rejection of such defective processes. More generally, for purposes of the
theory of justifiedness, it is not enough to know the global truth ratio of a
cognizer's actual beliefs. One needs to know the specific cognitive processes
that are employed and available for employment, since it is only such pro-
cesses that a right J-rule system will specify. Identification of the available
processes is an empirical task.
Finally, all of the foregoing questions concern reliability. But epistemol-
ogy is also interested in power and speed. Evaluation of basic cognitive pro-
.
cesses along these dimensions is not at all affected by the theory of
interpretation. (At least, no argument has been given for any such thesis.)
Thus, many questions of importance to primary epistemology would re-
quire contributions from cognitive science-ven if some of Davidson's
more radical suggestions were correct.
I conclude that any plausible interpretation theory is compatible with
the role assigned by my framework to empirical psychology. Let us pro-
ceed, then, to concrete illustrations of the contributions that psychology
might make.
part II
ASSESSING O U R C O G N I T I V E
RESOURCES
Perception
9.1. An Agenda for Primary Epistemics

The principal way that cognitive science can contribute to epistemology, I
claim, is to identify basic belief-forming, or problem-solving, processes.
Once identified, these processes would be examined by primary epistemol-
ogy according to the evaluative dimensions and standards adduced in Part I.
How do the processes fare in terms of justification, rationality, and intelli-
gence? How well do they stand up, as judged by the three standards of relia-
bility, power, and speed?
My extended treatment of justification, in Chapters 4 and 5, may suggest
to the reader that I regard justification as the central term of epistemic ap-
praisal. Indeed, the reader may be expecting a full-scale system of J-rules,
with permissions granted to a specified coterie of basic cognitive processes.
Any such impression, however, needs correction.
My extended attention to justification is somewhat disproportionate to
the importance I attach to it. I have devoted so much attention to it partly
because of its pivotal role in traditional epistemology, and partly because I
have a positive theory to advance. This does not mean that I regard evalua-
tive terms like 'rational' and 'intelligent' as less important. I simply have no
fully developed positive accounts of those notions to present. (I also find
them intrinsically vaguer notions.) Nor do I mean to assign the standard of
reliability a higher intellectual status than power or speed. I decline to offer
any such ranking.
As far as the construction of a right J-rule system is concerned, it would
be greatly premature to essay such a construction at this juncture. Cogni-
tive science is still groping its way toward the identification of basic pro-
cesses. Many different theories are afloat. This is understandable, since the
mind-brain is an enormously complex system, with a wide range of subsys-
tems, all of which are extraordinarily difficult to study experimentally. Fur-
thermore, it is no trivial task to relate the sorts of processes tentatively
delineated by cognitive science to questions on the epistemological agenda.
182 Perception
Even an initial survey of the field and sampling of the possibilities will oc-
cupy us at length. Finally, a truly complete set of J-rules would have to reg-
ister an interplay between different components, or functions, of the
mind-brain. I shall have enough on my hands in identifying relevant pro-
cesses individually, and exploring their properties.
For all of these reasons, Part I1 presents only some first steps toward a
full-fledgedprimary epistemics: a prolegomenon, if you will, to the subject.
Clearly, a full development of primary epistemics will require continued
research and collaboration by many people from several disciplines.
I have said that in this part of the book I will illustrate prospective con-
tributions of cognitive science to epistemology. The term 'illustrate' needs
emphasis. It is not crucial to my view that any particular item in Part I1
should prove to be correct. I am not wedded to any particular candidate as
a basic cognitive process; nor do I have a significant stake in the normative
assessment of any such candidate. My aim is to show what might be among
the basic cognitive processes and what the epistemological implications of
this may be. Relative to this aim, no specimen hypothesis I adduce about
psychological processes is critical. The structure of the inquiry concerns
me, not the correctness of specific details. However, I naturally select
themes taken seriously by at least part of the cognitive science community.
And I pay some attention to the evidence supporting various hypothesized
processes, although I do not do this systematically or rigorously. Weighing
evidence is the task of cognitive science itself (using currently accepted sci-
entific methods). The strict role of primary epistemology is to borrow the
results of cognitive science and assess the epistemic repercussions of these
results.
While the principal contribution of cognitive science is to identify basic
belief-forming processes, there is another contribution as well: refinement
of the descriptive categories with which epistemology operates. For exam-
ple, epistemology commonly deploys the category of belief. But belief is a
very coarse-grained category from the standpoint of mental science. Pri-
mary epistemology needs a more subtle and flexible set of doxastic state cat-
egories. Drawing on psychological materials, I shall introduce a few such
categorial refinements among doxastic states, distinguishing 'perceptual'
from 'central' beliefs (this chapter), and 'activated' from 'unactivated' be-
liefs (Chapter 10).The whole topic of uncertainty or degrees of confidence,
standardly handled with the category of subjective or 'judgmental' probabi-
lities, also needs to be addressed with increased psychological sophistica-
tion. I shall draw on psychological materials to tender a new approach to
accept-anceand uncertainty in Chapter 15.
Of course, any complication in the array of doxastic categories threatens
to complicate the picture of normative assessment. When activated and
unactivated beliefs are distinguished, questions arise about the meaning of
Perception 183
'reliability' and 'power'. Are these to be defined in relation to activated be-

liefs or unactivated beliefs? Such complications cannot be avoided in a seri-
ous and realistic theory. (I shall have other reasons also, in Part 11, to tinker
with some of our normative standards. But no major thematic changes will
be introduced.)
What are the implications of Part I1 for the global assessment of native
cognitive processes? Within the cognitive science community, perspectives
on human endowments span a wide spectrum, ranging from rosy to gloomy,
from admiring to depreciating. The admirers are struck with the remark-
able capacities of the system's endowments. Indeed, it is hard not to be im-
pressed when one reflects on the ease with which the mind learns
enormously complex language systems, or the mind's plasticity in problem-
solving activity (neither capacity being adequately understood). Human
mechanisms must be doing something-many things-right! But depreca-
tors within the cognitive science community point an accusing finger at the
'dirty' heuristics the mind seems to employ: procedures which do not mea-
sure up to statistical or probabilistic standards that are widely accepted as
normatively correct.
I shall present examples of both outlooks in this part of the book. But my
own tentative assessment will be neither warmly glowing nor severely dis-
paraging. It will be more qualified and moderate in tone. The mind uses all
sorts of cognitive processes and process variants. In terms of, say, reliability,
they are a mixed bag: some good, some not so good. Moreover, most of them
have significant trade-offs. The less reliable processes are often quicker.
They may be good from a pragmatic standpoint, even if they do not confer
maximal power. So I shall not be arguing for any sweeping conclusions here.
I adduce no reasons to think that the human mind, on the whole, is either
strikingly better or strikingly worse, epistemically speaking, than others
have supposed. The main conclusion, perhaps, is that it takes much careful
sifting and analysis both to identify the basic processes and to weigh their
strengths and weaknesses.
My discussion will sometimes suggest conclusions, however, that are at
variance with doctrines many epistemologists and methodologists have
held. For example, the appropriateness of a strong 'total evidence' require-
ment will be called into question (Chapter 10).The softening of principles
for belief revision will be proposed (also Chapter 10).Greater approval will
be tentatively conferred on 'imagistic' thinking (Chapter 12). 'Acceptance'
of hypotheses, rather than continued uncertainty, will be viewed more fa-
vorably than many theorists would condone (Chapter 15). And processes
that breed 'order effects' will be seen as nonetheless sanctionable (Chapter
16).
The importance of my analysis in Part 11, however, is not the depth of
disagreement with other epistemologists that it exemplifies. Its importance
184 Perception
lies in the exploration of largely uncharted problems. Epistemology simply

has not had the opportunity to inspect and assess fine-grained cognitive
processes, because the kinds of processes at work in the mind-brain have not
been dissected (in a scientifically adequate manner) until recent years. The
importance of Part 11, then, consists more in the formulation of new ques-
tions than in the answers I propose here.
A final comment is in order on what this part of the book does not seek to
accomplish. I have already stressed (in Chapter 8) that this is not an essay on
the philosophy of mind. Thus, I am not seeking to resolve such issues as the
nature and source of semantic properties for mental states. I now wish to
indicate with equal emphasis that this is not an essay on the methodology of
psychology. This means two things. First, I shall not spend time defendin
\\
'
the empirical cognitivist approach to the study of the mind, as opposed t

generic rivals, such as phenomenology, introspectionism, or behaviorism.
Second, I shall not discuss systematically the divergent research methodolo-
gies within cognitive science, nor seek to choose among them in a princi-
pled fashion. I assume that some variant or variants of cognitive science are
good ways to study the mind. Working on this assumption, I draw on any
materials that strike me as plausible and interesting for epistemological
purposes.
It may be observed that the methodological details of the study of mind
surely belong to epistemology, at least to that wing of epistemology called
'philosophy of science'. So why don't these issues fit squarely into Part II?
The answer follows from my distinction between processes and methods.
Insofar as epistemology deals with proper methods, especially methods of
the several sciences, the methodology of psychology does belong to episte-
mology. But this volume deals principally with primary epistemology,
which concerns itself with acceptable processes, not methods. That is why
the methodology of psychology falls outside my purview.
9.2. Bottom- Up and Top-Down Processing

Perception is a central topic in both psychology and epistemology, and
therefore a suitable starting point in exploring the interface between psy-
chology and epistemology. But there is a drawback to this strategy. The
psychology of perception is such a huge and complex topic that my treat-
ment of it will, of necessity, be extremely selective and shallow. Nonethe-
less, since it bears on such central issues in epistemology, it cannot be
neglected. I will start, then, with a brief foray into this domain.
A standard approach to perception is to view it as pattern recognition.
People regularly recognize presentations of a certain pattern-such as the
letter 'A '-as instances of that pattern. How is this recognition accomplished?
Before turning to this question, note that the language of recognition im-
184 Perception
lies in the exploration of largely uncharted problems. Epistemology simply

has not had the opportunity to inspect and assess fine-grained cognitive
processes, because the kinds of processes at work in the mind-brain have not
been dissected (in a scientifically adequate manner) until recent years. The
importance of Part 11, then, consists more in the formulation of new ques-
tions than in the answers I propose here.
A final comment is in order on what this part of the book does not seek to
accomplish. I have already stressed (in Chapter 8) that this is not an essay on
the philosophy of mind. Thus, I am not seeking to resolve such issues as the
nature and source of semantic properties for mental states. I now wish to
indicate with equal emphasis that this is not an essay on the methodology of
psychology. This means two things. First, I shall not spend time defendin
\\
'
the empirical cognitivist approach to the study of the mind, as opposed t

generic rivals, such as phenomenology, introspectionism, or behaviorism.
Second, I shall not discuss systematically the divergent research methodolo-
gies within cognitive science, nor seek to choose among them in a princi-
pled fashion. I assume that some variant or variants of cognitive science are
good ways to study the mind. Working on this assumption, I draw on any
materials that strike me as plausible and interesting for epistemological
purposes.
It may be observed that the methodologica1 details of the study of mind
surely belong to epistemology, at least to that wing of epistemology called
'philosophy of science'. So why don't these issues fit squarely into Part II?
The answer follows from my distinction between processes and methods.
Insofar as epistemology deals with proper methods, especially methods of
the several sciences, the methodology of psychology does belong to episte-
mology. But this volume deals principally with primary epistemology,
which concerns itself with acceptable processes, not methods. That is why
the methodology of psychology falls outside my purview.
9.2. Bottom- Up and Top-Down Processing

Perception is a central topic in both psychology and epistemology, and
therefore a suitable starting point in exploring the interface between psy-
chology and epistemology. But there is a drawback to this strategy. The
psychology of perception is such a huge and complex topic that my treat-
ment of it will, of necessity, be extremely selective and shallow. Nonethe-
less, since it bears on such central issues in epistemology, it cannot be
neglected. I will start, then, with a brief foray into this domain.
A standard approach to perception is to view it as pattern recognition.
People regularly recognize presentations of a certain pattern-such as the
letter 'A '-as instances of that pattern. How is this recognition accomplished?
Before turning to this question, note that the language of recognition im-
186 Perception
letters have many features in common-as with C and G-evidence sug-

gests that subjects are particularly prone to confuse them.' However, even
though our perceptual system may extract features, what we perceive are
patterns composed from these features. The feature-extraction and feature-
combination processes underlying pattern recognition are not available to
conscious awareness. What we are aware of are the patterns.5
The processing described above is referred to as bottom-up, because in-
formation flows from little perceptual pieces to larger units built from
them. Bottom-up processing is assumed to be operative in recognizing
words out of letters, sentences out of words, and so on. In addition to bot-
tom-up processing, there is considerable evidence for top-down processing,
in which higher-level beliefs, or background beliefs, influence the interpre-
tation of low-level perceptual units. Psychologists have particularly shown
that (knowledge of) a pattern's context influences how one perceives that
pattern.
Consider the example in Figure 9.1. We perceive this stimulus as THE
CAT, interpreting the middle letters of the words as H and A respectively.
But in fact the H and the A are identical stimuli. Presumably, these identi-
cal stimuli are interpreted differently because of our knowledge (or beliefs)
about adjacent letters, and our knowledge (or beliefs) about which letter se-
quences are most likely to be presented. These are pieces of 'higher-level'
knowledge.
Dramatic evidence for the role of context involves speech perception,
specifically, the phoneme-restoration effect. R. M . Warren had subjects lis-
ten to the sentence, The state governors met with their respective legislatures
convening in the capital city, with 120 milliseconds of pure tone replacing
the middle s in legislatures. Only one in twenty subjects reported hearing
the pure tone, and that subject was not able to locate it correctly.'
In this example context causes subjects to hear what is not there. But
there are plenty of cases in which context facilitates correct identification.
G. A. Miller and S. Isard presented sentences to subjects in noise; subjects
were to listen to the sentences and report them back.' Some sentences were
normal; others were semantically anomalous, though grammatical;. and still
others were ungrammatical. Successful identification was clearly best for
the normal sentences, second-best for grammatical but anomalous sen-
tences, and worst for ungrammatical sentences. Subjects obviously im-
proved their speech perception both through the use of meaning constraints
Figure 9.1
Perception 187
and grammatical constraints. That is, they were able to use their back-
ground knowledge about meaning and grammar to facilitate correct
identification.
Cognitive psychologists generally believe that perception uses a mixture
of bottom-up and top-down processing. The occurrence of top-down pro-
cessing is the basis for the widely made claim that perception is 'intelligent'.
All one's knowledge about the world--or a lot of it, at any rate--can aid the
construction of a percept.
One dissenting voice is that of Jerry odor.^ Fodor defends the thesis that
so-called input systems-including perceptual systems-are modular, and
one significant feature of these modules is that they are "informationally
encapsulated." This means that pieces of information represented else-
where in the organism are not available for use by such systems. Obviously,
the thesis of informational encapsulation conflicts with the claims of many
proponents of top-down processing.
However, Fodor's statements of his position make it less radical than it
initially appears, as is clear from this passage:
The claim that input systems are informationally encapsulated is equivalent
to the claim that the data that can bear on the confirmation of perceptual
hypotheses includes, in the general case, considerably less than the organism
may know. That is, the confirmation function for input systems does not have
access to all of the information that the organism internally represents.'
Fodor does not deny, then, that some top-down processing takes place.
Rather, he wishes to register three points. First,, he argues that some infor-
mation which the organism represents-in the 'central' systems-cannot
feed back on perception. That is, this information cannot influence the con-
struction of a percept. In a Miiller-Lyer illusion, one line looks longer even
when we know that the two lines are equal in length. What we know ap-
parently does not (in Zenon Pylyshyn's terminology)'0 cognitively "pene-
trate" the perceptual process. Second, Fodor disputes the interpretations of
some of the experiments standardly cited in support of top-down pro-
cessing. There is a question, for example, of whether subjects are really re-
porting how things look or sound, rather than, say, guessing. Third, Fodor
claims that where there is top-down information flow, it occurs within the
perceptual module. The input system cannot access outside information,
but it can deploy internal high-level information to help construct a
percept.
It seems hard to deny that memory information influences perception.
The most striking cases are with fragmented or incoherent figures, for ex-
ample, pictures that initially appear to be an assortment of dots, globs, and
unrelated lines. Once we succeed (perhaps with the help of a hint) in iden-
tifying such a figure as a particular object or group of objects, such as a dal-
188 Perception
matian or a rider on horseback, we have no difficulty seeing the figure in the

same way on subsequent occasions. A memory-schema obviously shapes
later perceptual acts." Of course, it needn't be that all knowledge lodged in
memory affects perception.12 But at least memory traces of previously per-
ceived figures seem to do so.
This is not the place to try to settle this delicate issue. It suffices for my
purposes to register the widespread acceptance of at least some top-down
processing in perception, which even Fodor accepts. Let me now begin to
explore some possible epistemological morals of this kind of psychological
material.
9.3. Reliability, Speed, and Power in Perception

The main question I want to raise in this section is whether top-down pro-
cessing heightens or lowers the reliability of perception. First, however, I
pause to consider a different question about top-down processing, which
has received lively discussion in the epistemological arena. This concerns
the possibility of 'objective', or 'neutral', theory choice in science.
Epistemologists have not been slow to draw conclusions from the ap-
parent influence on perception of background knowledge, theories, and ex-
pectations. The most prominent and extreme example, perhaps, is Thomas
Kuhn. Appealing to Gestalt and New Look psychology, Kuhn has con-
tended that scientists with different theoretical "paradigms" see things dif-
ferently. Hence there is no independent way to settle disputes among
holders of competing paradigms. l3
The tenability of this significant epistemological thesis obviously hinges
on facts of perception. It is therefore important to stress that several consid-
erations render the thesis dubious. First, as indicated in section 9.2, not all
theories and background beliefs are capable of shaping perception. In addi-
tion to the Miiller-Lyer illusion, there is the familiar stick-in-the-water case.
Even when a person knows that a particular stick is straight, it still looks
bent when immersed in water. His background theoretical beliefs do not
penetrate his perceptual processing so as to make the stick look straight.14
Second, even if there is a tendency to interpret what one sees in terms of
background theories, this interpretation is arguably a post-perceptual phe-
nomenon, a nonperceptual belief. If so, the percept itself may be a non-the-
ory-laden component of cognition. If that is right, there is a theory-free
element in the cognitive system to which appeal can be made in trying to
settle theoretical disputes.
Third, even if some percepts are theory-laden, it may be possible to re-
duce and even eliminate theory-ladenness. Many psychological experiments
involving top-down processing feature very brief exposure times, or large
quantities of 'noise'. Perhaps context and prior expectations predominate
Perception 189
over low-level determinants only when exposure times are brief, or when
there is a lot of noise, so that no perceptual conclusion can be reached by
bottom-up processing alone. Given more time and/or less noise, bottom-up
processing may predominate. For example, when one looks leisurely and
carefully at the stimulus in Figure 9.1, one sees clearly that the second let-
ters of each word are identical.
It is possible, then, that even if theory-ladenness sometimes occurs, it can
be excluded in scientific contexts by appropriate experimental designs. In
the case of scientific observation, there is (in general) no bar to ample ob-
servation time and noise reduction. Furthermore, the context of observation
can often be manipulated, so the impact of context can be minimized or
eliminated.
Finally, even when a percept is affected by categories stored in memory,
it is normally possible to switch back to a prerecognition stance: to see the
form independently of the familiar category. For example, even if one nor-
mally sees a drawing of a cube as a cube, with appropriate perspective, it is
also possible to see it in a prerecognition, two-dimensional mode, namely, as
a square and two parallelograms joined together at their edges, as suggested
in Figure 9.2. This again illustrates how perception is capable of releasing
itself from the influence of theory.
I now turn to the more central issue of this section: the impact of top-
down processing on the reliability of perception. Primary epistemology is
concerned with this point for the following reason. Let us assume that per-
ceptual processes can occur with variable amounts of top-down processing,
ranging from none at all to substantial amounts. If there is a substantial dif-
ference in reliability among these different processes, this could well have
an impact on approved systems of J-rules. Suppose, for example, that heav-
ily top-down perceptual processes-those using large amounts of contextual
information-are much less reliable than purely bottom-up processes. Then
perhaps no rules permitting these top-down processes would be included in
any right system of J-rules.
We have already encountered reasons for being suspicious of top-down
processing. In the phoneme-restoration effect, for example, subjects hear
Figure 9.2
190 Perception
legislatures even though the sound corresponding to the first s has not ac-
tually occurred. Only some nonspeech sound, such as a pure tone, has oc-
curred. Clearly, reliance on context can breed error. It does not follow,
however, that top-down processing generally promotes error. On the con-
trary, the contexts used by speech perceivers may generally be reliable in-
dicators of the target phonemes, and hence may generally foster reliability.
Some further experiments make this point clear. One important line of
research on top-down effects comes from experiments on letter identifica-
' ~ D. D. wheeler.'' In these
tion, especially experiments by G. ~ e i c h e r and
experiments subjects were given a very brief presentation of either a letter
(such as D) or a word (such as WORD).Immediately afterward they were
given a pair of alternatives and instructed to report which they had seen. If
they had been shown D, subjects might be presented with D or K as alter-
natives. If they had been shown WORD, they might be shown WORD or
WORK as alternatives. The two word choices, of course, differ only in the D
and K letter. Results of the experiments showed that subjects were about 10
percent more accurate in the word condition. They more accurately discri-
minated between D and K in the context of a word than as letters alone,
even though they had to process more letters in the word context.
Similar effects have been obtained at the multiword level. In an experi-
ment by E. Tulving, G. Mandler, and R. Baumal, for example, subjects were
given varying amounts of context before being shown a target word in a
brief exposure.17 The greater the number of context words preceding the
critical word, the higher the percentage of correct identifications.
As these experiments clearly indicate, top-down processing can increase
reliability, at least for some kinds of tasks. It is noteworthy, indeed, that the
initial example of legisla.tures, in which top-down processing leads to per-
ceptual error, was a deliberately contrived case (like all experimental
cases). One might well speculate that in more ecologically natural settings,
top-down processing does not often produce such errors. At present, then,
we have no reason to think that top-down processing per se is a negative
characteristic of perception, at least from the vantage point of reliability
and justifiedness.
It is worth elaborating this point in a slightly different way. In the litera-
ture on perception it is common to distinguish realist and constructivist
theories of perception. Roughly, constructivism holds that the mind adds
things to the incoming stimulus in the course of producing a percept,
whereas realism tends to deny such activity. For my purposes, the label 're-
alist theory7 is likely to be misleading. So let me substitute a preferable
phrase suggested by Rock: stimulus theory.18 On Rock's characterization
stimulus theories hold that nothing more is necessary by way of explanation
of perception than the relevant stimulus. Such theories would of course
allow that various processes transduce the physical stimulus, encode it
Perception 191
neurally, and transmit it deeper into the brain. But as long as these pro-
cesses are "perfectly correlated" with the sensory input, what goes on in-
ternally is a "direct function" of the incoming stimulus, and can therefore
be ignored by perceptual psychology. The most promising forms of stimulus
theory would stress, however, that proper identification of the effective
stimulus is a subtle matter. The stimulus for perceived size need not be the
size of the object's visual angle subtended at the eye: it could be the rela-
tionship (the ratio) of the visual angle of the object to that of other neigh-
boring objects. This emphasis on higher-order features of the stimulus
derives from the work of J. J. Gibson, the most prominent perceptual realist,
or in my terminology, stinlulus theorist.''
Contrasting with the stimulus theory are the various sorts of constructi-
vist theories. As indicated, the common ground of such theories is that some.
internal constructive process occurs that mediates between the incoming
stimulus and the percept; something gets added to the stimulus.
We might say, very roughly, that constructivist theories tend to postulate
top-down processing, whereas stimulus theories postulate purely bottom-up
processing. Furthermore, the contrast between constructivist and stimulus
theories seems to invite conclusions about reliability. If the perceptual pro-
cesses add things to the stimulus during perceptual processing, isn't this
likely to distort the perceptual output? Won't there be a tendency for the
percept to get things wrong? Conversely, if stimulus theories are correct,
doesn't this bode well for the accuracy of perceptual processes? Indeed, this
prospect of accuracy, or truth, is already intimated by the very term 'realist'
(which is why I prefer to avoid it).
However, these conclusions are completely unwarranted. As the experi-
ments I have cited indicate, the deployment of information obtained from
sources other than the target stimulus need not be distorting. On the con-
trary, proper use of accurate contextual information can enhance reliabil-
ity. Furthermore, real-world regularities may have been incorporated into
perceptual systems in the course of evolution. So if, as constructivism often
suggests, perceptual systems automatically add such information in the
course of processing--or reveal innate preferences for certain kinds of an-
swers or solutions to perceptual questions-this need not be detrimental to
reliability.
Until now my discussion has been restricted to the standard of reliability.
But other standards of appraisal should also be explored. Let me turn first to
the standard of speed (or, rather, the combination of reliability and speed).
The relevant literature is readily interpreted as demonstrating the contri-
bution of top-down processing to increases in speed, the ability to make
correct perceptual judgments in less time. In the Tulving, Mandler, and
Baumal study, for example, subjects were shown a target word, such as dis-
order, for varying brief exposure times. In one condition this word was pre-
192 Perception
ceded by no context at all; in another it was preceded by eight words of

context, for example, The huge slum was filled with dirt and. For an expo-
sure time of 40 milliseconds, the percent of correct identifications with no
context was about 15 percent, whereas the percent of correct identifica-
tions with the context was about 50 percent. Clearly, use of context facili-
tated accurate beliefs when speed was ~ritical.~'
This certainly seems to show that perceptual systems gain speed by using
auxiliary information. But, interestingly, Fodor argues that speed is ac-
quired by perceptual systems' being designed to ignore lots of the facts, by
their informational The argument runs as follows. If per-
ceptual systems were allowed access to all information possessed by the or-
ganism, and they tried to search out all relevant information in long-term
memory before making a perceptual 'decision', speed would be greatly re-
duced. So the very unavailability of auxiliary information fosters speed, ac-
cording to Fodor.
Although there appears to be a conflict in these contentions, they are in
fact reconcilable. Fodor is assuming that if information were available from
central systems, the accessing of this information would have to be very
slow. If this were right, then access to all information in central-system
memory would indeed sacrifice speed. But it is not clear that accessing of
information from central systems is so slow. Moreover, I have conceded (see
note 20) that the information actually used in top-down processing of lin-
guistic stimuli may not be in the central system, and hence can be very fast
on anybody's view.
Since Fodor makes so much of speed as one of several defining character-
istics of input systems, we should pause to question the idea that perceptual
systems always work faster than central systems. Perceptual systems, like
central systems, sometimes take their time in arriving at a final conclusion.
Consider Rock's experimental work in which an observer is placed inside a
large, opaque, rotating drum.22After a period, the drum appears to the ob-
server to stop turning and he experiences himself as rotating instead (in the
opposite direction). There is generally a transitional period in which the
drum appears to be slowing down and the observer experiences himself as
beginning to turn slowly. Thus, the perceptual system ultimately 'prefers'
the self-motion interpretation, rather than depicting the surrounding en-
vironment as moving. But it takes the system some time to arrive at this in-
terpretation; it does not happen virtually instantaneously.
In any case, everyone apparently agrees that perceptual systems have
mechanisms that contribute to their speed. It is also widely acknowledged
that speed is a sensible trade-off for (some amount of) reliability, at least for
perceptual systems. First, life-threatening objects in the environment need
to be identified quickly. Second, many stimuli are potential need satisfiers;
they deserve at least preliminary identification. If such identifications were
Perception 193
slow, it would take too much time to make them all. From a pragmatic
vantage point, then, some accuracy is worth sacrificing for speed.
Now I have not thus far endorsed pragmatic values, such as need satisfac-
tion or desire satisfaction, as species of epistemic values. But an appeal to
such pragmatic factors certainly makes sense. Although I rejected this prag-
matic approach to iustijiedness (in section 5.4), this does not bar the prag-
matic dimension from appearing in a more inclusive survey of epistemic
values. It is worth asking, then, how well the perceptual systems perform
from this pragmatic standpoint.
In particular, let us consider how perceptual pattern recognition may be
influenced by affect: by strong desires and aversions, or other emotions.
Suppose that pattern recognition proceeds in terms of matching sense-
derived materials to a stored pattern representation. Then make three other
assumptions: (1)A match can be effected even if there is only partial corre-
spondence with the stored pattern. (2) Greater activation of a stored pattern
facilitates the matching of that pattern. (3) Patterns with high affective
value--either positive or negative-receive a higher level of activation.*
Given these assumptions, there will be an enhanced tendency to perceive
stimuli in high-affect categories. For example, a young child who has a
strong desire to see her mother will recognize lots of women in the distance
as her mother. The parent in the shower, listening for its infant's cry in the
next room, may hear such a cry when in fact it is only the noise of the
shower. A gardener fearful of snakes will see a snake when it is really only a
piece of garden hose.
These characteristics of a cognitive system probably deserve high marks
from a pragmatic standpoint. Pragmatically, erring on the side of false posi-
tives is prudent. Nothing much is lost if the child mistakes another woman
for her mother, if you mistake your garden hose for a snake, or shower noise
for your infant's cry. It is cost-effective to accept some false positives in
order to minimize errors of omission.
But processes that apply desire and aversion parameters are not condu- ,
cive to reliability. Affect-caused increases of activation are not well corre-

lated with the actual probability of a pattern's instantiation. The
desirability or danger of a category is a poor indicator of whether an in-
stance of it is actually present. So the class of perceptual identifications that
result essentially from this sort of process will have a lower truth ratio than
other perceptual identifications. There is reason to suspect, then, that right
systems of J-rules would not permit these sorts of processes. In other words,
right systems of J-rules might not license perceptual belief formation that
includes affect heightening of pattern activation. When perceptual beliefs
are formed in this way, they may not be justified.
A few words, finally, on the standard of power. The concept of power
stressed in Part I, it will be recalled, is a goal-responsive sense of power:
194 Perception
A
Figure 9.3
getting answers to the questions you want to answer. Now perception is

often described as a highly automatic process, which is surely true. But
goal-responsiveness is built into perceptual operations in certain ways. One
kind of responsiveness to desire or aversion was just mentioned in the pre-
ceding paragraphs. But a more pervasive mode of goal-responsiveness in-
volves attention. Attention allows perceptual systems to be selective about
the information they gather.
Two sorts of attentional mechanisms can be distinguished. First, the orga-
nism can orient the sensory organs: it can turn the head, focus the eyes, ex-
tend its limbs for touching, and so on. Second, it has internal attentional
mechanisms that allow it to filter out certain sensory input in favor of other
input. The classical studies on this topic involve dichotic listening experi-
ments, in which subjects wear a set of headphones and receive a message
through each. When asked to follow one of the two messages, most subjects
have little trouble. They tune out one message and attend to the other, just
as a person at a cocktail party tunes out the conversation of the group he's
facing and tunes in the conversation of an adjoining group.
Another example of attention, this time in the visual modality, involves
the decision to focus on the global shape of a figure as opposed to specific
details. Dramatic differences in a percept can result from such differences
in focus. Consider Figure 9.3. If one attends only to the global characteris-
tics of the two figures, they will appear virtually identical. But this impres-
sion quickly recedes when their parts are examined carefully. Indeed, half
of the homologous parts are different. It is clear, then, that perceptual sys-
tems can be guided by goals. This speaks well of them in terms of power.
9.4. Foundationalism and Coherentism

In this section I turn to a slightly different application of perceptual psy-
chology to epistemological concerns. A long-standing controversy in episte-
Perception 195
mology revolves around the structure of epistemic justification. The classi-

cal contending theories are foundationalism and coherentism. In Chapter 5
coherentism was rejected as a criterion of J-rule rightness. In so rejecting it,
however, I noted its possible tenability at a different theoretical level: the
level of substantive J-rule systems. The reliabilist criterion of rightness is
neutral about the precise character, or content, of right J-rule systems. As
far as reliabilism goes, a system--or its rules-could be either foundationa-
list or coherentist in character. Assuming that at least one right J-rule sys-
tem is feasible for human beings, what the rules in such a system are
depends on psychological facts. I want to examine this issue now in connec-
tion with perception. Perception is relevant because perceptual beliefs
comprise one class of candidates for 'foundational', or 'basic', beliefs.
I am proposing that the foundationalist-coherentist controversy depends
for its resolution on psychological facts, but I do not maintain that it de-
pends on these facts exclusively. Part of the controversy hinges on one's
exact conception, or definition, of these doctrines. The terms 'foundationa-
lism' and 'coherentism' have been used for a wide variety of distinct doc-
trines. Greater terminological precision is therefore needed before the
controversy can be resolved. Once a clear choice of terminology is made,
however, psychological facts have an important bearing. Certainly this is so
if perceptual beliefs are one candidate class of basic beliefs. The viability of
their candidacy for justificational 'basicness' depends on the processes that
generate such beliefs. This is a matter to be settled by the psychology of
perception.
Let us begin with the terminology: What is meant by 'foundationalism'
and 'coherentism' (henceforth 'F-ism' and 'C-ism')? These doctrines are
commonly formulated as theses about the structure of reasons or evidence.
For example, F-ism can be construed as the view that some justified beliefs
are directly or immediately justified, that is, not justified in virtue of being
supported by other beliefs. And the relation of 'support' may be taken as a
relation in which one belief provides reasons or grounds for a n ~ t h e r . ' ~
The reasons or evidence formulations, however, are not congenial to my
theory of justifiedness. Within my approach, F-ism and C-ism are best
viewed as theses about the contents of right J-rule systems. Before turning
to this manner of construing the doctrines, though, let us survey more tra-
ditional formulations.
Descartes is the originator of the F-ist idea. He thought that knowledge
should be placed on a firm foundation, and he sought to identify the nature
of that foundation. Descartes maintained that beliefs can be epistemically
firm, or secure, only if they are based on "certain" or "indubitable7' founda-
tions. This is often reexpressed as the view that foundational beliefs must be
infallible or incorrigible, that is, incapable of being false. This position now
has few adherents. If there is a special class of foundational beliefs--so-
196 Perception
called basic beliefs--few theorists would hold that they must be incorrigi-
ble. The essential requirement for basicness, then, remains to be specified.
Another facet of Descartes's F-ism held that the class of foundational be-
liefs consists in beliefs about one's own current mental states (or one's own
existence-as a purely mental entity). And this aspect of Cartesianism is
also unfashionable nowadays. F-ism could allow other kinds of beliefs to be
basic, for example, perceptual beliefs about physical objects or events in the
environment. Current epistemologists tend to formulate F-ism as a doctrine
that is neutral about the specific subject matter of the basic beliefs. I will
follow this neutralist mode of formulation.
Even within this core, neutralist conception of F-ism, there are two dis-
tinct variants: strong F-ism and weak F-ism. Strong F-ism says that the jus-
tificational status of basic beliefs is wholly independent of the rest of one's
doxastic corpus. Basic beliefs are justified no matter what other beliefs the
cognizer currently holds. Weak F-ism says that the justificational status of
basic beliefs is partly independent of the rest of one's doxastic corpus. There
is some nondoxastic source of justificational status, or warrant-but perhaps
some doxastic source as well. This characterization would still allow a dis-
tinction between basic and nonbasic beliefs, since nonbasic beliefs would
get whatever warrant they have entirely from other beliefs.25
We come now to the formulation of C-ism. One approach to C-ism sees it
as the denial of F-ism, another sees it as an autonomous doctrine. Viewed in
the first way, C-ism is the doctrine that there are no basic, or foundational,
beliefs. But this doctrine admits of different interpretations, given different
versions of F-ism. Thus, one version of C-ism is the denial of strong F-ism,
another is the denial of weak F-ism. The former doctrine would hold that no
beliefs receive their justificational status wholly from nondoxastic factors.
The latter would hold that no beliefs receive their justificational status even
partly from nondoxastic factors: all justificational status derives from rela-
tions with the cognizer's other beliefs.
The second approach to C-ism, the autonomous approach, would avoid
characterizing the doctrine as merely the denial of F-ism. It might give the
following sort of generic formulation:
A belief is justified if and only if it belongs to a coherent system of beliefs of
the cognizer.
This leaves at least two questions to be answered: (a) what is it for beliefs to
'cohere'? and (b) which system of the cognizer's beliefs must cohere in order
for a member belief to be justified?
Conceming question (a), answers might take a negative or a positive
form.26A negative answer might be: beliefs cohere just in case they are not
inconsistent. A positive answer might be: beliefs cohere just in case they
mutually support one anotherF7 Concerning (b), answers might again be of
Perception 19 7
two sorts: holistic and nonholistic. Holistic C-ism says that the relevant sys-
tem for purposes of the coherence theory is the cognizer's entire doxastic
system. Nonholistic C-ism would admit proper parts of an entire doxastic
system.
Under the influence of Quine's holism, many current formulations of
C-ism tend to be holistic. But there is a problem with this approach. Sup-
pose that the construal of 'coherence' implies that the lack of inconsistency
of the relevant corpus is a necessary, if not sufficient, condition for coher-
ence. Then if one requires holism, coherence of the entire doxastic corpus,
someone with an inconsistent set of beliefs cannot have any justified beliefs.
This is strikingly counterintuitive. If some inconsistency has arisen in one
comer of your doxastic corpus, it should not disqualify all the rest of your
beliefs from being justified. So it is doubtful that an optimal formulation of
C-ism is one that requires holism. Notice too that C-ism viewed as the de-
nial of F-ism need not involve holism.28Clearly, not all varieties of C-ism
need be holistic.
Tuning now to my own theory of justifiedness, the question is, What is it
for a given J-rule system to be F-ist or C-ist in character? Again, the differ-
ent conceptions of F-ism and C-ism sketched above will dictate different
replies. On one conception the critical question is whether any rules in the
set of J-rules permit the use of belief-forming processes whose inputs are
wholly nondmastic. If so, this set of J-rules would instantiate strong F-ism.
If not, it would instantiate C-ism viewed as the denial of strong ~-isrn."
On an autonomous and holistic conception of C-ism, however, the fore-
going would not be the critical question. According to holistic C-ism, a set
of J-rules will be C-ist only if the rules authorize only holistic processes,
processes that take as inputs all current doxastic states. In other words, a
process must be (in Fodor's terminology) "Quinean," in the sense of being
sensitive to the cognizer's entire belief system.30
My reliable-process criterion of rightness implies that whether a set of
J-rules is right depends on two things: (1)the cognitive processes available
to human beings, and (2) their reliability. Whether there are any right F-ist
J-rule systems, or any right C-ist J-rule systems, also depends on these facts.
Hence, whether right J-rule systems are F-ist or C-ist in character depends
on psychological facts.
Let me apply these points to the area of perceptual processes. It should
be recalled, in this context, that I interpret a percept as a belief, a percep-
tual system's judgment about environmental objects (see section 9.2).Such a
belief can be rejected by the central system. This is just what occurs when
vision 'says' that the stick immersed in water is bent, and the central system
'says' that it is straight. Admittedly, calling perceptual outputs 'beliefs' di-
verges from ordinary usage; but it is theoretically fruitful. It makes sense of
the intuitively compelling feeling that the central system disagrees with the
198 Perception
visual system when it judges the stick to be straight. Disagreement is best

construed as conflict in belief.
Thinking of percepts as beliefs, we confront the question, Are these be-
liefs justificationally basic beliefs? This must be investigated by reference to
perceptual processes. First, are any of these processes sufficiently reliable to
be included in a right set of J-rules? If so, the second question is: What is the
nature of these processes? Do they take prior doxastic states as inputs? In
particular, do they take all prior doxastic states as inputs? Are they sensitive
to the entire doxastic corpus?
If perceptual processes take no prior doxastic states as inputs, the result-
ing beliefs would qualify as basic according to both strong and weak F-ism.
However, if my analysis of top-down processing is correct, it is unlikely that
perceptual processes ignore other doxastic states entirely. Thus, the psycho-
logical facts are not congenial to strong F-ism, at least if the candidate class
of basic beliefs are perceptual beliefs.
However, several versions of C-ism also fare poorly if such theories of
perception are correct. First, holistic C-ism cannot be very plausible. As we
have seen, it is doubtful that perceptual processes have access to the cog-
nizer's total doxastic corpus: it is doubtful that they are Quinean. But if, de-
spite this, perceptual processes are sufficiently reliable to be sanctioned by a
right J-rule system, then holistic C-ism is undermined.
Another form of C-ism I defined also holds scant prospects for vindica-
tion. C-ism as the denial of weak F-ism says that there is no source of justi-
fiedness except other doxastic states. Within my framework this implies that
the only processes licensed by right J-rule systems would be processes that
take only doxastic states as inputs. But surely perceptual processes take
some sort of sensory events, or transducer activity, as inputs. And these sorts
of inputs presumably are not doxastic. Hence, assuming that some of these
perceptual processes are permitted, some licensed processes would contra-
vene the constraints of this version of C-ism.
I need not review all the variants of F-ism and C-ism to sustain my prin-
cipal conclusion. It should be amply clear how facts of perceptual pro-
cessing have a direct bearing on the viability of various forms of these
doctrines. Thus we see how epistemological conclusions in the area of justi-
fiedness hinge on facts that need to be determined by cognitive science.
chapter 10
Memory
10.1. Two Kinds of Belief

Another central component of cognition, and a main source of belief, is
memory. Admittedly, to call memory a 'source' of belief is a bit misleading,
since memory, when veridical, does not originate belief. But we can think of
the matter as follows. Belief is a dated affair. Though you believe a proposi-
tion now, there is no guarantee you will believe it later. That depends on
memory power. (This sense of 'power' is slightly different from that of
problem-solving power, a distinction I pursue in subsequent sections.) Simi-
larly, though you believe a proposition now, because you seem to remember
it from an earlier time, it does not follow that you did believe it earlier.
That depends on the accuracy or reliability of your memory. So what you
believe at any given time, whether correctly or incorrectly, often depends
upon memory. Epistemology, then, is rightly concerned with properties of
memory, in particular, with the power and reliability of memory. Memory
is an important item for epistemology's agenda.
Before proceeding, one matter of terminology should be mentioned. The
term 'remember' is often used in a factioe sense. That is, 'S remembers that '
p' entails the truth of p; and 'S remembers event E' entails that E really oc-
curred. For present purposes, though, I set aside this factive sense. When I
speak of remembering a proposition or event, I do not mean to imply that
the recalled proposition is true, or the recalled event actually occurred.
'Memoryy will refer, roughly, to ostensible memory.
Questions about the power and reliability of human memory will be
treated in later sections. First I wish to address another issue to which the
psychology of memory is highly germane. This concerns the concept of
belief.
One way psychology can impinge on primary epistemology, I have sug-
gested, is by refining its descriptive resources. In this chapter I shall illus-
trate this point by arguing that memory research demands a revision--or
enrichment-in the concept of belief. Twentieth-century epistemology has
largely rested content with the conceptual tools of ordinary language, in-
cluding the belief concept. But this concept is really psychologically inade-
quate. In the first two sections of this chapter, I will identify this
inadequacy and suggest how the psychology of memory can improve the
situation.
Many recent writers have worried about the content aspect of the belief
concept, as we saw in Chapter 8. The inadequacy of the belief concept on
which I focus here is quite different. There are (at least) two different kinds
of belief, and failure to distinguish these kinds and heed this distinction sys-
tematically undermines the attempt to provide satisfactory and realistic
epistemic principles.
In Chapter 9 I proposed a dichotomy between perceptual and nonper-
ceptual beliefs. In this chapter, though, I will ignore perceptual beliefs en-
tirely. The distinction examined here is a further distinction, wholly within
the more commonly acknowledged class of beliefs, so-called central-system
beliefs.
To believe a proposition at a given time, a person need not think of it, or
entertain it, at that time. Five minutes ago you doubtless had a belief about
where you were born, but you probably were not ruminating on your birth-
place five minutes ago. Yet a person often does entertain beliefs; they oc-
cupy attention and get actively invoked in cognitive tasks. These two
classes of belief are sometimes noted by philosophers of mind, who use the
phrases 'dispositional belief and 'occurrent belief to mark the two classes.'
But cognitive psychology would have us think of these two classes in a
slightly different way.
According to the popular 'duplex' theory of memory, the mind has two
memories: long-term and short-term (or actiue) memory. (A third kind of
memory, sensory information storage, is ignored here.) Long-term memory
(LTM) is a place where information can be stored for very long periods of
time, perhaps indefinitely. Information in LTM decays very slowly. Short-
term memory (STM), by contrast, has a much faster decay period. Unless
refreshed, or rehearsed, information fades fairly rapidly from STM. There is
also a big difference in the a m n t of information retainable in LTM as
compared with STM. LTM is virtually unlimited in capacity, whereas STM
has a very restricted capacity. Furthermore, whereas information lies dor-
mant in LTM and is not capable of influencing cognitive activities, STM is a
sort of 'warkplace', where various operations and transformations can be
performed on data, and new items can be constructed. In order to utilize
material stored in LTM, that material must be retrieved and copied into
STM. That's what it is to 'think of something one has previously known or
believed.
Although this duplex portrait of memory has been popular for about a
decade and a half, experiments during the 1970s have posed difficulties for
Memory 201
it, particularly for the notion of STM as a single entity. The properties used
to define STM have been split apart, so it is no longer clear which of them
(if any) should be taken as constitutive of a short-term store.2 ow ever, I
shall stick with the older model in my arguments. If anything, the new re-
search suggests that belief-state phenomena are even more complicated
than the older model suggests, an idea that would strengthen my case for
making distinctions. For expository purposes, it will be convenient to rest
content with the older model. It should be recalled, throughout, that my
discussion is only illustrative of the potential impact of psychological re-
search. Not all details of contemporary findings need detain us.
John Anderson's recent treatment of memory in The Architecture of Cog-
nition departs slightly from the duplex theory.3 Anderson views active
memories not as events occupying a special location, but as activated states
of LTM. Activation is something like a light bulb lighting up. Indeed, on
Anderson's theory, it is like a light bulb governed by a dimmer switch: acti-
vation can take on a continuous range of levels, not just on and off. For pre-
sent purposes, though, I will confine my attention to on and off states:
'activated' and 'unactivated' states. These can be identified with the two
sorts of belief states I spoke of earlier, which I will call activated beliefs-
A-beliefs-and unactivated be1iefs-U-beliefs4
Two objections might be lodged against the attempt to identify so-called
dispositional beliefs with states of LTM. First, it may be claimed, not all
states of LTM are belief states. Second, not all dispositional beliefs are
states of LTM.
The first objection has three elements. (1)Not all LTM states are belief
states because some LTM states are goal states. Our long-range goals, or in-
tentions, are stored in LTM; but these are not belief states. (2) Some things
may be stored in LTM, but neither as beliefs nor as goals. For example, a
sentence may be salted away in LTM because we like the sound of it, or just
because it has a sort of staying power in the imagination.5 This does not
mean that we believe the sentence. (3)Some contents stored in LTM are so
weak-whether through decay, repression, interference, or what have
you-that they are no longer retrievable, even with the best of retrieval
cues. It is not very plausible to say that these contents are still believed.
Argument (2) is not entirely persuasive. When a sentence is stored, there
may always be one or more beliefs associated with it, such as 'I have heard
this sentence before', or 'This sentence was uttered on such-and-such an oc-
casion'. However, elements (1) and (3)seem decisive enough. But what do
these points establish? Only that not all states of LTM are belief states; we
may still retain the view that some-presumably many-states of LTM are
belief states.
Turn now to the second objection, which claims that not all dispositional
beliefs are LTM states. Daniel Dennett, for one, is inclined to say that we
202 Memory
have indefinitely many beliefs that are not explicitly stored in LTM (or else-
where). His examples include: "New York is not on the moon, or in Vene-
zuela"; "Salt is not sugar, or green, or oily"; "Tweed coats are not made of
salt"; "A grain of salt is smaller than an elephant."6 These are all proposi-
tions a person may never think of, and hence are not stored in LTM. Still,
they are things he probably believes.
Personally, I find this claim unconvincing. If a person has never thought
whether New York is in Venezuela, then he does not already believe that
New York is in Venezuela. If he has never asked himself whether zebras
wear codpieces in the wild, then he does not already believe that zebras
don't wear codpieces in the wild. He may be disposed to assent to these
propositions unhesitatingly, the moment they are queried. But this just
shows he has beliefs from which these conclusions would be readily in-
ferred. It does not show he already believes each conclusion, prior to the
question being raised.
I acknowledge, however, that many philosophers do not share my intui-
tions. They hold there are non-occurrent beliefs that also are not explicitly
represented in LTM.' Whether or not they are right is not critical here,
however. I can accept the notion of nonexplicitly represented beliefs, as
long as we also retain the two categories of explicitly represented beliefs,
activated and unactivated beliefs. Since I want to restrict attention to these
two categories, I will cheerfully leave the issue of nonexplicitly represented
beliefs unresolved.
My main point about the distinction between A-beliefs and U-beliefs
concerns normative epistemic principles; I will come to this in section 10.2.
First let me comment, though, on the descriptive importance of this dis-
tinction. If the notion of belief is to be even moderately useful for predictive
and explanatory purposes, the distinction between A-beliefs and U-beliefs
must be observed.
Here is how a failure to observe this distinction can derail a prediction.
Suppose Melanie is deciding, one Sunday morning, whether to go to the
university library to work. She seldom works there on Sundays, but today
she needs to read a certain important article there. Now Melanie knows
that the library normally opens at 7:00 A.M. She has also noticed, on many
occasions, that the library's Sunday hours are different: it only opens at 1:00
P.M.If she were to recall the latter piece of knowledge, she certainly would
not make the trek to the library this morning; it would be a waste of time.
Alas, she forgets this piece of information and decides to go there after all.
When, upon arrival, she sees the posted hours, she upbraids herself as fol-
lows: 'Wasn't that silly of me! I knew the library doesn't open on Sundays
until 1:0OY.
Now suppose we worked with a belief-desire model of human decision
that countenanced only one kind each of belief and desire. Within each
Memory 203
class the only differences in efficacy would rest on strength. Then such a
model would presumably predict that Melanie would decide not to go to
the library this morning. After all, her corpus of beliefs and desires include
the following:
(B1) The library normally opens at 7:00 A.M.
(B2) Today is Sunday.
(B3) On Sundays the library opens at 1:00P.M.
(Dl) It is most desirable to read a certain article soon, which is only ob-
tainable at the library.
(D2) It is desirable to do a good bit of other work too today.
(D3) It is undesirable to waste time going places that don't allow essential
work to be done.
Since Melanie believes today is Sunday, and believes that the library does
not open before 1:00 on Sundays, the model should predict that she will de-
cide not to go this morning. This would be a wrong prediction, of course,
but one to which such a model seems committed.
Once we draw the distinction between A-beliefs and U-beliefs, and take
account of the fact that only A-beliefs are efficacious in decision making,
this prediction can be reversed. While the items listed above constitute
U-beliefs and U-desires-they are part of Melanie's LTM store-they do
not all get activated at the time of decision. At the moment of decision,
Melanie fails to retrieve the belief that today is Sunday; or fails to retrieve
the belief that the library opens at 1:00 on Sundays. The critical activated
belief at the time is the belief that the library normally opens at 7:00 A.M.
Now if a belief-desire model incorporates the A-belief/U-belief distinction,
and if it portrays decisions as governed by A-beliefs only, such a model
could correctly predict Melanie's decision to go to the library. Clearly, this
would be a better variant of the model.
It is a familiar occurrence that people often do what Melanie does-they
fail to access part of what they know, and therefore perform silly actions.
Failure to distinguish U-beliefs and A-beliefs makes it impossible to account
for these occurrences.
Turn now to theoretical inference, where the situation is perfectly analo-
gous. Which inferences a cognizer makes is determined not by his beliefs
stored in LTM, but by those that get activated, by his A-beliefs. A person's
U-belief corpus is enormous, and only a relatively small portion of it is acti-
vated at any one time. Cognizers often fail to make connections between
elements in this corpus, even though they are relevant to the theoretical
tasks at hand.
For example, at least a decade before Fleming's discovery of penicillin,
many microbiologists were aware that molds cause clear spots in bacteria
cultures; and they knew such a bare spot indicates no bacterial growth. Yet
they did not consider the possibility that molds release an antibacterial
Memory 205
credence to the hypothesis that molds release an antibacterial agent. This

was despite the fact that they knew both (a) molds cause bare spots in bac-
teria cultures, and (b) bare spots indicate inhibited bacterial growth. Does
this show that they were irrational? Again, this would be a poor diagnosis
of the episode. They probably failed to put these facts together because
they filed them under different categories in LTM. That molds cause bare
spots may have been filed under the category of practical laboratory lore,
as information on undesirable contamination; whereas the fact that bare
spots indicate inhibited bacterial growth may have been filed under the
category of microbiological theory." Their failure to retrieve these facts
jointly was probably due, therefore, to properties of their memory network;
or perhaps to procedures they used--or neglected to use-in searching that
network. But it is doubtful that this should be chalked up to a failure of
reasoning.
Let us not fuss too much over the words to use in characterizing the
breakdown. The more interesting question is why a person fails to retrieve
relevant facts from LTM. Of course, this question cannot be separated from
the question of why people regularly succeed in retrieving relevant facts. It
is only because memory retrieval is normally so powerful and efficient that
breakdowns like Melanie's are disappointing. Obviously, the two must be
studied together.
To explain what transpires in any particular case, at least three things are
relevant: (1)the general properties of LTM, (2) the particular individual's
history of encoding units into LTM, and (3) the particular sources of activa-
tion (which trigger retrieval processes) present on the occasion in question.
Although there is no unanimity among theorists on the general nature of
LTM, there is wide agreement that it is a highly organized body of material.
A popular approach depicts LTM as a complex network, containing many,
many nodes-each containing some sort of cognitive unit-and a complex
array of associative pathways, or links, connecting these nodes. These path-
ways, or links, are established by environmental stimuli and by the cog-
,
nizer's own internal processing. The particular history of such stimuli and
processing can strengthen or weaken memory traces. In particular, addi-
tional presentations of a single environmental stimulus tend to strengthen
the trace representing that stimulus. Stronger traces have a higher probabil-
ity (ceteris paribus) of retrieval. Paths, or links, also can have different
strengths, which also account for retrieval probability. A recent model
along these lines is found in Anderson's The Architecture of Cognition. I will
follow his version of this kind of theory.
Retrieval takes place by means of a process of spreading activation.
When a given cognitive unit is activated, the energy embodied in this acti-
vation spreads through the memory network along paths from original
source or sources to associated materials. Anderson theorizes that each node
has a limited capacity for spreading activation, so as more paths are at-
tached to it, the amount of activation that can be spread down any path is
reduced. He also assumes that activation spreads from multiple sources at
one time; activation converging on one node from multiple sources will add
UP.
Let us return to the case of Melanie. Since memory traces can have dif-
ferent strengths, it is plausible that the trace underlying the proposition
that the library normally opens at 7:00 is much stronger than the trace un-
derlying the proposition that it opens at 1:00 on Sundays. Melanie often
goes early to the library on weekday mornings but seldom goes there on
Sundays; nor has she had occasion to rehearse the posted hours, although
she has learned them and could reproduce them under some circumstances.
It would not be surprising, then, if the latter trace is much weaker than the
former. Furthermore, it would not be surprising if the paths between the
normal-hours representation and the Sunday-hours representation is not
particularly strong. If Melanie has seldom used the library on Sundays, and
perhaps never before tried to use it on a Sunday morning, she may have
done nothing to strengthen these paths. None of these facts points to any
epistemic flaw in Melanie's cognitive conduct.
Is there anything Melanie could-and should-have done at the time of
decision to activate the Sunday hours trace? Doubtless, had she directly
probed her memory with the question 'When does the library open on Sun-
days?' this would have activated the answer. But if she was not already tak-
ing note of the fact that Sunday is special as far as library hours are
concerned, why should she have asked herself that question? In short, al-
though she could have done things that would have activated the relevant
belief, there is no clear indication that she was epistemically culpable or
negligent in this failure. In particular, there is no indication that she was
irrational.
Although I have given much attention to this example of behavioral de-
cision, the same points carry over to cases of credal decision. In determining
the acceptability of a hypothesis, a cognizer must typically appeal to ante-
cedent beliefs. But these beliefs need to be activated before they can be
used in inference.
This point conflicts with the epistemological tenets of certain philoso-
phers, specifically with the view I will call inferential holism. Some of
Quine's writings hint at such a view, but it is most clearly endorsed by Gil-
bert Harman:
We have implicitly supposed that inductive inference is a matter of going
from a few premises we already accept to a conclusion one comes to accept
. . . But this conception of premises and conclusion in inductive inference is
mistakeii . . . The suggestion that only a few premises are relevant is wrong,
since inductive inference must be assessed with respect to everything one
believes . . . Our "premises" are all our antecedent beliefs; our "conclusion"
is our total resulting view.12
Memory 207
.
This doctrine of inferential holism, which sees inductive inference as in-
volving all our prior beliefs and all our resulting beliefs, is psychologically
most unrealistic. To be sure, there is a complex state of the cognizer that
comprises all his prior beliefs and another complex state that comprises all
his beliefs in the next moment of time. It is possible to conceive of inference
as a transition from one state to the next. But this departs dramatically from
the ordinary conception of inference (as a psychological affair, which is
what Harman is discussing). We normally think of inference as a causal pro-
cess, in which the premises are all causally operative in producing a new
belief (or other doxastic change). Yet it is psychologically impossible for all
one's beliefs to be causally operative in any process of inferential transfor-
mation. Only activated beliefs are causally operative; and one cannot (si-
multaneously) activate all of one's beliefs.
Normative principles of inductive inference should not, therefore, re-
quire cognizers to utilize all of their belief corpus. Inductive inference
should be divided into two parts: (a) search of LTM, in the effort to acti-
vate, or retrieve, relevant evidence; and (b) weighing of the activated evi-
dence. Epistemologists have attended almost exclusively to (b), but (a) also
needs attention. Specifically, what cues are given to memory during the
search process, and what is the duration of the search process? Normative
principles might well be introduced in these areas. Furthermore, there is a
third kind of activity, which occurs prior to (a) and (b), and which heavily
influences them. This is (c) the encoding process, which lays down the path-
ways through which activation spreads in the memory network,
If a person reflects on an item of information, and tries to connect it to
other things he knows, this can establish a richer network of associative con-
nections, which will be operative in subsequent retrieval tasks. In other
words, a person can deliberately integrate an item of information with
either a smaller or wider body of belief. Memory research indicates that
deeper integration (sometimes called 'depth of processing') can increase re-
call of an item. But I am here concerned with the range of connected items .
that can be activated by a given item. The scientist who learns a new fact
may either (1)simply file it away, or (2) deliberately trace out its implica-
tions (if any) for all sorts of other things he knows. This effortful inquiry will
often pay off in later inferences. The pathways established thereby will
broaden the range of potentially relevant evidence that can subsequently be
activated. By the same token, a scientist who fails to engage in such effortful
inquiry will fail to activate relevant evidence on later occasions. This failure
may result in negligence because he had earlier failed to lay down a suffi-
ciently wide set of memory connection^.'^
I am not prepared to propose specific normative principles in this area. I
merely emphasize the implausibility of any principles that ignore the differ-
ence between A-beliefs and U-beliefs, or the psychological mechanisms that
govern activation.
208 Memory
10.3. Elaboration and Reconstruction

I have been assuming, in the foregoing discussion, that activation of mate-
rial from LTM is always veridical, that what we retrieve from LTM revives
episodes or information that were indeed once encoded there. This assump-
tion is untenable. Memory can certainly yield mistakes. The question is:
How pervasive is the unreliability of memory, and what specific mecha-
nisms account for such unreliability?
In speaking of memory unreliability, I am speaking of the production of
false A-beliefs, somehow produced through memory mechanisms. It is pos-
sible, of course, to acquire false beliefs in the first place, and then retain and
retrieve these beliefs accurately. In such cases the falsehoods are not due to
memory. I am interested in the ways that error may creep into a doxastic
corpus specifically through memory mechanisms.
As elsewhere, though, I am not concerned with just reliability; I am
equally concerned with power. Even infallible memory mechanisms would
not be very valuable, on the whole, if they generated a mere trickle of recol-
lections. Clearly, powerful memory mechanisms are valuable, even at some
sacrifice in reliability. Let us inquire into this mix of traits among human
memory mechanisms.
Before proceeding, a word about the term 'power'. In talking of memory
power, I am using the term slightly differently than before. At least four dif-
ferent dimensions, or components, seem relevant to the power of a memory
system. Although these dimensions are generally relevant to the question-
answering sense of 'power-the ability to retrieve a stored item of infor-
mation is one way to answer a question-the dimensions are not severally
reducible to question-answering power.
Here are the four dimensions of memory power I have in mind: (1)How
easy is it to store information initially in memory? (According to standard
theories, not all information acquired by the mind forms a trace in LTM.)
Presumably, the easier material can be stored, the greater the memory's
power. (2) How long does the memory system retain the information? The
longer the storage period, the greater the system's power. (3)How much in-
formation can be concurrently retained in memory? The greater the quan-
tity, the greater its power. (4) How easy is it to access stored materials?
More specifically, how many different cues or probes will activate the mem-
ory trace? The greater the access range, the greater the power. For exam-
ple, if the only probe that will activate the trace of a given stimulus is a
re-presentation of the same stimulus, this is very restricted accessibility
(mere 'recognition memory'). It is commonly desirable to activate a trace
with other cues or probes, and the more the better.
There are many aspects of human memory mechanisms that bear on reli-
ability and power. One such aspect is 'interference', an oft-cited cause of
Memory 209
both inability to recall and errors of recall. But in this section I shall con-
centrate on two other memory processes that psychologists have investi-
gated: elaboration and reconstruction.
J . D. Bransford, J. R. Barclay, and J. J. Franks studied the ability of sub-
jects to recognize previously presented sentences, such as The ants were in
the kitchen, The ants ate the sweet jelly, The ants in the kitchen ate the jelly
which was on the table, and The jelly was sweet.l4 When subjects were sub-
sequently asked to try to recognize sentences, some of which had been pre-
sented and others of which had not, they tended to respond on the basis of
how much of the total information a test sentence conveyed. They were
more likely to misrecognize a nonpresented sentence involving all the pre-
sented semantic information-for example, the ants in the kitchen ate the
sweet jelly which was on the table-than a shorter sentence which had in
fact been presented. Apparently, subjects encode the sentences in a seman-
tic structure that goes beyond the individual sentences actually presented.
An interpretation is constructed that elaborates upon the information ac-
tually presented.
Another example of this elaborative process occurs in an experiment by
R. A. Sulin and D. J. ~ o o l i n ~ .One
" group of subjects read the following
passage :
Carol Harris was a problem child from birth. She was wild, stubborn, and
violent. By the time Carol turned eight, she was still unmanageable. Her par-
ents were very concerned about her mental health. There was no good insti-
tution for her problem in her state. Her parents finally decided to take some
action. They hired a private teacher for Carol.
A second group of subjects was given the same passage, except that the
name Helen Keller was substituted for Carol Harris. A week after reading
the passage subjects were given a recognition test in which they were pre-
sented with a sentence and asked to judge whether it was in the passage.
One of the critical sentences was She was deaf, dumb, and blind. Only 5 .
percent of the subjects who read the Carol Harris passage accepted this
sentence, but 50 percent of the subjects who read the passage as about
Helen Keller thought (mistakenly) they had read this sentence. Apparently,
the latter subjects had elaborated the story they actually read with facts
they previously knew about Helen Keller.
These experiments leave it indeterminate whether construction occurs at
the time material is originally encoded into LTM or whether a reconstruc-
tive inference occurs at the time of test. Other experiments, however, seem
to demonstrate that both of these phenomena commonly occur.
An additional experiment on the Carol Harris passage by Dooling and
R. E. Christiaansen shows that inferences can be made at the time of test.16
They had subjects study the passage and told them a week later, just before
210 Memory
test, that Carol Harris really was Helen Keller. Subjects still made many
errors, accepting such sentences as She war deaf, dumb, and blind. These
subjects were clearly making an inference at the time of test (or recall).
But elaboration at the time of encoding is also well accepted and docu-
mented. T. S. Hyde and J. J. Jenkins had subjects read groups of twenty-four
words presented at the rate of 3 seconds per word." One group was asked
to check whether each word had an e or a g. The other group was asked to
rate the pleasantness of the words. It is reasonable to assume that the pleas-
antness rating was a task that required more elaborate processing than the
letter verification task. In rating a word for pleasantness, subjects had to
think about its meaning, which gave them an opportunity to elaborate on it.
For instance, a subject presented with duck might think, 'Duck--oh yes, I
used to feed the ducks in the park; that was a pleasant time'. Subjects rating
for pleasantness showed much higher recall than those with the letter veri-
fication task. This was apparently due to elaborative processing.
In the experiments just reviewed elaboration and reconstruction often led
to false recall judgments. So elaboration and reconstruction seem to foster
unreliability. This fact is important not only in abstract, theoretical ap-
praisal of memory mechanisms, but also in practical contexts, where we
need to decide whether to trust someone else's memory judgments. Such
practical exigencies arise in the courtroom, where decisions appeal in part
to the memories of eyewitnesses. The memory mechanisms I have been de-
scribing offer opportunities for errors in eyewitness testimony, as Elizabeth
Loftus has documented.18
One problem in eyewitness testimony is simply the period of retention.
R. N. Shepard tested thirty-four clerical workers for recognition of pictures
after intervals of two hours, three days, one week, and four months.'' He
found that retention of the picture material dropped from 100 percent cor-
rect recognition after two hours to only 57 percent correct after four
months. The latter figure essentially represents mere guessing. However,
not concerned at the moment with retention intervals, I want to look at the
impact of reconstruction and elaboration on recall. Loftus cites relevant
work on this topic, including some of her own studies. Specifically, she dis-
cusses how post-event information-information given after a witnessed
event occurred--can distort one's recollection of that event.
An early illustration of this phenomenon was presented by C. ~ i r d . ~ '
During the course of a routine classroom lecture, the instructor was discuss-
ing the results of a series of experiments. A well-meaning but not very
thoughtful reporter on the local newspaper printed an account of the lec-
ture riddled with errors. Many students read the newspaper account, and
nearly ail of these thought it was accurate. The instructor gave an exam at
the end of the week and after the usual set of exam questions asked students
to indicate whether they had read the press account. Those who had read
Memory 21 1
the article made many more errors on the exam. They remembered the er-
roneous information from the newspaper, assuming they had learned it
from the instructor's original lecture.
In several experiments Loftus and colleagues tested subjects to see
whether post-event misinformation would distort their later recollection. In
one experiment nearly 200 subjects viewed a series of thirty color slides de-
picting successive stages in an accident involving a car and a pedestrian.21
The car was a red Datsun shown traveling along a side street toward an in-
tersection, where there was a stop sign for half of the subjects and a yield
sign for the others, and where the accident occurred. Immediately after
seeing the slides, the subjects were asked some questions, one of which was
critical. For half of the subjects the critical question was, Did another car
pass the red Datsun while it was stopped at the stop sign? The other half of
the subjects were asked the same question with stop sign replaced by yield
sign. For some subjects, the sign mentioned was the one they had actually
seen: the question gave them consistent information. For the remaining
subjects, the question contained misleading information.
Later a recognition test was administered. Pairs of slides were presented
and subjects had to indicate which member of each pair they had seen be-
fore. The critical pair was a slide depicting the Datsun stopped at a stop
sign and a nearly identical slide depicting it at a yield sign. When the inter-
vening question had contained consistent information, 75 percent of the
subjects responded accurately. When the question had contained mislead-
ing information, only 41 percent responded accurately. If subjects had been
simply guessing, they would have been correct about 50 percent of the
time; so the misleading information contained in the question reduced their
accuracy below the guessing level,
Apparently, people tend to use post-event information to reconstruct
episodes they have observed. Furthermore, they are not always able to dis-
entangle the original sources of their information. So even though they ac-
quire certain information (or misinformation) after the event, they often
think it was contained in what they saw.
Is this another piece of evidence for the unreliability of the reconstruc-
tion process? Superficially, this is so. But we should proceed cautiously. In
Loftus's cases, error is introduced into the subjects' cognitive systems by
misleading questions provided by the experimenter. The initial fault, then,
does not reside in memory, at least not in the unreliability of memory. It is
true that the mistakes could be averted if the subjects confidentIy recalled
exactly which sign-a stop sign or a yield sign-they originally saw. If so,
they presumably would not be misled by the question, for they would reject
what the question presupposes. But failure to recall this accurately is a defi-
ciency in memory power, rather than in memory accuracy. Similarly, any
failure to disentangle or isolate the original sources of information could
212 Memory
more plausibly be attributed to memory power than to memory accuracy.

More precisely, it is a limit on power that, in these special circumstances, is
partly responsible for inaccuracy. Still, since the error is introduced by an
extramemorial source, it is not clear how badly this reflects on the recon-
struction process.
Even if the reconstruction process is deemed inaccurate, and the elabora-
tion process is as well, we should not neglect the fact that the elaboration
and reconstruction processes make substantial positive contributions to-
ward memory power. Even if these processes have flaws as judged by the
reliability standard, they provide substantial trade-offs in power.
One way of explaining this is in terms of the so-called depth of processing
principle, developed by F. I. M. Craik and R. L. ~ o c k h a r tThey
. ~ ~ and other
researchers suggest that more fully elaborated material results in better
memory, in my terminology, more powerful memory. Manipulations that
increase the 'depth' to which information is processed-roughly, the extent
to which the subject processes the material's meaning-result in better
memory. The Hyde and Jenkins experiment cited above is one such case.
Another demonstration is recounted by S. Bobrow and G. H. ~ o w e rThey .~~
had subjects try to commit to memory simple subject-verb-object sentences.
In condition 1subjects were provided with sentences written by the experi-
menters. In condition 2 subjects had to generate a sentence themselves, to
connect the subject noun and the object noun. In condition 1 the level of
recall was 29 percent; in condition 2,58 percent. Presumably, in generating
their own sentences, subjects had to think more carefully about the mean-
ing of the two nouns and their possible interrelationships. This additional
mental effort, or deeper processing, seems to have led to more elaborations
concerning the nouns, which apparently produced the dramatic difference
in recall level.
To illustrate how elaboration can improve recall, consider this example
presented by ~ n d e r s o n . 'Suppose
~ a subject must commit to memory the
following sentence:
1. The doctor hated the lawyer.
A subject presented with this sentence is unlikely to deposit only this struc-
ture in LTM. He will probably have other thoughts while studying the sen-
tence, which might also be committed to memory. So, he might store the
following propositions in addition:
2. The subject studied this sentence in the psychology laboratory one
dreary morning.
3. The lawyer had sued the doctor for malpractice.
4. The malpractice suit was the source of the doctor's hatred.
5. This sentence is unpleasant.
Memory 213
He may have thought of proposition 4 because he had another proposition

in memory, namely,
6. Lawyers sue doctors for malpractice.
In storing all the above items in LTM, the subject creates a much more
complicated memory structure, which incorporates the elaborations on the
original sentence.
How can these elaborations lead to better memory? In two ways. First,
they provide additional retrieval routes for recall. Suppose that at the time
of test, the subject is given the word doctor and asked to retrieve the origi-
nal sentence. The link from the node representing doctor to the target node,
representing the sentence The doctor hated the lawyer, may be too weak to
revive the latter. But because elaborations have led to additional associative
links, recall may still be possible. From the prompt doctor, the subject might
recall the proposition that the lawyer sued the doctor for malpractice. And
from here the subject might be able to recall the target node. Thus, an al-
ternative retrieval route may succeed if the more direct one fails.
A second way in which elaboration aids memory is that it can enable
someone to infer the target information. For example, the subject who can-
not immediately recall the target sentence from the prompt doctor might
think as follows:
I cannot remember the target sentence but I can remember conjecturing
that it was caused by the lawyer suing the doctor for malpractice and I can
remember it was a sentence with a negative tone.
From this he might infer that the target sentence was The doctor hated the
lawyer. L. M. Reder and R. J. Spiro have argued that in real life a great deal
of inference is involved in Talk of retrieval or activation may be
somewhat misleading, then, if the so-called remembered beliefs are heavily
influenced by inferential reconstruction.
In looking for epistemological consequences, we should not rush to the
conclusion that elaboration and reconstruction yield unjustified beliefs.
Whether the processes would be licensed by a right J-rule system depends
on their contribution to such a system's truth ratio. If both elaboration and
reconstruction are well constrained by prior information in LTM, if the lat-
ter information is true, and if the inferential processes are suitable, the con-
tribution to the rule system's truth ratio may well be acceptable. (Of course,
this will also depend on the chosen degree of reliability needed for
justifiedness.)
Notice too that there are all sorts of amounts and styles of elaboration and
reconstruction that could be discriminated. Some of these amounts and
styles would have a more deleterious effect on truth ratios than others. A
214 Memory
right J-rule system might license processes incorporating some amounts and
styles of elaboration and reconstruction, but not others.
The processes of elaboration and reconstruction exemplify a situation
that we have encountered before and will encounter again: a trade-off be-
tween reliability and power. It is clear that these processes (can) purchase
increases in memory power, though perhaps at some cost in reliability.
Our native, uncritical procedures seem content with the purchase price,
and nothing in my principles of epistemological appraisal dispute its
appropriateness.
Actually, though, it is not clear that there is always a net trade-off, that
power and reliability always conflict. Although I often speak of acquiring
power at the cost of reliability, power can also foster reliability in a J-rule
system. At least a case for this suggestion is readily made, in connection
with inductive inference. First, the more experiences a person correctly re-
calls, the larger the number of true beliefs available as premises for induc-
tive inference. Second, the larger the number of true premises employed in
inductive inference (that is, the greater the total evidence that gets used),
the greater the likelihood of arriving at true conclusions. So there are ways
in which recall power can promote the overall reliability of a cognitive sys-
tem, or the reliability of a set of J-rules. To some degree, then, power sup-
ports reliability, and does not conflict with it.
10.4. Belief Perseverance

In this section I examine mechanisms of belief 'perseverance': mechanisms
that retain beliefs even after supporting evidence has been totally under-
mined. Certain psychologists have not only studied perseverance empiri-
cally, but evaluated it epistemologically. I explore their empirical findings
and raise questions about their epistemic evaluation. On the surface perse-
verance has little to do with memory. But perseverance may be largely the
result of the process of elaboration, discussed in the previous section.
Hence, epistemic evaluation of perseverance must go hand in hand with
evaluation of elaboration.
Perseverance has been studied mainly by Lee Ross and his colleagues. I
draw extensively on the discussion of the phenomenon in his and Richard
Nisbett's Human Inference: Strategies and Shortcomings of Social Judg-
ment.2sIn an early study by Ross, M. R. Lepper and H. Hubbard subjects
were given the task of distinguishing authentic from inauthentic suicide
notes.27As they worked, they were provided with false feedback after each
trial. This feedback indicated that, overall, they had performed at close to
an average level, at a level much above average (success condition), or at a
level much below average (failure condition). This feedback was simply
manipulated by the experimenters, and had nothing to do with a subject's
Memory 215
actual performance. Subjects were later thoroughly debriefed concerning

the random nature of their feedback. They were told that their feedback
had been false, and were shown the experimenter's instruction sheet assign-
ing them to the success, failure, or average performance condition. Subse-
quent to this debriefing, subjects were asked to fill out a postexperimental
questionnaire on which they had to estimate their actual performance at
the task, to predict their probable success on related future tasks, and to
rate their ability both at the suicide discrimination task and at other related
tasks involving social sensitivity.
Nisbett and Ross report that this first experiment revealed a "remarkable
degree" of postdebriefing perseverance. Even after debriefing, subjects who
had initially been assigned to the success condition continued to rate their
performance and abilities far more favorably than did subjects whose initial
feedback had indicated average performance, while subjects initially as-
signed to the failure condition showed the opposite pattern of results.
To replicate this perseverance phenomenon, a second experiment was
undertaken, which included both actor subjects and observer subjects
(viewing the events from behind a one-way mirror). The observer subjects
witnessed the initial (false) feedback presentations to the actors and the
later debriefings given by the experimenter. They were later asked to rate
the actors' success and ability. Also, two additional experimental conditions
were incorporated, yielding three in all. One group of subjects was given
no debriefings. A second group of subjects was given outcome debriefings,
the same sort of debriefings as in the first experiment. A third group of sub-
jects was given process debriefings, which not only told them about the
random assignment and false feedback, but provided an extensive discus-
sion of the perseverance phenomenon, the processes that might contribute
to it, and the potential personal relevance and costs of erroneous impression
perseverance.
Results from this experiment are shown in Figure 10.1. Among the sub-
jects who were given outcome debriefings, there was still substantial perse-
verance of the initial impressions for both actors and observers. But much of
the initial impression was changed as a result of the debriefings. When sub-
jects were given process debriefings, the initial impressions were eliminated
(or overridden) much more completely, except for observer subjects. Ross
and colleagues have done similar experiments involving other tasks, and
these too show the perseverance effect to some degree or other.*
W h y does perseverance occur, when it does? And how should the phe-
nomenon be evaluated epistemically? These two questions are intertwined,
on my approach to epistemic evaluation, because the evaluation of a belief
state is a function of the processes that generate it.
Nisbett and Ross consider two kinds of possible explanations of the perse-
verance phenomenon. The first, emotional commitment to one's beliefs,
216 Memory
No Debriefing
Outcome Debriefing
Process Debriefing
-
Actors Observen Actors Observers Actors Observers
Estimated Predicted Rated Task
lni tial Future Ability
Success Success
Figure 10.1
they find unsatisfactory, preferring more purely information-processingex-

planations. The second explanation, which they endorse, is that subjects
search for additional information in memory to support the initial (false)
feedback, and such information, once found, continues to support the initial
I
impression even after the feedback is discredited.
Another possible explanation not considered by Nisbett and Ross is that
the subjects do not fully trust the experimenter's debriefing. Such distrust
would be reasonable. After all, the experimenter admits he has lied to them
already. Why not entertain the possibility that the debriefing is itself a de-
ception? I will not pursue this suggestion further, however.
Let us return to Nisbett and Ross's explanatory hypothesis, which they
amplify as follows. Consider a hypothetical subject Jane, who receives
feedback suggesting that she is uncannily successful at the suicide note dis-
crimination task. She has no trouble generating additional 'evidence' that
seems consistent with her apparent social sensitivity. Her reasonably good
performance in her abnormal psychology course, her ability to make new
friends easily, and her increasing sense of confidence and assurance as she
progressed in the suicide note task, all might be seen as further evidence of
Memory 217
such powers. She can also easily find causal explanations of her perform-
ance: her familiarity with the writings of a famous novelist who recently
committed suicide, or her part-time job as a paramedical assistant, or her
open relationship with her parents all might explain her high level of ability
at a task requiring social sensitivity. Once these additional facts or explana-
tions are generated, they are found convincing, and they remain to support
her assessment of her ability even after she receives a debriefing from the
experimenter.
Although there is no direct support for this account of the perseverance
phenomenon, I agree that it is quite plausible. In the remainder of my dis-
cussion I will assume that this sort of process underlies the phenomenon.
Notice, though, that this process can be redescribed as an instance of elabo-
ration. Jane is finding other things in memory with which her apparent good
performance on the suicide note task coheres. She is finding a meaning in
this good performance by linking it up to other (putative) facts in her life or
traits she possesses.
Let us now turn to the normative appropriateness of this sort of process.
Nisbett and Ross are highly critical of it, seeing it as an instance of two
habits, or propensities, which they deem normatively wrong. First, they see
it as an instance of confirmation &us. They cite a rich research literature
"that shows the operation of a variety of encoding and decoding biases that
favor confirmation of prior hypotheses or beliefs over disconfirmation . . .
People tend to recognize the relevance of confirming cases more readily
than that of disconfirming ones, and therefore tend to search for such cases
in evaluating their hypotheses."29 They cite experiments by P. C. Wason
and P. N. Johnson-Laird, M. Snyder and W. B. Swann, and Snyder and M.
Cantor, in support of this claim.30
Second, they see the hypothetical process as an instance of searching for
causal explanations, which are too readily generated and accepted even
when the evidence for them is skimpy or nonexistent. Such explanations
have a marked effect, they say, even when later undermined. They cite an
experiment by L. Ross, M. R. Lepper, F. Strack and J. L. Steinmetz in which
subjects were asked to place themselves in the position of a clinical psychol-
ogist trying to understand and predict a patient's b e h a ~ i o r . Subjects
~' were
given an authentic clinical case history, and some were asked to use this
case history to explain a critical event in the patient's life, whereas other
subjects were not asked to give any such explanation. After the explanation
task subjects were informed that the events they had been asked to explain
were purely hypothetical, and that there was no available information
about the later life of the patient. Then subjects were asked to assess the
likelihood of a number of possible events in the patient's later life, inchding
the critical events. The subjects' likelihood estimates revealed a marked ef-
fect of the explanation task. Subjects who had explained an event rated it as
218 Memory
more probable (even after debriefing) than subjects who had not explained
it.
We should not accept Nisbett and Ross's characterizations of people's
habits too uncritically. Concerning confirmation bias, for example, they
themselves admit that some research indicates that confirming instances are
not always more 'available' to a theory holder than disconfirming ones. As
R. Hastie and P. A. Kumar noted, surprising or incongruent events may be
attended to and stored in memory more often than expected or hypothesis-
confirming events.32 Concerning causal explanation, it is not clear just how
readily people manufacture causal explanations. That they readily invent
causal explanations when given such a task by an experimenter hardly
shows that they always invent them spontaneously, even on the slimmest of
evidence. Clearly, much empirical work remains to be done on these
matters.
Turning to the normative question, suppose that people regularly search
for other material in memory that coheres with, or even explains, some
newly acquired belief. Is this good or bad? An epistemologist might disagree
with Nisbett and Ross's negative assessment on the grounds that search for
coherence is a desirable epistemic trait, and that the search for possible
causal explanations is one piece in such a coherence search. Such searches
are a legitimate-perhaps even a required-phase of belief formation. Un-
less a cognizer checks for coherence, his chances of true belief formation are
greatly reduced. (For some further discussion of this point, see Chapter 15.)
Of course, we should distinguish between attempts to assess coherence
and the specific procedures used for assessing coherence. Perhaps Nisbett
and Ross mean to be critical of the specific procedures people use, not their
proclivity to investigate the question of coherence. But three responses
should be made here. First, since many debriefed subjects do not suffer from
belief perseverance, as the studies show, these subjects may be using per-
fectly acceptable procedures (even by Nisbett and Ross's standards). Sec-
ond, among the subjects who do persist with their initial belief, despite
debriefing, their coherence-determining procedures may be all right.
Maybe they legitimately found reasons to bolster this belief (even if the be-
lief is false). Third, we should be quite prepared to learn that subjects differ
from one another in their coherence-determining procedures; that some of
them fail to conform to any right system of J-rules whereas others do so
conform. After all, subjects differed in their post-debriefing test responses,
so they should not all be lumped together. Nisbett and Ross have a tendency
to make blanket statements about human traits, without full allowance for
observed variation in subject responses.
However, there is a further point, specially connected to memory, that I
have yet to stress. I remarked above that the tendency to look for anteced-
ent beliefs which cohere with a new item of information may be classified as
elaboration: the cognizer seeks to connect the new item with other things
previously stored in memory. In the previous section we saw that elabora-
tion serves to strengthen memory, that is, to improve recall at later junc-
tures. This increase in memory power is surely an epistemically desirable
outcome. It should be weighed as one factor in the overall normative assess-
ment of the process in question. If it is significant enough, it may render the
process, on balance, normatively acceptable.
There are several complications that need mentioning here. First, I have
not offered any principles for overall normative assessment. I have only of-
fered a criterion of justifiedness, and this does not incorporate the element
of power. (Arguably it should; but for now let's put that aside.) However,
we also saw at the end of section 10.3 that power can positively contribute
toward reIiability. So elaboration may enhance reliability via its enhance-
ment of power, and this could bolster its claim for admissibility by a right
system of J-rules. Nisbett and Ross, although they speak frequently of
"human judgmental failings" and even "perversity," offer no principles for
normative evaluation^.^^ Perhaps they would not count power as a relevant
factor at all, for normative purposes. And perhaps they would not accept
the truth-ratio consequentialism that I have proposed. This just underlines
the need for general criteria of epistemic evaluation.
Suppose we focus, though, on my own criterion of justifiedness. It is a
very delicate matter to decide just which processes, among those presently
at issue, should be approved by a system of J-rules. Since it is always possi-
ble that a newly acquired belief is false, how much should such beliefs be
strengthened in memory? How much should they be combined with other
material in memory to generate new beliefs? In the case of Jane, should the
subject combine the (apparent) fact that she has performed well on a cer-
tain task with other facts in memory to make inferences about her general
traits (for example, her social sensitivity)? When focusing on these experi-
mental cases, where it is a datum-to the reader!-that the newly acquired
belief is false, it is all too easy to say that the belief should not have been
combined with others to generate further inferences. We are inclined to say .
this because we know that the belief is false. But a J-rule system needs to
authorize general processes. A general process of finding coherent beliefs
and making new inferences may have a high truth ratio, as long as the class
of input beliefs has a high enough truth ratio. In short, a right J-rule system
might well sanction just what many (even most) of the subjects in Ross's ex-
periments do. Many of those subjects arrive at errors; but these errors are
only a tiny sample of the set of cases that determine the relevant truth ratio.
I have been discussing the normative status of belief perseverance in
terms of the (hypothetical) processes that are actually at work. But perhaps
this is the wrong angle. Instead of focusing on what subjects do that they
shouldn't, perhaps we should focus on what subjects fail to do that they
should.
At time t,, Jane forms a belief (based on feedback) about her task per-
220 Memory
formance; call this p. At time t2, she infers from p (plus other beliefs that p
activates) something about her general ability in a certain domain; call this
q. At time t3, she is debriefed: she is told that p is really false, that her per-
formance at the task did not have the property she was led to believe. What
should she do now? There is an initial feeling that she should go back and
revise both her belief in p and her belief in q. In section 10.5 I will raise
queries about what such revision could consist in. The question to be ad-
dressed now is: Is Jane obliged to go back in order to revise her belief in p
and her belief in q?
When Jane is debriefed, this activates her belief in p, and it is plausible to
hold that this belief should be revised. But the debriefing may not activate
her belief in q, or her beliefs in any other propositions that are partly trace-
able to her belief in p. So it may not occur to her to revise any of these other
beliefs. Is this a culpable omission? Does her failure to backtrack and re-
cover those other beliefs, and then revise them, render those beliefs
unjustified334
Notice that my basic theory of justifiedness does not imply that merely
because Jane's belief in q at t, is justified, it is also justified at t3. My frame-
work does not guarantee automatic continuity in justifiedness, even for a
belief that remains inert in LTM. Possibly, a right system of J-rules would
require a cognizer to backtrack wheneuer new beliefs are acquired, to make
selected revisions in antecedently entrenched beliefs. If these old beliefs are
retained, contrary to this requirement, they become unjustified.
But would a right rule system really make such a requirement? It should
first be stressed that a rule system cannot simply say: 'Revise all U-beliefs
that have been undermined by new evidence'. For this injunction does not
specify any basic cognitive processes, and only such processes are allowed in
the rules. What a rule may do is instruct the cognizer to search LTM for old
beliefs that could be undermined by the new evidence, or perhaps to search
LTM for old beliefs that were originally generated by beliefs one has now
abandoned. The rule could then say that any beliefs so identified should be
abandoned.
A rule formulated in one of these ways would certainly be admissible. But
one could conform with such a rule and not succeed in weeding out all (or
even many) of the beliefs that are the proper targets of the search. First,
even if one searches for old beliefs under the heading 'beliefs that the new
evidence undermines', the search might not yield much. Second, if one
searches under the heading 'beliefs that originated from beliefs now aban-
doned', then again the search might yield sparse results. For then one must
be guided by beliefs about the origins of one's beliefs; yet one might not
keep track of the origins very well, or might not activate these beliefs. (Re-
call my discussion to this effect in section 10.3.)There might be any number
of beliefs satisfying this description that one does not identify as such and
therefore does not identify as a belief that needs to be abandoned. In short,
a person could abide by rules of this sort, fail to abandon some of the in-
directly undermined beliefs, yet not be justificationally remiss.
Although not uncongenial, this is a result that may seem counterintuitive
until one considers the feasibilities of available cognitive operations. When
one considers feasibilities, it is an open question whether a right rule system
would oblige a cognizer continually to search for old beliefs in LTM that
might be weeded out in light of new evidence. Since one acquires massive
quantities of new evidence (new perceptual beliefs) at every waking mo-
ment, conformity to such a requirement would be very costly in cognitive
resources; so costly that it could well reduce the overall truth ratio of one's
belief corpus. And it would almost certainly put a crimp in other cognitive
tasks. Hence, it is a dubious normative precept.35
Notice that much more modest belief revision principles are possible
than the ones scouted above. A simple rule might say that if one activates an
old belief in q, and if one (actively) believes that this belief wholly stems
from now abandoned evidence, then one is required to abandon q. But this
is a far cry from the general maxim to search LTM continually for possibly
discredited beliefs. Here is another case, then, in which facts about the
structure of memory, and what it takes to activate materials in memory,
have direct bearing on normative precepts.
10.5. Retaining, Forgetting, and Revising

As emphasized in section 10.3,primary epistemology is interested in mem-
ory power as well as reliability. As was indicated there, memory power is a
multidimensional affair. It can be measured by ease of storage, by duration
of retention, by quantity of retainable information, and by range of accessi-
bility. Without trying to construct a single measure from these dimensions,
1wish to inquire further into the determinants of memory power. I have al-
ready commented on the impact of elaboration. But there is much research
on other influential factors, such as the context of encoding and the spacing
of multiple encodings. As a sample of this work, let us look at the 'spacing
effect'.
When facts are studied on multiple occasions, the temporal spacing of
these occasions is important, as S. A. Madigan found in an experiment on
the free recall of single words.3BForty-eight words were presented at the
rate of 1.5 seconds per word. Some were presented once and others twice.
After study, subjects were asked to recall as many words as they could, and
of course they recalled a higher proportion of those studied twice. Of inter-
est here are the data for the twice-presented words. In Figure 10.2 the data
for the twice-presented condition are plotted as a function of lag, the num-
222 Memory
Lag between repetitions

(numbers of intervening words)
Figure 10.2. Recall probability of twice-presented words.
ber of intervening words between the two presentations. Probability of re-

call increases systematically with lag. It is a very robust and powerful phe-
nomenon. There is a rapid increase over the initial lag, the benefit increases
to lags of forty intervening items, and there is no reason to think this is an
upper limit. The improvement of memory with increase in lag between
study episodes is known as the spacing efect.
In Chapter 1 I distinguished regulative from nonregulative styles of epis-
temic appraisal, the former designed to guide deliberate conduct and the
latter designed to appraise cognitive states and operations from a purely
theoretical standpoint. Most of my discussion has been, and will continue to
be, confined to the nonregulative style. However, results of cognitive psy-
chology are sometimes usable to devise practical, action-guiding rules. The
spacing effect is one such result. One obvious application is that a student
who wishes to prepare for an exam during a semester, and who has time to
study the material twice, should maximize the lag between the two study
periods. Other such pieces of advice can be derived from other experimen-
tal results on encoding variability and spacing; but I will not explore these
any further.37
I wish to turn now to the nature of forgetting. It is natural to suppose that
Memory 223
forgetting simply consists in losing material from memory. Many memory

theorists do indeed postulate that traces in LTM decay, but decay is just a
weakening of memory material, not its wholesale disappearance. A popular
conjecture among cognitive psychologists is that one never really loses in-
formation from LTM; one only loses access to it. Loss of access, it is also
conjectured, is primarily due to interference. To give these theses some
plausibility, let us begin by looking at interference effects.
Interference effects have been extensively studied with paired-associate
materials. One group of subjects learns two lists of paired associates, in
which the first item of a pair occurs in both lists. For example, the first list
might contain cat-43 and house-61, and the second list cat-82 and howe-37.
The control group studies the same first list; but its second list does not
contain the same initial words. In general the first group does not do as well
as the control group on retention tests administered after a considerable
delay. This is atrributed to interference: members of the first group learn
interfering associations to the initial items.
It is difficult to test the conjecture that all forgetting is due to interfer-
ence, not to the total loss of memory traces. W. Penfield's well-known stim-
ulation of the temporal lobes led patients to report memories that they were
unable to report in normal recall-such as events from their ~hildhood.~'
But it is hard to know whether these reports uncovered genuine memories.
More persuasive evidence comes from experiments by T. 0. Nelson, who
studied relearning of old items that had been forgotten.39 His results are
consonant with the suggestion that forgotten memories are nevertheless still
there, only not easily accessible.
If we suppose that this popular conjecture is correct, then there are im-
plications to be traced for certain epistemic norms philosophers often enun-
ciate. It is common for epistemologists and philosophers of science to say
that when an old belief has been undermined by new evidence, it should be
abandoned. It is also common to talk about a 'corpus' of belief as if it were a
blackboard with a list of sentences written upon it. When new evidence in-
firms a sentence, it should be stricken or deleted from the corpus; it should
be erased from the blackboard. However, a long-standing belief is just a
state of LTM. If psychologists are right that there is never any real loss of
material from LTM, all this talk of abandoning, deleting, striking, or erasing
prior beliefs cannot be literally realized.
This idea is explicitly embraced by D. E. Rumelhart, P. Lindsay, and
D. A. Norman:
A basic axiom followed in the development of the system is that, although
new information can always be added to the memory, nothing already pres-
ent can be erased. Excluding physical or physiological damage, we find no
evidence whatsoever to support the notion that the material in LTM can
ever be deliberately forgotten.40
224 Memory
If this is right, what does belief revision, or change in belief, consist in?
Surely there is some phenomenon of belief revision, eSen if it is not literally
the erasure or replacement of an old belief. What it is, I suggest, is the 'su-
perimposition' of new content over, or alongside, old. New associative links
are established that are intended to be stronger than the old, and to have
contrary contents. For example, suppose you believe for a long time (as all
the books used to show) that brontosauruses had very small heads. Later
evidence is unearthed to show that they did not have such small heads after
all. What happens, when you learn this, is that a new associative path is
added, leading from the node for brontosaurus to the node for moderate-
sized head. Perhaps you also add a note, or correction, to the old small-hea-
dedness node, which says: 'No, no, this is wrong'. Later, if asked the size of
brontosauruses' heads, you might initially retrieve 'small head'. But then
you will retrieve the correction as well and inhibit any such verbal re-
sponse. When 'moderate-sized head' is also retrieved, that response is made
verbally. Still later, after the new material is frequently repeated, for exam-
ple, the 'small-head' node may not be activated at all; it might be totally
interfered with by the 'moderate-sized head' node.
All this conforms with the discussion by Rumelhart, Lindsay, and
Norman:
If material cannot be erased from LTM, then it will contain contradictions
and "temporary" structures. The intermediate incorrect steps one follows in
solving a complex problem (one that takes a considerable amount of time
and thought) will be remembered permanently, along with the final, correct
path. The retrieval processes must evolve to deal with these contradictions
and irrele~ancies?~
If the connection between 'brontosaurus' and 'small head' is still there, in
LTM, is it a belief? We better not say so, else we will have to ascribe in-
consistent beliefs to the cognizer, and this is not the sort of situation that
invites ascription of inconsistency. But this is not a serious problem. As long
as the connection between 'brontosaurus' and 'moderate-sized head' is
stronger than the old connection, and as long as the note, or correction,
pathway remains strong, only the latter materials deserve to be called be-
liefs. But don't we need a label for the LTM state that consists in the old,
uneradicated connection between 'brontosaurus' and 'small head'? Al-
though there is no good provision for this sort of state in ordinary language,
I have frequently stressed the inadequacies of ordinary-language taxono-
mies of the mental. I will call this sort of state a credal residue.
Credal residues are important to acknowledge because, like a phoenix,
they can be reborn-as beliefs. Suppose a recently added correction and
newly added content become weaker than the old associative pathway.
That is, upon first hearing the new evidence, you resolve to make the new
Memory 225
content be your real opinion, but unfortunately, your memory does not co-
operate; it does not retain the new content in sufficient strength. 'Bronto-
saurus' cues now activate the old node ('small head') without activating
either the correction ('No, no; this is wrong') or the new node ('moderate-
sized head'). The new belief is forgotten, and the credal residue resumes its
ascendancy. What can we now say about this credal residue? I think the
only reasonable thing to say is that it has reacquired its old status as a belief.
What is the justificational status of a credal residue that reemerges as a
belief? There is some temptation to call it unjustified. After all, its content
has been rejected on a previous occasion, and only the failure to retrieve
this rejection allows the content to reemerge in an ascendant form. This an-
swer would make sense if it were possible to expunge beliefs. A theory of
justification might say that striking such a belief from one's corpus is man-
datory; and if this is not done, then the preserved belief is unwarranted.
However, since we have denied the psychological feasibility of belief era-
sure, it is not clear how we could accommodate the view that the reascen-
dant credal residue is unjustified.
One way of trying to rationalize the label 'unjustified' within our frame-
work is to point out that the target case of belief recreation crucially in-
volves a process of forgetting (forgetting some other material). It might be
urged that the process of forgetting generally inhibits reliability, and there-
fore would not be authorized by a right system of J-rules.
How would it be argued that forgetting inhibits, or reduces, reliability?
There are really two forms, or manifestations, of forgetting. One manifesta-
tion is simply a failure to retrieve a once stored item. This manifestation
does not involve the production of a false A-belief; rather, it merely in-
volves the absence of an A-belief. The second manifestation of forgetting,
though, is erroneous retrieval. This arises from interference. Suppose a sub-
ject has studied a paired-associate list, including the item cat-43. Later he
studies another list that includes cat-82. Learning of the second list inter-
feres with the first. So when he is subsequently asked to recall items on the
first list, he erroneously retrieves cat-82. This process of interference is re-
sponsible for error.
The present case of credal residue is really a form of such interference
(only it features 'proactive' inhibition rather than 'retroactive' inhibition).
Earlier learning of the connection between 'brontosaurus' and 'small head'
interferes with the subsequent learning of the connection between 'bronto-
saurus' and 'moderate-sized head', so that the latter is no longer retrieved.
When it is recognized that such a process of interference works against reli-
ability, we can see that a rule permitting such a process might be a poor
candidate for inclusion in a right system of J-rules. Hence, beliefs formed by
such a process would not be justified.
All this is a way of supporting the claim that a reascendant credal residue
could be counted as an unjustified belief. But there is a problem with this
analysis. Although interference clearly can produce false A-beliefs, it can
also work to promote true beliefs. According to my earlier account, the way
in which the mind revises its old beliefs is by adding new informational
pathways and getting them to dominate the old pathways by means of in-
terference. In many of these cases the beliefs resulting (in part) from intet-
ference will be true. So it is not evident whether interference generally
promotes unreliability, or how seriously it breeds unreliability.
It is thus premature to expect that interference processes generally would
be excluded by right J-rule systems. Hence, it is questionable whether my
theory implies that reascendant credal residues are always unjustified. I do
not think it obvious that this is the intuitively correct judgment anyway.
The whole matter needs further reflection and investigation. At this junc-
ture I leave this as one among many problems for future explorations of pri-
mary epistemics to pursue.
chapter 11
Constraints on Representation
11.1. Epistemology and Representation Formation

In order to believe something, it must first be present in the mind; some
mental representation must be constructed that is a candidate for doxastic
acceptance or nonacceptance. The question is, How does the mind perform
such constructions? Is the construction of representations a random, hap-
hazard affair? Or is it governed by a fixed set of rules and principles? Is there
at least some set of preferred patterns and styles of constructional activity?
Of all philosophers, Kant laid the greatest emphasis on the mind's con-
structional activity. Forms of intuition and categories of understanding
were viewed as constraining what the mind could represent. Recent trends
in cognitive science have been drawing similar conclusions. The sorts of
constraints in question are different from the Kantian ones, and the methods
of studying them form a sharp contrast. Where Kant used transcendental
methods, cognitive science uses theoretico-empirical methods. But there is
a striking similarity of perspective. In this chapter I shall discuss certain
constraints on representation that seem to emerge from work in cognitive
science. Then I shall assess the epistemological ramifications of such
constraints.
There are, of course, many domains of cognition and arguably many
forms of representation. Representational constraints need not be similar,
nor present to the same degree, for different domains or for different forms
of representation. Every perceptual system, for example, might have its
own peculiar constraints; and constraints on nonperceptual, purely propo-
sitional, representations may have little in common with constraints on per-
ceptual representations. There may be specialized constraints on
grammatical representations, with no parallel in other domains. All this is
open to investigation. But in this chapter I shall propose a number of simi-
larities in representational constraints that appear to cut across several
domains.
228 Constraints on Representation
Suppose it were established that the mind's activity in forming represen-

tations is fairly sharply constrained. How might this be evaluated along the
dimensions I have sketched? First, there is a danger that such constraints
would be distorting. This might reduce reliability. Second, there might be
truths that the system is incapable of representing. In Fodor's terminology
the system might be "epistemically bounded."' This apparently would
imply a limitation on the system's power. These, among others, are the sorts
of liabilities that representational constraints might engender. But there are
also significant assets that constraints might possess. All these will be sur-
veyed after I present some styles of constraints that current research
suggests.
11.2. Partitions and Hierarchies

The neorationalist approach to language acquisition, inspired by Noam
Chomsky, is undoubtedly the arena of cognitive science that most inten-
sively explores constraints on representation. Chomsky describes the un-
derlying idea as follows:
It is clear that the language each person acquires is a rich and complex con-
struction hopelessly underdetermined by the fragmentary evidence available
. . . This fact can be explained only on the assumption that these individuals
employ highly restrictive principles that guide the construction of grammar
. . . Powerful constraints must be operative restricting the variety of
languagesS2
Language acquisition ,however, is not the main area that I shall explore, for
several reasons. First, the literature there is vast and technical, and would
take us far afield. Second, precisely because it is so intensively researched, it
may be instructive to choose examples from less familiar domains. Third,
language acquisition may have very specialized properties, which reduce its
usefulness for illustrative purposes. So I shall concentrate my efforts else-
where. Nonetheless, a few remarks on grammatical representation will be
ventured later in the chapter.
Virtually all theories of mental representation assume that the mind
somehow decomposes experience into parts, that it somehow partitions or
segments the world into convenient units. It is also assumed that the mind
engages in 'synthesizing' activities: constructing new units and structures
out of antecedently formed units. Kant, of course, laid great emphasis on
the mind's synthesizing activity, specifically the mind's propensity to struc-
ture experience in terms of material objects. In this section I shall focus on
some selected unitizing and constructional operations that cognitive psy-
chologists are investigating. In particular, I shall look at some hypotheses
Constraints on Representation 229
about .the ways in which human beings partition objects and events into
subunits, hypotheses which (in many cases) attribute a strong preference for
groups of representations that admit of hierarchical organization.
Let me begin with hierarchical structuring. The literature is replete with
descriptions of data structures, or cognitive units, that are organized hierar-
chically. Units are embedded in higher-level units, which are in turn em-
bedded in yet higher-level units.
Many examples of hierarchically structured representations are linked to
natural language.
(I) The phrase structure of sentences is hierarchical. A sentence is the
highest level unit, subordinating noun phrases and verb phrases, which in
turn subordinate nouns, verbs, and other grammatical units. While these
units are categories of word sequences, not of mental units, the universality
of these categories in public codes implicates a hierarchical organization of
associated mental representations.
(2) Another domain in which hierarchical structuring is widely postulated
is semantic memory. The popular semantic memory 'networks' stress a
class-inclusion structure in commonsense concepts.= Class-inclusion arrays
such as <physical object - living thing - animal - mammal - dog - collie>
are ubiquitous in natural languages. These taxonomical systems are proba-
bly manifestations of deep-seated representational mechanisms of the
human cognitive system.4
Hierarchicalization is by no means confined to representations associated
with natural language. It is found in the representation of all sorts of visual
and temporal materials.
(3)In the case of visual cognitive units smaller visual parts are embedded
in larger ones.'
(4) In the case of temporal strings shorter units are subordinated to longer
ones.6
Detailed discussion of these domains of representation will follow shortly.
Many other domains are also the subjects of postulated hierarchicalization.
(5)In story understanding plots are said to be represented hierarchically,
with smaller episodes embedded in larger ones.'
(6) The organization of behavior is achieved by hierarchical networks of
motives and submotives, plans and subplans.'
(7) Problem-solving tasks are structured by means of problems and
subproblems?
Turning now to some details, let us begin with visual materials. It is easy
to portray visual representation-in both perception and imagery-in
terms of multiple layers of embedding. Take the representation of a stand-
ing person. The whole is an elongated, ellipse-shaped object, oriented ver-
tically. At a finer level of resolution, the parts of the body are delineated: a
head, a torso, two arms, and two legs. Each of these parts has global proper-
ties. But each can also be considered a whole with further parts. The head,
for example, has eyes, ears, a mouth, and a nose.
An example from A. Stevens and P. Coupe testifies to the storage of imag-
istic information by hierarchically layered parts.10When asked which is far-
ther west, Reno or San Diego, the typical subject mistakenly chooses San
Diego. The natural explanation is that San Diego and Reno are encoded as
subunits of California and Nevada respectively, and California is pictured
as west of Nevada on one's mental map of the continent.
Several of the experimental studies I shall review are not only concerned
with hierarchical organization; they also postulate preferences for patterns
that accord with Gestalt principles. So before turning to these studies, let
me mention and illustrate these well-known Gestalt principles, originally
postulated for perception. Some of the principles are illustrated in Figure
11.1. In part A we perceive four pairs of lines rather than eight separate
lines. This illustrates the principle of proximity: elements close together
tend to organize into units. Part B illustrates the same principle. Five col-
umns of dots are seen rather than five rows, because the dots in a column are
closer together than the dots in a row. Part C illustrates the principle of
similarity: objects that look alike tend to be grouped together. In C we tend
to see five rows of alternating 0's and X's, even though the rows are spaced
similar to those in B. Part D illustrates the principle of good continuation.
We perceive part D as two lines, one from x to y and the other from w to z,
although there is no reason why this could not represent another pair of
lines, one from x to z and the other from y to w. But the line from x to y
displays better continuation than the line from x to z, which has a sharp
turn. Part E illustrates the principles of closure and good form.We see the
drawing as one circle occluding another circle, although the occluded ob-
ject could have many other possible shapes.
It is time to consider some representative studies of hierarchical organi-
zation and partition of visual materials. G. H. Bower and A. L. Glass studied
subjects' styles of partitioning unfamiliar line drawings." In principle, a
given line drawing could be mentally dissected into all kinds of parts. But in
fact, they proposed, only certain subpatterns are 'naturally7 represented as
parts. Bower and Glass predicted that a natural part should serve as a strong
retrieval cue for the original, whole pattern; whereas an unnatural frag-
ment, even one of equal size, should lead to poorer recall of the original
whole. After their subjects studied a set of original drawings, they were
tested for recall on fragments which the experimenters had categorized as
'good', 'mediocre', or 'bad' (misleading).The experimental results were that
good cues had about five times more retrieval power than bad cues.
Bower and Glass chose the good, bad, and mediocre fragments by appli-
cation of Gestalt rules of common direction and minimal angle. As we have
seen, the rule of common direction, or good continuation, states that line
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
X X X X X
X X X X X
Figure 11.1
segments continuous in the same direction are combined and encoded as a

single structural unit. The rule of minimal angle says that when three or
more segments intersect at a point, the two with the minimal angle be-
tween them are combined to form a single structural unit. Referring to Fig-
ure 11.2 and assuming that part segmentation accords with these rules, the
figure in A is parsed into the parts shown in B, and the figure in C is parsed
Good Cue Mediocre Cue Bad Cue
Figure 11.2
into the parts shown in D. Using this theme, Bower and Glass constructed
picture fragments of C such as E, F, and G. E should be a good cue, F a
mediocre cue, and G a bad cue. Experimental results confirmed this
prediction.
A similar set of results was reported by Stephen palmer.12 Like Bower
and Glass, Palmer suggested that only certain subparts are 'natural', or
'good', decompositions of a larger figure. He devised a measure of goodness
based on Gestalt principles of grouping: proximity, closedness, connected-
ness, continuity, and so forth. Figure 11.3displays some of the whole figures
and respective parts he used, where the parts are rated for degrees of good-
ness: high-(H), medium high (MH), medium (M), medium low (ML), and low
(L). Palmer then conducted a series of experiments to see whether observed
results conformed with the part ratings his scheme predicted. One experi-
.
ment asked subjects to divide figures into their natural parts. A second task
was to rate goodness of identified parts within figures. A third experiment
was a reaction-time experiment. A fourth experiment involved a 'mental
synthesis' task, in which subjects were instructed to construct a figure men-
tally from parts. Palmer found substantial evidence for the kinds of selective
organization in perception and imagery that he had postulated.
Although one might suppose these styles of representation are peculiar to
vision, quite similar styles are found in the representation of sequences of
events. With an eye toward studying hierarchical representation of serial
patterns, Frank Restle devised a formalism congenial to such representa-
tion.13Let E = 112 3 4 5 61 be a set of elementary events, where 'l', '2', and
so on, comprise the alphabet for representing these elementary events. Let
X be a sequence of events from this set; for example, X = <1 2>. A set of
operations on this set is then introduced: repeat (R) of X , mirror image ( M ) of
X, and transposition +1 (T) of X. Each operation generates a new sequence
from a given sequence and then concatenates the old and the new. For ex-
ample, R<1 2> = <1 2 1 2>; relative to the alphabet specified above,
M<2 3 4> = <2 3 4 5 4 3>; and T<2 3> = (2 3 3 4>. In this fashion a
lengthy and complex series can be represented economically by multiple
embedding. For example, T<1> = <1 2>; R<T<l>> = <1 2 1 2>;
T<R<T<l>>> = <12 1 2 2 3 2 3>; and M<T<R<T<l>>>> = <12 1 2
2 3 2 3 6 5 6 5 5 4 5 4>. Using sequences constructed in this way, Restle and
Brown found evidence that subjects mentally encode temporal patterns in
accord with a hierarchical structure indicated by the recursive format.14
A number of writers have recently proposed that musical cognition re-
veals hierarchically organized representations. F. Lerdahl and R. Jacken-
doff, for instance, hold that at least four components of the experienced
listener's musical intuitions are hierarchical.'' 'Grouping structure' ex-
presses a hierarchical segmentation of a piece into motives, phrases, and
Whole
Figures H MH M ML L
Figure 11.3
Figure 1 1.4
sections. 'Metrical structure' expresses the intuition that the events of the
piece are related to a regular alternation of strong and weak beats at a num-
ber of hierarchical levels. And so on.
Diana Deutsch and John Feroe also suggest that people represent pitch
sequences in tonal music in abstract form.'' To illustrate their argument,
consider the pitch sequence shown in Figure 11.4. A musically sophisticated
performer (or listener) would represent this sequence at two or more levels
of description. At the lowest level there is a representation of the entire
presented sequence. At the next (higher) level the representation would
feature an arpeggio that ascends through the C major triad (C-E-G-C). This
more abstract representation registers the fact that the entire sequence can
be viewed as the four notes of this triad each preceded by a neighbor em-
bellishment. At this higher level of representation, the embellishments are
deleted. The resulting hierarchical structure may be presented in tree form
as in Figure 11.5.Drawing on ideas of earlier authors, Deutsch and Feroe
propose that tonal music is represented with the help,of several different
(mental) 'alphabets': major scale alphabets, minor scale alphabets, the chro-
matic alphabet, and the triad alphabets. Different alphabets can be used at
C- E-G- C
B-C D#-E FH-G B-C

Figure 11.5
different levels of representation of the very same passage. Whereas the

first-level representation in the foregoing example occurs in the chromatic
alphabet, the second-level representation could use the major triad alpha-
bet. In the major triad alphabet, C-E-G-C is a sequence of successive ele-
ments. Employment of multiple alphabets facilitates conformity with laws
of figural goodness. (Like other authors I have cited, Deutsch and Feroe
postulate Gestalt-like laws of goodness.) For example, pitch sequences are
more efficiently perceived when their components combine to produce uni-
directional pitch changes and changes proximal in pitch. By treating the
sample passage in the major triad alphabet, it can be represented as a se-
quence of successive-proximal-notes.
A major hypothesis of Deutsch and Feroe is that music sequences are
stored hierarchically, and that a performer generates a learned composition
by accessing it in a top-down fashion. There are close analogies here, which
Deutsch and Feroe note, with Chomsky's system of grammar. In music
these ideas were anticipated by Heinrich ~chenker," who in turn credits
C. P. E. Bach's Essay on the T w Art of Playing Keyboard ~nstruments.l8
In addition to these studies of hierarchicalization in specific domains,
John Anderson postulates hierarchical organization as a pervasive cognitive
phenomenon.'g Anderson points out that even more complex structures can
result when a single element occurs in several different hierarchical struc-
tures, thereby joining them to produce a "tangled hierarchy." Furthermore,
hierarchies can also contain units of multiple representational types: ima-
ginal, temporal, and propositional. Finally, he postulates, people can re-
trieve elements from a hierarchical structure either through a top-down
procesjstarting with the top structure and unpacking successively lower
layers of elements-or in a bottom-up manner-starting with a bottom
node and retrieving higher-level units.
Not all studies of partitioning invoke Gestalt principles or a preference
for hierarchicalization, however. Donald Hoffman and colleagues have pro-
posed principles of shape partitioning independent of either Gestalt ideas
or hierar~hicalization.~' Hoffman's initial proposal is based on the observa-
tion that when two arbitrarily shaped surfaces are made to interpenetrate,
they always meet at a contour of concave discontinuity with their tangent
planes. A straw in a soft drink, for example, forms a circular concave discon-
tinuity where it meets the surface of the drink. A candle in a birthday cake
is another example. Hoffman then proposes, as an initial rule of partition-
ing: divide a surface into parts along all contours of concave discontinuity
with the tangent plane.
This rule does not always define a unique partition. But that fact leads to
a prediction: when no unique partition is defined, it should be perceptually
ambiguous how the object is partitioned. This prediction is confirmed with
reference to the elbow shaped block in Figure 11.6. The only concave dis-
Figure 11.6
continuity is the line in the crook of the elbow. The rule does not define a
unique partition of the block. As predicted, there are three plausible ways
to cut the block into parts, all relying on the contour defined by the parti-
tioning rule.
The foregoing rule does not help with entirely smooth surfaces. Hoffman
therefore develops the following partitioning rule: divide a surface into
parts at loci of negative minima of each principal curvature along its asso-
ciated family of lines of curvature. As applied to the shape in Figure 11.7,
this yields the indicated partitioning contours (dashed lines).
Figure 11.7
Let me close this section with a slightly different topic: not constraints on
the representation of parts, but constraints on the representation of motion.
Shimon Ullman has studied principles that underlie the visual system's in-
terpretation of a sequence of two-dimensional stimuli (for example, dots of
light on a movie screen).21How does a subject 'see' such a succession of
stimuli? One issue is how the subject 'matches' the parts of one frame in a
motion picture with parts of a succeeding frame. Which parts of the subse-
quent frame are viewed as identical to parts of a preceding frame? And
does the subject interpret these event sequences as representing a single
three-dimensional object in motion or in change? If so, what governs the
choice of a three-dimensional interpretation?
Ullman claims that the visual system displays at least three innate prefer-
ences: (1) In interpreting which dot elements match in successive arrays,
there is a preference for seeing motion to nearest neighbors. (2) In deciding
what matching to impute, motion along straight lines is preferred. (Notice
that these first two principles echo Gestalt principles of proximity in space
and common direction.) (3) Any set of elements undergoing two-dimen-
sional transformation that has a unique interpretation as a rigid body mov-
ing in space should be so interpreted. This third principle is a version of the
preference for seeing visual arrays as solid 'bodies', a theme that many phi-
losophers have stressed.22
Earlier work had shown how humans can retrieve structure from motion.
For instance, if a transparent beach ball with tiny light bulbs mounted ran-
domly on its surface is set spinning in a dark room, one immediately per-
ceives the correct spherical layout of the lights. When the spinning stops,
so does the perception of the spherical array. How does one see the cor-
rect three-dimensional array when infinitely many three-dimensional struc-
tures are consistent with the moving two-dimensional retinal projection?
Ullman showed mathematically that if the visual system exploits the
laws of projection, and 'assumes' that the world contains rigid objects,
then in principle a correct interpretation can be obtained. In particular,
three views of four noncoplanar light bulbs are enough to solve the
problem.
However, in cases of 'biological motion' studied by Gunnar Johansson,
one does not have four noncoplanar points.23When small light bulbs are at-
tached to the major joints of a walking person, there are at best only pairs of
rigidly connected points, such as the ankle and the knee, or the knee and the
hip. Nonetheless, the visual system does recognize a walking person from
such information. Hoffman and Bruce Flinchbaugh suggested that if the vis-
ual system makes a further assumption-the 'planarity' assumption--dis-
plays of gait of this sort can be correctly interpreted from three snapshots of
just two points.24Thus, the visual system appears to have very specific con-
straintjprobably innate computational mechanisms-for constructing
shape representations.
11.3. Operations and Analogies

These last examples illustrate specific principles, or operations, that the
mind seems to realize in constructing visual representations. But even the
earlier examples, identifying preferred partitions, seem to reflect specific
operations or transformations for the generation of representations.
The proximity rule might reflect an operation that searches for adjacent
elements within a pattern and that constructs partitions which group these
elements together. Another preference cited by the Gestalt psychologists,
but not previously mentioned here, is the preference for symmetry. This
preference might reflect an operation that looks for, or tends to generate,
inversions or mirror images. I have noted the role of such an operation in
musical patterns, but it also seems to underpin aesthetic responses to sym-
metry in ornamental and even intellectual creations. Again, the preference
for hierarchical orderings may reflect such operations as 'abstraction' and
'globalization', or their inverses, 'specification' and 'localization'. All of
these operations may function in one or more of the following contexts: (a)
perception, (b) storage and retrieval from memory, and (c) imaginative
construction of new representational units out of preexisting units.
For my purposes, it is not crucial to fix the representation-forming opera-
tions precisely. What interests me here is the hypothesis that there is a de-
terminate and fairly narrow set of such operations, roughly of the kind I
have culled from the surveyed literature. Indeed, it is equally worth as-
sessing the implications of a weaker hypothesis, namely, that there is a
fairly wide set of available operations, or transformations, but also a prefer-
ence ordering among them. On this hypothesis the system's control pro-
cesses effect a priority ordering among the transformations, so that
higher-ranked ones get applied earlier and therefore have a greater fre-
quency of application.
Intimations of a narrow set of operations, at least for certain domains, are
contained in the research I have presented. Deutsch and Feroe suggest that
a great deal of observed pitch sequences in traditional music can be recon-
structed in terms of elementary operations such as sameness in pitch, next
in pitch (within a specified scale, or alphabet), predecessor in pitch, trans-
position of a phrase, inversion (mirroring) of a phrase, and so on. Similarly,
of course, Chomsky and other grammarians postulate a fairly narrow set of
operations to account for grammatical structures. Chomsky also suggests
that analogous representational constraints may underlie art forms.
Certain conditions on the choice and arrangement of linguistic expressions
characterize literary genres intelligible to humans, with aesthetic value for
humans; others do not . . . In these and many other domains, a certain range
of possibilities has been explored to create structures of marvelous intricacy,
while others are never considered, or if explored, lead to the production of
work that does not conform to normal human capacities.=
Citing Marshall ~ d e l s o nChomsky

,~~ notes that Freud made similar pro-
posals in his work on dreams. In Edelson's interpretation Freud's psycho-
analysis was a sort of science of semiology, a dream being created by
principles he calls "the dreamwork," which produce the manifest content
of dreams from latent dream thoughts. Ultimately, Chomsky proposes a
"universal grammar" of scientific theorizing, by which he apparently means
a set of hypothesis-constructing operations, which I intend to include in my
discussion. Although the literature I have adduced focuses mainly on opera-
tions concerning visual or auditory patterns, the class of representation-
forming operations I mean to address include operations that take concepts
or hypotheses as inputs and generate new hypotheses as outputs.
Notice that my conjecture is independent of the issue of modularity of
mind. If the mind is modular, it would be natural to expect that distinct
modules have distinct sets of operational, or transformational, repertoires.
But a nonmodular theory is also compatible with a highly constrained set of
operations. Nothing said here is intended to take sides on the modularity
issue.
Distinct repertoires of operations might be associated with distinct re-
presentational codes, as opposed to distinct modules. Anderson, who op-
poses mental modularity, endorses a multiplicity of codes and takes these
codes to have different sets of operations?' Specifically,he postulates three
types of representation: temporal strings, spatial images, and abstract prop-
ositions. For temporal strings, he suggests three probable operations: (a)
combination of objects into linear strings, (b) insertion, and (c) deletion. For
imagery, he mentions synthesis of existing images and rotation. For abstract
propositions, he lists insertion of objects into relational slots and filling in of
missing slots. He clearly thinks that the representational types, or codes, dif-
fer in their transformational repertoire.
The next theme I wish to discuss is analogy, which I shall link with the
theory of natural operations. Analogy is a large topic, which has burgeoned
in recent years. No attempt at comprehensive coverage of the literature is
therefore contemplated. Furthermore, I will lump together the notions of
analogy, similarity, and metaphor, though these can doubtless be
distinguished.
The ubiquity of the similarity construct in psychology has been remarked
upon by Amos Tversky: "The concept of similarity is ubiquitous in psycho-
logical theory. It underlies the accounts of stimulus and response general-
ization in learning, it is employed to explain errors in memory and pattern
recognition, and it is central to the analysis of connotative meanings."2s
Other cognitive scientists have emphasized the centrality of analogy in
(among other areas) the phraseology of everyday language, and the strate-
gies of problem solving.
A recent book by George Lakoff and Mark Johnson traces the pervasive-
ness of analogy--or metaphor-in ordinary language and thought.'' A sam-
ple metaphorical pattern involves orientational metaphors: expressions

rooted in spatial orientations, such as up-down, in-out, front-back, on-off,
deep-shallow, and central-peripheral. Innumerable expressions for nonspa-
tial properties and events are built upon these orientational notions. For
example, happy is up, sad is down ('I'm feeling up', 'That boosted my spir-
its', 'My spirits rose', 'You're in high spirits', 'I'm feeling down', 'He's really
low these days', 'My spirits sank'); health and life are up, sickness and death
are down ('He's at the peak of health', 'He's in top shape', 'He's sinking fast',
'His health is declining'); having control or force is up, being subject to
control or force is down ('I have control over her', 'I am on top of the situa-
tion', 'He's at the height of his power', 'He is under my control', 'He fell
from power', 'He is low man on the totem pole'); good is up, bad is down
('Things are looking up', 'We hit a peak last year, but it's been downhill
ever since', 'Things are at an all-time low', 'He does high-quality work').
According to Lakoff and Johnson, a vast expanse of our conceptual territory
is rooted in analogical extensions of more primitive conceptual contents.
The linguistic patterns, then, may be taken as evidence for the pervasive
employment of some sort of analogy-constructing operation in the cognitive
system.
The use of analogy in problem solving is also widely remarked upon.
G. Polya recommends searching for analogies as a strategy for solving math-
ematical problems.30 Analogical test items of the form A:B as C:D are
widely used in intelligence testing, since a grasp of these sorts of relation-
ships seems to develop quite early.
Experimental studies of problem solving also reveal a strong influence of
similarity judgments in problem-solving strategies. In many problems a
person's task is to find a way of getting from one state in a 'problem space'
to the goal state. Typically, a long series of states must be traversed. How
does the human problem solver choose, in any given state, what state to
move to next? Apparently, people are strongly influenced by similarity.
They choose to move into new states that are more similar to the goal state
than the present state, and more similar to the goal state than are other
possible candidates.
Here is an illustration reported by M. E. Atwood and P. G. ~ o l s o nSub- .~~
jects were given the following water jug problem.
You have three jugs, which we will call A, B, and C. Jug A can hold exactly 8
cups of water, B can hold exactly 5 cups, and C can hold exactly 3 cups. A is
filled to capacity with 8 cups of water. B and C are empty. We want you to
find a way of dividing the contents of A equally between A and B so that
both have 4 cups. You are allowed to pour water from jug to jug.
Atwood and Polson asked which move subjects would prefer, starting from
the initial state. Would they prefer to pour from A into C, thereby yielding
state 2 (below); or would they prefer to pour from A into B, thereby yield-
ing state 3?
State 2: A(5) B(0) C(3)

State 3: A(3) B(5) C(0)
Twice as many subjects preferred the move to state 3 as preferred the move
to state 2. Notice that state 3 is quite similar to the goal, since the goal is to
have 4 cups in both A and B, and state 3 has 3 cups in A and 5 cups in B. By
contrast, state 2 has no cups of water in B. This illustrates a tendency to
move to states that are as similar as possible to the goal state.
Recent work on scientific theory construction, both by cognitive scien-
tists and by historians of science, also explores the role of analogy, or simi-
larity. For example, several authors trace the development of Darwin's
theory to his finding an analogy between Malthus's theory of human popu-
lation growth and the growth of species.32
Analogy is often treated as a separate representation-forming device, and
we could certainly place it alongside the other operations tentatively listed
earlier. In other words, where it is assumed that one representation (be it a
concept, a theory, an idea for a problem solution, or what have you) is ob-
tained from another by some sort of operation, one candidate operation
might be that of analogy. This might be regarded as a primitiue, irreducible
operation. But I want to advance the proposal that the analogy operation is
analyzable in terms of other operations.
One way of trying to flesh out this proposal would be to treat analogical
representation formation as a substitution operation. Take Bohr's model of
the atom, for example. This representation might be derived 'by analogy'
from a representation of the solar system. Wherein does the analogical con-
struction consist? The suggestion would be that it consists in substituting (a)
a nucleus for the sun, (b) electrons for the planets, and (c) nuclear forces for
the gravitational force.
Is this approach plausible? A criterion of adequacy for any such theory
would seem to be this: if representation ROis obtained from R 'by analogy',
then R* (or, rather, its putative referent) should be judged quite similar, or
analogous, to R, at least by the cognizer in question. But not every (mental)
substitution yields a representation of a subjectively similar object. If R is a
representation of a swan, and ROis obtained by first deleting everything in
R except (a representation of) the eyes and then substituting (a representa-
tion of) a tiger, the new representation ROis hardly very similar to, or analo-
gous to, the original.
What further constraints can be placed, then, on substitutions to yield a
match with intuitive judgments of resemblance or similarity?One possibil-
ity is to employ the notion of 'natural parts'. The greater the naturalness of
the parts preserved under replacement, the greater the subjective similarity
of the resulting representation.
I doubt, however, that judged analogies or similarities can be adequately
recovered by the substitution operation, even conjoined with this additional
device. Other high-ranking or 'preferred' operations, in addition to substi-
tution, might have to be invoked. For example, assume that reversal, or in-
version, relative to some natural axis or alphabet is a high-ranking
transformation. Then an object obtained from another object by such a re-
versal would be judged rather similar to it.
With this in mind, we might try to generalize the approach along the fol-
lowing lines:
The contents of representations R and R' are similar-receive a high simi-
. larity rating-if and only if the content of one is obtainable from the content
of the other by application of a high-ranking cognitive transformation (or
combination of transformations).
Lance Rips points out, in a personal communication, that one source of evi-
dence for this account of similarity is that different transformations yield
representations similar to the original one but not themselves similar. For
example, the concept 'cold' might be considered similar to 'hot' (from one
perspective), via an 'opposites' transformation. Also, 'warm' is similar to
'hot', via a 'next-to' transformation. But 'warm' and 'cold' are not subjec-
tively similar.33
Notice the analogy between this idea and one employed by topologists.
In topology two figures are said to be topologically equivalent (homeomor-
phic) just in case one can be obtained from the other by some sort of topo-
logical operation.
If this line of analysis proved promising, it might be exploited by turning
it on its head. (I owe this suggestion, and the example that follows, to
Charles Chastain.) Instead of studying similarity in terms of transforma-
tions, one could study transformations in terms of judged similarities. For
example, given perceived similarities among faces, one might try to identify
high-ranking transformations by seeing which would account for those simi-
larity judgments.
These are some tentative thoughts about analogy. Their relevance stems
from my interest in representation-forming operations. If analogy is related
to such operations in the manner proposed, we gain new avenues for getting
at the repertoire of operations. Furthermore, given the apparent ubiquity of
similarity, or analogy, in cognitive phenomena, identification of this reper-
toire-and the most preferred members of it-is quite important for cogni-
tive science.
11.4. Episternic Assets and Liabilities

It is time to turn to epistemological issues, which will occupy the rest of the
chapter. My epistemological reflections will be directed at the sorts of rep-
resentation-forming operations presented in the preceding sections. To be
more specific, I will inquire into the epistemic implications of having a nar-
rowly constrained set of representation-forming operations of these kinds. I
start, in this section, with a survey of assorted assets and liabilities that
might be associated with such a set. In the next section I will explore the
role of such operations in the evaluative dimension of originality. Finally, in
section 11.6, I will examine the relevance of such operations to epistemic
iustifiation.
Consider the mind's propensity to segment objects and events into parts,
such as spatial and temporal parts, and its possession of a rather narrow set
of partitioning preferences. What assets might be associated with such rep-
resentation-forming properties? Hoffman and Richards discuss a number of
such assets.34
First, an articulation of an object into parts is useful because one never
sees an entire object in one glance. Excepting transparent objects, back
sides are never visible. Moreover, even an object's front side is often par-
tially occluded by things interposed between it and the observer. By seg-
menting an object into parts, and storing associations between the object
and its several parts, one can effect a recognition of the object by the visible
subset of its parts.
Another virtue of the mind's propensity to decompose an object into a hi-
erarchy of parts concerns the nonrigidity of objects. A hand can assume dif-
ferent postures, for example, a clenched fist, an outstretched hand, or a
hand giving a victory sign. How well would the mind manage recognition if
it only had a representational device like a template? Not so well. A tem-
plate of an outstretched hand would correlate poorly with a clenched fist or
a hand giving a victory sign. Templates would have to be proliferated to
handle the many possible configurations of the hand. This would be unpar-
simonious and a waste of memory. By comparison, decomposition of objects
into parts promotes efficient recognition.
These points pertain to the general strategy of decomposition into a hier-
archy of parts. But what about the hypothesized narrowness of part prefer-
ences? To understand the advantage there, imagine the demands on
memory and on the matching process if every time one looked at an object
one saw different parts. A face, for example, which at one instant appeared
to be composed of eyes, ears, a nose, and a mouth, might later metamor-
phose into a potpourri of eye-cheek, nose-chin, and mouth-ear parts--a
gruesome and unprofitable transmutation. No advantage would accrue for
allowing such repartitions. In fact, they would be uniformly deleterious to
the task of recognition.35Hence, the narrowness of constraints on partition-

ing also seems to be useful.
In terms of my list of epistemic virtues, the foregoing advantages fall
under the heading of aids to power (or perhaps power and reliability). The
mechanism of decomposition into a hierarchy of parts helps one answer the
question 'What is that?' asked of a perceived, or partially perceived, object.
The more such questions can correctly be answered, the greater the power
of the cognitive system.
Thus far we have reflected on parts, and hierarchies of parts. But hierar-
chicalization in general must be considered. Not all hierarchies are part
hierarchies. The mind's general affinity for hierarchicalization, however,
has advantages in power. Many intellectual tasks benefit from flexibility in
the 'grain', or 'level of resolution', of analysis. Which details about an object
are relevant and which are irrelevant varies from task to task. If I want to
use a computer for word processing, I shall of course be concerned with its
microcharacteristics. But if I only want to use it as a paperweight, I can re-
strict my attention to its global size and weight. The mechanism of hierar-
chicalization facilitates flexible transitions from level to level; and it creates
the potential for the creation of handy superordinate and subordinate
groupings ad libitum.
Hierarchicalization also contributes to power through its associated
mnemonic deployment. This is dramatically illustrated by the case of a
Carnegie-Mellon student, reported by K. A. Ericsson, W. G. Chase, and
S. allo on.^ After training an hour a day for eighteen months, this student
increased his memory span for digits to 79. That is, he could memorize and
accurately reproduce a random string of 79 digits! According to the authors,
there was still a short-term working memory store with a capacity of 4
items, a memory space that had not been affected by practice. The other 75
items were held in a complex hierarchical mnemonic structure of subunits
in LTM, each of which held 3 or 4 items. The key unit at the bottom level of
the hierarchy was a three- or four-digit string usually associated with the
time for running a race. (The subject was a keen runner with a large number
of categories of races, from half mile to marathon.) For example, the string
'3492' would be stored as '3 minutes 49.2 seconds-nearly a world record
for the mile'. The mnemonic system for digits would not generalize to other
materials, of course, and his span for consonants remained at 6. Clearly, it
was the inventive use of hierarchical organization that made possible his
mnemonic feat for digits.
The mnemonic assets of hierarchicalization are also well displayed in the
execution of extended courses of action. This concerns the timeliness of
mental representation. For effective guidance of action, a plan or represen-
tation of the action needs to be brought into working memory. But working
memory has a severely restricted capacity, so a large, detailed plan would
not 'fit'. To execute such a plan, however, a cognizer can partition the plan
into subplans. If the relevant subplans are activated at suitable junctures,
the entire plan can be executed piece by piece. But this requires that the
larger, superordinate plan be suitably linked in memory to appropriate
subplans. This is achieved by hierarchical representation of plan and sub-
plans, and the retrieval of microunits from macrounits in accordance with
this hierarchical organization.
Musical performance is a vivid case in point. The piece to be performed
is represented in LTM at various levels, encompassing different sized seg-
ments of the work. As the musician proceeds through the performance, rep-
resentations appropriate to each passage are accessed. These are detailed
enough for guidance of behavior but small enough for accurate retention in
working memory. Deutsch and Feroe underline this point as follows:
Several investigators have shown that for serial recall of a string of items,
performance levels are optimal when such a string is grouped by the ob-
server into chunks of three or four items each . . . When segments of tonal
music are notated on the present [hierarchical]system, there emerges a very
high proportion of chunks of three or four items each . . . As pitch sequences
become more elaborate, they are represented as on a larger number of hier-
archical levels, but the basic chunk size does not appear to vary with changes
in sequence complexity. This chunking feature therefore serves to reduce
memory load.37
Thus far I have mentioned only virtues of the partitioning and hierarchi-
calization mechanisms, But one might ask, aren't there any vices? For ex-
ample, doesn't the presence of restricted styles of decomposing wholes into
parts reduce power and flexibility? This issue may be discussed in terms of
Fodor's notion of "epistemic boundedness": the inability of a cognitive sys-
tem even to represent-and a fortiori its inability to believe--certain
truths. If a cognitive system has a fixed set of representation-forming opera-
tions, there must be representations it is incapable of generating. This ap-
pears to imply a limitation on power.
Actually this is not so, at least on my definition of power. On my defini-
tion a system is completely powerful if it can answer (correctly) every gues-
ti072 it can ask. This does not imply that it can answer every question. In
other words, complete power does not entail unboundedness. Therefore,
boundedness does not entail the absence of complete power.
Setting this point aside, how serious a defect is the absence of total
power? Not much. When talking about finite minds, it is not an alarming
defect that omniscience is beyond reach. That's just the nature of the beast.
The more important question is: Is it capable of getting true answers to the
sorts of questions it actually wants to answer, to the sorts of problems it ac-
tually wants to solve? On this point there is no glaring deficiency in the re-
presentational operations under discussion.
Notice, in this connection, that the Gestalt principles might be best in-
terpreted as preferred operations or default procedures, rather than manda-
tory operations. Take the wholes and parts that Palmer used in some of his
experiments (shown in Figure 11.3).The parts are rated for goodness as
either high, medium high, medium, medium low, or low. Subjects gave rat-
ings to the parts that accorded with Palmer's Gestalt-inspired metrics. But
subjects can 'see' each of them as a part of the whole, even the lowest rank-
ing parts. It is just harder to see them as parts. So people are not incapable
of parsing the whole into the inferior segments; those parsings are just less
'natural' or obvious.
Still, isn't it a vice of the system that it has preferences for certain pars-
ings? Doesn't it inevitably betoken a distortion in the system's noetic rela-
tion with the world? Doesn't it mean that the mind can't mirror the world,
and doesn't this necessarily breed error as well as ignorance?
Representational preferences do preclude mirroring, but mirroring is not
required for truth. You do not have to get every truth to get some truths.
Partitioning and synthesizing preferences may indeed imply that some
'correct' parts and wholes do not get constructed. But the preferred parts
and wholes may be perfectly accurate as far as they go.
Sometimes, however, preferences do produce error, not just partial igno-
rance. Gestalt preferences induce the visual system to complete, or fill in,
the stimuli it encounters, thereby achieving 'good continuation', even when
the real figure has discontinuities or gaps. This is indeed distortion. Inter-
pretation of two-dimensional arrays as three-dimensional solid objects can
also lead to error-when the real stimuli are only dots of light.
But how serious are these error possibilities? As long as the organism's
environment is full of solid objects, which generally do instantiate good
continuation, these error prospects will not be systematic. In terms of the
actual environment of human beings, the preferences seem to be highly re-
liable. Indeed, the mathematical proofs by Ullman, Hoffman, and their col-
leagues, mentioned in section 11.2, show that under certain conditions like
rigidity and planarity, the hypothesized visual principles are perfectly ac-
curate. (To be more precise, the probability of error has a measure of O-
which does not imply the impossibility of error.)
Even if highly constrained representation-forming devices can lead to
error, there may be compensating gains in speed. This is worth illustrating
in connection with language learning. In acquiring a first language, the
learner may be thought of as generating grammatical hypotheses on the
basis of linguistic 'texts' (discourses) to which he or she has been exposed.
Language-learning theorists in the Chomskyan tradition conjecture that
human dispositions of hypothesis generation-conceptualized as functions
from texts to hypothesized grammar-are quite restricted. Such disposi-
tions could not lead to correct identification of all possible languages.38
Constraints on Representation 24 7
However, although this implies a limit on power, it may be accompanied by

gains in speed. A learning function that cannot learn every language may
nonetheless be able to learn many languages-including actual natural lan-
guage-much more rapidly than more powerful learning functions. Specif-
ically, it may be able to do what the normal child can do: learn a first
language in three years. This trade-off between speed and power may well
be worthwhile, given our limited lifespan.
Returning now to constraints on partitioning, is there anything useful
about the specific partitioning principles human beings seem to manifest (as
judged by the evidence in section 11.2)?I have no inkling of the benefits or
costs to the organism taken singly, but a definite plus can be assigned to the
sharing of a narrow set of principles by species members, no matter what
the principles. The plus is associated with social coordination. Shared parti-
tions promote successful communication; and successful communication
provides enormous epistemic advantages.
The sort of point I have in mind is made, in its essentials, by Quine, who
stresses the uniformity of people's innate quality spacings.39 We might
carry this over to a uniformity of innate partitioning principles. If the young
child had a very different set of innate quality spacings from its elders, it
might associate different mental units with its elders' words than the elders
do. This would make it hard to achieve mutuality in sense or reference.
Where the parent individuates in terms of rabbits, the child might indivi-
duate by rabbit-stages. Or the discrepancy might be even greater. This
would certainly make successful intercourse hard to achieve. (I here as-
sume, unlike Quine, that it is determinate which style of mental individua-
tion in fact characterizes a person.) But where innate partitioning
principles are shared and relatively small in number, the prospects for suc-
cessful communication are greatly enhanced. Such communication has im-
measurable epistemic benefits: innumerable truths, and truth-acquiring
methods, can be learned from others via language. A noncommunicating
creature would be at a severe epistemic disadvantage. Thus, the mere shar-
ing of highly constrained partitioning operations-whatever they may be
(within limits)-seems to be a great epistemic boon.
1 1.5. Originality
Although it is scarcely touched by philosophical epistemology, 'originality'
is a widespread term of evaluation in intellectual affairs. It is a dimension
commonly invoked in the evaluation of both intellectual product-books,
articles, ideas, and hypotheseeand intellectual producers. Since the
progress of knowledge depends on the production of new knowledge, and
since this characteristically depends on original ideas (not merely new evi-
dence for old ideas), originality is a proper subject of epistemological study.
Of course, originality is also present in the sphere of the arts, which may not
fall under the rubric 'intellectual', and may therefore fall outside epistemol-
ogy. But that needn't perturb us.
What do we mean when we call an idea 'original'? One broad meaning is
that it is an idea nobody else has had before; at any rate, it is an idea that
the thinker did not get from anyone else. But there is another, narrower
sense of the term. Within the class of ideas that are not directly borrowed
from others and are not mere duplicates, some strike us as more inventive,
clever, surprising, imaginative, or creative than others. These we deem
'original'. Other ideas, which are equally nonduplicates, seem pedestrian,
obvious, and trite. The term 'original' is withheld from these. But how do
we distinguish the inventive ideas from the pedestrian, the imaginative
from the hackneyed?
Apparently we have some tacit metric by which we measure originality.
Some ideas seem more 'distant' from socially available ideas, harder to think
up on one's own. Other ideas, even if they are original in the broad sense,
seem easy to think up; they seem 'closer' to what has already been formu-
lated. But what underlies this metric? What makes some ideas seem close to
their predecessors, and easier to think up, and others more distant and diffi-
cult? The answer is related to human representation-forming operations.
Suppose, as I have suggested, that human beings have a fairly delimited
set of representation-forming operations, and a (substantially)fixed ranking
of these operations. In addition, assume that in any scientific or cultural mi-
lieu the stock of ideas antecedently available to cognizers is essentially the
same. Then all new ideas must result from transformations of the same ini-
tial stock of ideas. If this is correct, one should expect certain new ideas to
occur independently to many workers in a given field. Members of the dis-
cipline all have roughly the same stock of initial information and previously
tried ideas; and they share a set of problems they wish to solve. So when the
highest-ranking-or most 'natural'-operations are applied to their initial
stock of ideas, many people will come up with the same (or very similar)
new thoughts. But these will be the most obvious thoughts, the most
straightforward extensions of preexisting ideas.
The sociologist of science Robert Merton has remarked on the prevalence
of multiple discoveries in science: the same discovery being made at
roughly the same time by several independent researcher^.^' In fact, I
would add, the number of historically signijicant multiples surely underes-
timates the prevalence of the phenomenon. Scientists and scholars are con-
stantly coming up with similar ideas; but many of these do not get
published or do not get published by everyone who thinks of them. People's
minds &n in very much the same channels.
How, then, are truly original ideas created? And why do certain individu-
als seem particularly adept at such inventiveness?Four possibilities come to
mind. (1)Perhaps the ranking of operations is not completely fixed, or uni-

versal. Perhaps certain operations, which are unusual or low-ranking for
most people, are higher in rank for others. The latter individuals apply
these 'unusual' operations more readily, and hence come up with ideas that
others find surprising and inventive. (2) Although there may be no interper-
sonal difference in the ranking of operations, perhaps some individuals try a
larger number of operations, turning eventually to unusual operations that
most people do not try at all. This could result in ideas that strike others as
departures from the commonplace. (3) Perhaps the important difference
lies in the combinations, or permutations, of operations that certain people
employ. What distinguishes the creative thinker from the humdrum thinker
may be the transformational combinations, or permutations, he or she
utilizes.
All three of these possibilities are ways in which control mechanisms that
select representational operations may differ interpersonally. A fourth pos
sibility is that one of our prior assumptions is inaccurate, namely, the as-
sumption that people have the same stock of initial ideas on which to
operate. While there is doubtless much overlap in such antecedent re-
sources, there are also differences. Hence, (4) perhaps the distinguishing
character of creative people is that they apply the same operations to un-
usual sets of initial ideas, thereby producing unusual outputs.
Whether this is an exhaustive list of possibilities does not much matter. It
is clearly important to the understanding of creativity to determine which
of these-or other-possibilities are primarily responsible for differences in
creativity. This is a future task for cognitive science, though continuous
with work already undertaken on basic representation-forming operations.
But how is this relevant to epistemology? One dimension of epistemolo-
gical appraisal is the assessment of problem-solving procedures. But prob-
lem-solving procedures, as I stressed in Chapter 6, include the formation of
representations that are candidate solutions to problems. When problems
are difficult and previously unsolved by others, this may mean that unusual
solution candidates need to be constructed. Hence, cognitive procedures
involved in such construction fall under the scope of epistemology.
11.6. Justification
Originality and creativity are nonstandard topics for traditional epistemol-
ogy. Other, more traditional, dimensions of appraisal to which the study of
representation-forming operations is also germane include justification,
which can be illuminated by the psychology of representation formation
slightly differently from the reliability considerations already adduced.
It is widely suggested that a person is justified in believing hypothesis p if
p would be a good explanation of certain observed phenomena and there is
no equally good alternative explanation. But this principle is not quite

right. Even if there is an alternative explanation, we would not deny that a
person's belief in p is justified unless it is reasonable to expect him to think
up the alternative explanation. A person is justified in placing credence in a
good explanatory hypothesis that occurs to him, as long as any alternative
explanation is excessively hard to construct (from his prior belief corpus).
Suppose a child (a ten-year-old, say) has never paid close attention to ob-
jects partly immersed in water. Suppose also that he has never heard of laws
of optical refraction or other such physical laws. Suppose he now sees an
oar half immersed in water and notices that it looks bent. He concludes that
it is bent. It never occurs to him that there might be a law of physics which
implies that a straight object partly immersed in two different media, such
as water and air, should look bent. He does not form a representation of any
such alternative explanation of the oar's apparent bentness. Is his belief in
the oar's being bent justified or unjustified? I think it is justified. Given his
background information, it would be very difficult to think up any alterna-
tive hypothesis to explain the oar's apparent bentness.
Suppose you too see the apparently bent oar, but you are familiar with
laws of optical refraction; yet you too conclude that the oar is bent. Then
we would say that your belief is unwarranted. You ought to be able to con-
struct the alternative explanation, which should inhibit your belief (at least
the 'central-system' belief).
Consider another example. Is Othello justified in believing that Desde-
mona was unfaithful? Apparently, Othello does not construct any alterna-
tive explanation of Iago's insinuations, of Desdemona's loss of her
handkerchief, and so on. It does not occur to Othello that Iago might be a
villainous fellow, that Iago's being passed over for promotion might make
him vindictive and lead him to deceive Othello about Desdemona. Most of
Shakespeare's viewers doubtless find in Othello a singular paucity of imagi-
nation. It is easy to construct alternative hypotheses to the unfaithfulness
hypothesis, but Othello fails to do so. Hence, his belief is unwarranted.
The point I wish to stress in all such cases is that judgments of justified-
ness or unjustifiedness depend on the ease or dificulty of thinking up rival
hypotheses. When a scientist's belief in a certain theory is subsequently
shown wrong by the invention of a new and vastly more powerful theory,
this need not undermine the justifiedness of the earlier belief (that is, the
justifiedness of the first theory being believed at the earlier time). It depends
on how hard it would have been for him to create the rival theory then.
The ease or difficultyof creating alternative hypotheses is a function of
both ones antecedent conceptual and doxastic corpus, and the naturalness,
or likelihood, of a suitable combination of representation-forming opera-
tions producing these alternative hypotheses from this corpus. Ow intuitive
judgments of justifiedness and unjustifiedness, I believe, rest partly on a tacit
grasp of the ease or difficulty of constructing relevant rival hypotheses.

However, if we want to spell out the basis for these intuitive judgments, we
need to identify the repertoire of operations and their standard preference
ordering. That would give us a systematic ground for the indicated judg-
ments. But this is a job for cognitive psychology.
How do these comments on justifiedness relate to reliability? The idea is
this. A belief-forming process that incorporates consideration of competing
explanatory hypotheses is much more reliable than a similar belief-forming
process without concern for competing hypotheses. So only the former can
be expected to be licensed by suitable rules of justifiedness.
It is not clear, however, that such rules would require a cognizer to con-
struct all possible competing hypotheses. At any rate, on the 'resource-
relative' conception of justifiedness, a person's belief is justified just in case
it is produced by a process that is reasonably reliable relative to the re-
sources of the human cognitive system. Justifiedness is judged by the use of
processes that are as reliable as can be expected, given the native endow-
ments of human cognizers. Among the relevant endowments are those for
hypothesis construction. Relative to these resources (and some initial corpus
of concepts and beliefs), some rival hypotheses are easy to construct and
others are difficult. The aforementioned conception of justifiedness is sensi-
tive to this relative ease or difficulty. A detailed identification of the dimen-
sions of ease and difficulty, however, rests with psychology, for only
psychology can illuminate the representation-forming repertoire available
for these tasks.
chapter 12
Internal Codes
12.1. Sense-Linked and Sense-Neutral Codes

It is natural to think of mental representation as occurring in different in-
ternal codes, or code systems, some of which are associated with particular
sense modalities such as vision. In addition to visual perception, there is vis-
ual recollection (for example, visually recalling one's lover's face) and visual
imagination (for instance, picturing a possible move in a chess game). These
cognitive events ostensibly share a representational form: they are all, in
some sense, 'visual'. Other cognitive events similarly form natural families:
a family of auditory representations (recalling a passage from Brahms third
symphony), a family of motor representations (recalling 'by feel' what finger
is used to type the letter f), and a family of sense-neutral representations
(recalling the meaning of untersagen).
A given code might be employed in at least four different phases of cog-
nition: (1)perception, (2) LTM, (3) recall, or activation, from LTM, and (4)
manipulation or transformation, in active memory, of materials generated
from LTM, When a code is sense-linked-for example, visual-then phases
(3) and (4) would ordinarily be called 'imagistic'. One generates from LTM
a visual image of one's lover's face, or transforms imagistically a certain
chess configuration.
The questions I wish to raise in this chapter are the following. Assuming
that there are, indeed, different internal codes or code systems-in particu-
lar, sense-linked code systems-what are the epistemic properties of these
systems? Suppose that, in addition to some sense-linked code systems, there
is a sense-neutral code system (what psychologists often call a 'proposi-
tional' code). Then how do epistemically relevant properties of sense-linked
systems compare with those of the sense-neutral system? How do they com-
pare on such dimensions as reliability, power, speed, and so on?
Since I include perception as one stage of a given code's employment, the
topic of the present chapter in principle includes perception. But I have al-
ready touched on perception in Chapter 9. And since no sense-neutral code
Internal Codes 253
occurs in perception, there is no basis for comparing sense systems with

sense-neutral systems at the perceptual stage. I shall therefore concentrate
on the three other stages at which different codes might be employed: re-
taining material in LTM, recalling material from LTM, and constructing
new ideas, scenarios, and possibilities in imagination.
The history of epistemology bears witness to the distinction I have
sketched between sense-linked and sense-neutral codes or code systems.
Often, of course, different terminology was used in describing these code
systems. The language of 'faculties', for example, was quite common. In
whatever terminology, oppositions were drawn between 'sensibility' and
'understanding', 'intuition' and 'reason', 'percepts' and 'concepts', 'sensa-
tion' and 'judgment'. Admittedly, many of these oppositions are primarily
concerned with perception, rather than the other cognitive stages I have
delineated, Still, it is not too much of a stretch to view these oppositions in
the guise I have just been presenting.
The British empiricists and associationists certainly viewed the later
stages I have mentioned-retention, recall, and imagination-as closely re-
lated to perception. Locke, for example, thought of ideas lodged in memory
as the "same" as ones previously perceived:
This further is to be observed, concerning ideas lodged in the memory, and
upon occasion revived by the mind, that they are not only (as the word re-
due imports) none of them new ones, but also that the mind takes notice of
them, as of a former impression, and renews its acquaintance with them as
with ideas it had known before.'
The notion that images are like weak percepts is found in Hume:
Those perceptions, which enter with most force and violence, we may name
impressbas, and under this name I comprehend all our sensations, passions
and emotions, as they make their first appearance in the soul. By idem I
mean the faint images of.these in thinking and rea~oning.~
Similarly for James Mill:
. . . by shutting my eyes see him no longer, I can still think of him. I have still
a feeling, the consequence of the sensation, which, though I can distinguish
it from the sensation, and treat it as not the sensation, but something differ-
ent from the sensation, is yet more like the sensation, than anything else can
be; so like, that I call it a copy, an image, of the sensation; sometimes, a rep-
resentation, or trace, of the ensa at ion.^
The classical empiricist theory, however, is not a perfect example of the
model of mind I favor. The empiricists apparently rejected sense-neutral
codes altogether. They viewed all thought as derivative from, or decom-
posable into, sense-produced materials. In contemporary terminology the
only internal codes they countenanced were sense-linked codes. If this view
254 Internal Codes
were accepted, there could be no epistemic comparison between sense-

linked and sense-neutral codes, because there would not be any sense-
neutral codes.
Although I part company with classical empiricists in their rejection of
sense-neutral codes, I agree with them in their acceptance of sense-linked
codes. This view, of course, contrasts with that of many philosophers and
psychologists of the twentieth century. Many theorists posit only a single,
sense-neutral code and reject the presence of multiple codes, especially
sense-linked ones. They claim that all cognition takes place in a 'proposi-
tional', or 'symbolic', code. If this claim were correct, there would again be
no basis for attempting to compare the epistemic properties of different
codes, for there would not be different codes4
To get a feel for the dispute between single-coders and multiple-coders,
let us briefly survey some of the sample controversies, first among philoso-
phers and then among psychologists. Although the dominant view of twen-
tieth-century philosophy sharply distinguished sense-experience from
belief, dissenters began to arise in the last twenty years. D. M. Armstrong
and George Pitcher proposed a belief theory of perception, that perceptual
experience is the acquisition of beliefs (or dispositions to believe).' In a simi-
lar vein J. M. Shorter and Daniel Dennett held that imagery is a matter of.
describing rather than picturing.e These philosophers, then, proposed a uni-
fication of the sensuous and the cognitive. But unlike the empiricists, who
sought unity by reducing the cognitive to the sensory, these philosophers
sought unity in the opposite direction: by reducing the sensuous to the cog-
nitive (or propositional).
Parallel developments have occurred in psychology. As we saw in Chap-
ter 9, cognitive psychologists tend to view perceptual and percept-derived
events as 'information'-bearing events. In the early 1970s this was made ex-
plicit in the theory that all cognitive processing is essentially propositional,
a view championed by (among others) John Anderson and Gordon Bower,
Herbert Simon, H. H. Clark and W. G. Chase, and M. Minsky and S. Pa-
pert.7 Part of the impetus for the propositional view was the computer anal-
ogy. Computer feats impressed theorists with the power of digital coding. It
showed what could be done, say, with lists and list processing. Further mo-
mentum came from the neurobiological work on feature detectors cited in
Chapter 9. This research suggested that even at low levels of processing, the
nervous system selects slices of information about environmental objects.
The conclusion drawn by the aforementioned group of researchers was that
the mind processes information with a single code-often called a 'proposi-
tional' code. Even perception and imagery could be handled in terms of
propositional, or digital, encoding. Alternatively, imagery was dismissed as
a mere 'ephiphenomenon'.8
Meanwhile, other psychologists were postulating and unearthing evi-
Internal Codes 255
dence for multiple codes. J. S. Bruner, R. 0 . Oliver and P. M. Greenfield

postulated three developmentally linked forms of thought: enactive (or
motor), iconic, and symbolic.9 In their scheme the baby uses only the enac-
tive format; the young child uses imagery as well; but in the adult abstract,
symbolic thought predominates. Extremely influential research on imagery
was done by Roger Shepard and associates.1° In famous experiments
Shepard and colleagues gave subjects problem-solving tasks involving pairs
of perspective line drawings of three-dimensional objects, as in Figure 12.1.
Subjects were asked whether the objects in each pair were the same or not,
that is, whether they could be made congruent by a physical rotation. For
'same' cases, subjects' reaction times increased according to a remarkably
linear function of the angular difference in portrayed orientation. The
strong suggestion-confirmed by introspective reports-was that subjects
'image' one of the objects rotating in space, and then imaginatively deter-
mine whether the two objects would coincide. This mental rotation seems
to involve a kind of mental object-an image-quite different from a set of
propositions. And it seems to involve a process-rotation--quite different
from one that would be appropriate to propositions.
This sort of research has fueled a lively controversy that continues un-
abated. Does imagery constitute a distinct form, or code, of mental repre-
sentation, over and above a generally acknowledged propositional form of
representation? Does the mind have a pictorial, or 'analog', way of repre-
senting states of affairs, as well as a descriptional, or depictional, mode of
representation? The most persistent critic of imagistic representation,
Zenon Pylyshyn, has offered theoretical and empirical reasons to question
this form of representation, and has (along with many others) produced al-
ternative explanations of experimental findings using a propositional, or
digital, model." Stephen Kosslyn and colleagues have generated intriguing
new findings on imagery and a set of theoretical replies to ~ ~ l ~ s hThus,
~ n . ' ~
the battle lines have been drawn.
Figure 12.1
256 Internal Codes
Since I regard the multiple code theory as extremely natural and attrac-
tive, and since I (therefore) wish to explore the epistemological significance
of multiple codes, I want to put in some good words for the multiple code
hypothesis. In other words, I do not want to merely assume its correctness;
rather, I want to lend support to it. There are different ways of formulating
the central hypothesis, however, and I do not mean to endorse the hypoth-
esis in all--or even the most popular--of its renderings. Much of the debate
has centered on topics I regard as incidental to epistemological purposes. I
shall endorse only a modest variant of the multiple code approach. To ex-
plicate this variant, and set it against the backdrop of the recent contro-
versy, is the task of the next section.
12.2. The Imugery Controversy: Some Ambiguities

There are two main ways of stating the nub of the imagery dispute. In one
formulation the dispute is whether imagery is (in some sense) 'pictorial' or
'descriptional'. On this formulation proponents of imagery claim that it in-
volves something like picturing, while opponents try to reduce the experi-
ence of imagery to propositional or sentencelike representations. Variants
of this formulation are (a) that imagery involves a spatial, or quasi-spatial,
medium of representation; and (b) that it constitutes an analog representa-
tion of what it depicts. These variants are not obviously equivalent, either
to each another or to the original, pictorial characterization. There are also
difficulties in stating just what an 'analog' mode of representation consists
in. Still, despite these ambiguities, this family of formulations contrasts
sharply with the second kind of formulation.
In the second main formulation imagists maintain that imagery is essen-
tially like perception. It employs the same code, medium, or system of rep-
resentation as the corresponding form of perception; at least it shares some
of the resources and machinery of the corresponding sense modality. We
may call this the perception-similitude formulation of the dispute.13
The perception-similitude formulation is significantly weaker than the
pictorialist formulation. In particular, it does not imply that imagery is
either pictorial, spatial, or analog. It just says that imagery is akin to per-
ception-leaving it open what perception is like.
Can descriptionalists accept the perception-similitude thesis? Can they
accept the claim that imagery shares the same code, medium, or system of
representation as its corresponding perceptual faculty? It might seem, at
first, that they can. After all, descriptionalists maintain that even percep-
tion is descriptional. If imagery is like perception, it too is descriptional. So
why need they dispute the perception-similitude thesis?
They do need to oppose this thesis if they want to retain their insistence
on a single code. According to the perception-similitude thesis, visual imag-
ery shares the code or resources of visual perception. But ordinary 'abstract'
Internal Codes 257
(sense-neutral) thinking presumably does not share that code or those re-
sources. Nor do auditory images. Hence, if the perception-similitude thesis
is correct, there are distinct codes or code-systems. Thus the descriptional-
ist, to the extent that he wants to cleave to a single code, should not be re-
ceptive to the perception-similitude thesis.
For my purposes, I do not care to endorse or defend the pictorial formu-
lation of the imagist position, although this has received the lion's share of
attention. It suffices for my purposes to endorse and defend the perception-
similitude thesis-or, for that matter, any other thesis that guarantees mul-
tiple, infmmationally distinct code systems. However, let us see just what
imagists have been claiming. We will find considerable ambiguity in their
writings, but no matter. It is open to us to select whichever aspects of their
position seem most secure, and most significant from an epistemological
point of view.
The most sustained and detailed defender of the imagist position is Ste-
phen Kosslyn. Many of his writings contain phrases and models that
strongly insinuate an acceptance of the pictorialist thesis. He has intro-
duced the idea that visual images are like displays on a cathode ray tube
(CRT), and he and his colleagues characterize images as "spatial displays"
in active memory.14 In his book I m g e and Mind Kosslyn talks explicitly
about images as occurring in an "internal spatial medium,"15 and he calls
his position "quasi-pictorialist."ls
Yet Kosslyn and colleagues also say that the CRT model is just that-a
model-and they do not mean that images are exactly like a display on a
CRT.In Image and Mind Kosslyn says that picture talk in connection with
imagery is just a metaphor:
Images are not physical objects. Unlike a picture, an image cannot be

dropped, carried under one's arm, and so on . . . If images pass through in-
termediate orientations when they seem to "rotate," for example, they do
not do so for the same reasons that a rotating physical object passes through
intermediate states."
No researcher in the field would seriously argue that images are pictures;
pictures are concrete objects that exist in the world, while images are ethe-
real entities that occur in the mind.''
What researchers usually mean when they talk of having in one's
head is that one has retrieved, or generated from memory, representations
like those that underlie the +ence of seeing . . . Image representations are
like those that underlie the actual experience of seeing something, but in the
case of mental imagery these representations are retrieved or formed from
memory, not from immediate sensory stim~lation.'~
Whereas the earlier phrases I quoted strongly suggest the pictorialist the-
sis, these passages--especially the italicized clauses-support only the per-
258 Internal Codes
ception-similitude thesis. The latter interpretation is bolstered by passages

in which Kosslyn indicates what he means by 'quasi-pictorial':
We can think of the mind's eye as a processor that interprets quasi-pictorial
representations (that is, those underlying visual perceptual experiences) in
terms of "conceptual" categories.*
Unfortunately, it is by no means clear that the perception-similitude inter-
pretation is Kosslyn's final, considered intent. In a very recent paper, "The
Medium and the Message in Mental Imagew." he certainly seems to en-
dorse the analog and spatiality themes:
The distinction between a representation and a medium has proven impor-
tant in the study of visual mental imagery. Although no serious researcher
today maintains that images are actual pictures in the head, some still find it
reasonable to posit quasi-pictorial representations that are supported by a
medium that mimics a coordinate space. On this view, images are not lan-
guagelike "symbolic" representations but bear a nonarbitrary correspon-
dence to the thing represented. Partly because of the primitive origins of this
idea, many people seem wary of it. But the idea that images are a special
kind of representation that depicts information and occurs in a spatial me-
dium is not patently ridiculous, and in fact can be developed in a very coher-
ent way that violates neither philosophical nor empirical considerati~ns.~'
Ambiguities in the imagists' position are found in Shepard's work as well
as Kosslyn's. Shepard's talk of "rotating" mental images suggests a spatial
posit. Shepard and Chipman's endorsement of "second-order isomorphism"
may also appear to confirm a spatial analog construal of their position. The
term 'analog' is explicitly used by Shepard. But more recent passages invite
only a perception-similitude interpretation. In a 1978 article Shepard
writes:
Most basically, what I am arguing for here is the notion that the internal pro-
cess that represents the transformation of an external object, just as much as
the internal process that represents the object itself, is in large part the same
whether the transformation, or the object, is merely imagined or actually
perceived.w
Similarly, Shepard and Cooper write:
The isomorphism that we do claim is a more abstract, "second-order" one
. . . namely, that the brain is passing through an ordered series of states that
(whatever their neurophysiological nature) have much in commun with the
perceptual states that would occur if the appropriate physical object were
presented in successively more rotated orientations in the external
Turning now to Pylyshyn, it is clear that his favorite targets in the imagist
position are the analog and spatiality themes, or the general idea that there
is some nonsymbolic, nonpropositional medium or code. As far as the per-
ception-similitude thesis goes, his stance is more ambiguous. In his most re-
Internal Codes 259
cent statement, "The Imagery Debate: Analog Media versus Tacit Knowl-
edge," Pylyshyn appears to accept a large chunk of the perception-simili-
tude view:
It is my view that there is only one empirical hypothesis responsible for the
predictive success of the whole range of imagistic models . . . When people
inuagine a scene or an event, what goes on in their minds is in many ways
similar to what goes on when they obseroe the corresponding event actually
happening . . . The claim that imagery is (in some ways) like perception has
predictive value because it enables us to predict that, say, it will take longer
to mentally scan longer distances, to report the visual characteristics of
smaller imagined objects, to rotate images through larger angles, . . . and so
However, Pylyshyn indicates that this (the italicized statement) is not all
that Kosslyn and colleagues maintain. They also claim that imagery in-
volves use of a distinctive form of representation, or a distinctive medium or
mechanism. And he is unpersuaded of the alleged characteristics of such a
form of representation.
Since even the chief antagonist of the imagist position seems to accept
the perception-similitude thesis, perhaps I need not offer an extended de-
fense of it. However, let me cite some of the experimental evidence that
supports a close connection between vision and visual imagery, and sup-
ports the contention that they both utilize some of the same specialized
First, there is C. W. Perky's old demonstration that people can confuse

imaging with perceiving.2s Then there is recent work by Shepard, R. A.
Finke, and others, showing that imagery and perception share much of the
same physiological machinery.27Yet another impressive experiment is a re-
finement of the McCollough Effect. In the standard McCollough Effect test
subjects first see patterns of black and red vertical stripes, and black and
green horizontal stripes, for about 10 minutes. They are then presented
with test patterns of black and white horizontal and vertical stripes. Sub-
jects report seeing green on the vertical white stripes and red on the hori-
zontal white stripes. In a variation on the McCollough Effect experiment
Finke and J. Schmidt found that similar results can be obtained when sub-
jects merely imagine black vertical stripes on the red patch and black hori-
zontal stripes on the green patch!2BAnother striking effect of a similar sort
is this. It is known that visual acuity for vertical stripes is better than for
oblique stripes. If subjects are asked to say when the stripes blur as the ex-
perimenter moves them farther away, they report being able to distinguish
vertical stripes from farther away than oblique stripes. N. Pennington and
Kosslyn found-amazingly-that this effect occurs with imaginary stripes
as well!* This is rather dramatic evidence that imagery is subserved by the
same code as visual perception.
All this evidence strongly supports the contention that some distinctive
260 Internal Codes
code system is shared by visual perception and visual imagery, a code not
deployed by nonsensory thought or by perception and imagery in other mo-
dalities. Of course, at some deep level, all internal representations, or code
systems, have a common basis, namely, electrochemical activity of neurons.
But there can be distinct code systems at a higher level, even if there is a
common bottom-level realization. (The same piece of computer hardware
can run distinct programs and implement many programming languages.)
Moreover, for the purposes of (primary)epistemology, it is sufficiently note-
worthy that, at some level, there are code systems with informationally dis-
tinctive properties. That is the idea I shall be elaborating in the rest of the
chapter.
The next questions we need to ask are: What are the informationally dis-
tinctive properties of sense-linked code systems, and what are the episte-
mological ramifications of these properties? How does the visual code
system, for example, differ from the nonsensory code system? Here the
theories developed by Kosslyn and others should prove helpful.
12.3. Code Systems and Cognitive Tasks

Before turning to sensory and nonsensory code systems, let us consider some
theoretical aspects of code systems in general. Notice that I use the phrase
'code system', not just 'code'. By 'code system', I mean to designate not sim-
ply a symbol system, possessing some formal and semantic properties, but
also a set of resources associated with the employment of a symbol system,
for example, the medium in which it occurs, the operations or processes
that can be applied to it, and so on. Examples of such resources will be pro-
vided shortly.
It is standardly assumed that a code has both a syntax and a semantics.
Although both of these notions become especially problematic when ap-
plied to cognitive codes, let us start with some possible semantic differences
between different codes. The question is not how codes (cognitive or other-
wise) get their semantic properties, but what k i d of semantic properties
they can have. Stephen Palmer proposes three ways in which codes, or re-
presentational schemes, can differ semantically.30
First, two codes can differ in terms of the type of i n f m t i o n they can
represent, specifically, the sorts of 'dimensions' they can represent. Dimen-
sions might be height, color, distance, pitch, or handedness. One code might
be capable of representing a certain group of dimensions while another
code can represent only a proper subset of those dimensions or a completely
disjoint -groupof dimensions.
Second, even if two codes are capable of representing the very same set
of dimensions, they can differ in terms of resolution. In one code, for exam-
ple, only two possible values of length might be representable: short and
long. In another code, the length dimension may be representable in a hun-
dred, or infinitely many, different values.
Third, codes can differ in terms of uniqueness. A code might provide
order information, making it ordinal. Or it may only offer information about
same-different relations, making it merely nominal. Other kinds of unique-
ness determinations may be borrowed from measurement theory.
Semantic properties of codes are obviously of epistemic interest, but they
are not the only properties important to epistemics. Other important prop-
erties connected with codes, however, are really properties of code systems
rather than codes themselves. For example, how much information and
what kinds of information can a code represent at a given time? What are
the limits, if any, on the information that can be represented conjointly or
concurrently? Constraints of this sort are imposed not by the syntax of the
code itself, but by the code's syntax plus the medium in which the code is
contained. Suppose the medium is restricted to a single page of print. Then
although the code is capable, in principle, of representing, say, a million dif-
ferent pieces of information, perhaps only thirty of these are conjointly rep-
resentable in the medium.
There are other important features of code systems of epistemic interest.
For instance, how easily do representations in a specified code get pre-
served, or retained, and how easily do they get lost? Again, how easily is one
representation mistaken for another or confused with it? Answers to these
questions might differ from code system to code system.
It should be obvious that the theory of representation must concern itself
with code systems and not just codes. This is clearly recognized by cognitive
scientists, although the phrase 'code system' is my own. (Others might pre-
fer the term 'computational system', but this has fewer of the right conno-
tations, I believe, for my purposes.) One attempt to develop a theory of
code systems is found in Kosslyn. Since he uses it to discuss visual imagery,
presentation of that theory should be i n s t r ~ c t i v e . ~ ~
In characterizing a code system, Kosslyn distinguishes two kinds of .
'structures'--data structures and media-and two kinds of 'processes'-
compt.ism and transformations. Data structures are the information-bear-
ing representations in any processing system. They can be specified by ref-
erence to three properties: their format, content, and organization. The
format is determined by (a)the nature of the 'marks' used in the representa-
tion (such as ink, magnetic fluxes, or sound waves) and (b) the way these
marks are interpreted. The format specifies whether a representation is
composed of primitive elements and relations and, if so, specifies their na-
ture. The content is the information stored in a given data structure. Any
given content can be represented using any number of formats. The organi-
zation is the way elementary representations can be combined. The format
of a representation constrains the possible organizations but does not deter-
262 Internal Codes
mine them. For example, propositional representations can be ordered into

various kinds of lists and networks.
A medium does not carry information in its own right; rather it is a
structure that supports data structures. A page, a television screen, and
even the air are media-supporting ink, glowing phosphor, and sound pat-
terns, respectively. Media can be specified by reference to their formatting
and accessibility. The formatting places restrictions on what sorts of data
structures can be supported by a medium. A short-term store, for example,
might have five 'slots' that take 'verbal chunks'-but not visual images. Ac-
cessibility characteristics dictate how processes can access data structures
within a medium. The slots of a short-term store, for instance, might be ac-
cessible only in a given sequence.
Kosslyn indicates that all properties of the media and the data structures
are defined in the context of a particular processing system. Even though
structures have an independent existence, and their nature imposes con-
straints on the kinds of processes that can be used, structures attain their
functional properties only vis-A-vis the operation of particular processes.
Among the classes of processes, it will be recalled, are comparison and
transformation processes. Comparison processes compare two data struc-
tures, or parts thereof, and return a match or mismatch decision, or a mea-
sure of their degree of similarity (defined over a specific metric).
Transformationprocesses can either alter a given data structure by chang-
ing its contents (for instance, by adding or deleting items on a list), by reor-
ganizing the structure, or by leaving the structure intact but using it as an
impetus to replace or supplement it with a new data structure.
There are many questions one might ask about this general theory, espe-
cially concerning its treatment of interpretation or content. But let us pro-
ceed directly to Kosslyn's theory of visual imagery, a particular code
system. Images, says Kosslyn, have two major components. The surface rep-
resentation is a quasi-pictorial representation that occurs in a spatial me-
dium: this representation 'depicts' an object or scene and underlies the
experience of imagery. The deep representation is the information in LTM
that is used to generate a surface representation.
The properties of the surface image are in part a consequence of the
properties of the medium in which it occurs, which Kosslyn calls the visual
bufe7. The visual buffer functions as if it were a coordinate space. This
'space' is not an actual physical one but a functional space defined by the
way processes access the structure. The processes that operate on an image
access the medium in such a way that local regions are separated from each
other by different numbers of locations. Information is represented by se-
lectively activating local regions of the space. The visual buffer has a lim-
ited extent and specific shape, and hence can support only representations
depicting a limited visual arc. (This makes sense if this medium is also used
in perceptual processing.) The visual buffer has a grain, resulting in a lim-
Internal Codes 263
ited resolution. Thus,portions of subjectively smaller images are more diffi-

cult to classify. Representations within the visual buffer are transient and
begin to decay as soon as they are activated. This property results in the
medium's having a capacity defined by the speed with which parts can be
generated and the speed with which they fade; if too many parts are
imaged, the ones activated initially will no longer be available by the time
the later ones have been imaged.
Concerning the format of the data structure, Kosslyn says this. The pri-
mary characteristic of representations in this format is that every portion of
the representation must correspond to a portion of the object such that the
relative interportion distances on the object are preserved by the distances
among the corresponding portions of the representation. One implication of
this characteristic is that size, shape, orientation, and location information
are not independent in this format-in order to depict one, values on the
other dimensions must be specified. (When a term like 'distance' is used to
refer to surface images, it refers to functional relations among regions in the
visual buffer. Increased distance will be represented by increased numbers
of locations in the visual buffer.)
Kosslyn postulates two types of representations in LTM that can be used
to generate images: literal-and propositional representations. Literal infor-
mation consists of encodings of how something looked; an image can be
generated merely by activating an underlying literal encoding. Proposi-
tional information describes an object, scene, or aspect thereof and can also
be used to generate a surface image. The literal LTM medium does not
function as if it were a coordinate space. Rather, it stores information in
nonspatial units analogous to the files stored on a computer. In Kosslyn's
computer simulation model files store lists of coordinates specifying where
points should be placed in the surface matrix to depict the represented
objects.
Turning to the processes postulated by the theory, Kosslyn lists four
classes of imagery processing: (1) those involved in image generation; (2) .
those involved in image inspection; (3)those involved in image t r a n s f m -
twn; and (4) those that determine when imagery will be used spontaneously
in retrieving information from LTM. For example, images are often trans-
formed incrementally, passing through intermediate points along a trajec-
tory as the orientation, size, or location is altered. This is called a shift
transformation.
This completes my summary of Kosslyn's theory. As indicated in section
12.2, its most controversial ingredient is the spatial medium, or visual
buffer. But for my epistemological purposes, this element is not of para-
mount importance. What is of interest to me are answers to questions like
the following: What can the visual code system contribute to cognitive
problem solving? What can it do that another (for example, nonsensory)
code system cannot do? How reliable is the visual code system; or how reli-
264 Internal Codes
able are specific operations, or combinations of operations, within this sys-

tem? Since Kosslyn himself and other participants in the controversy have
been so concerned with the ontological issues (especially the nature of the
medium), there has not been sufficiently systematic treatment of these epis-
temological issues, though they have not been entirely neglected. I wish to
urge a reversal of the emphasis. Let me first survey some of the epistemolo-
gically relevant materials provided by writers on imagery, and then identify
questions for future research.
There are two kinds of ways in which visually coded materials can help in
cognitive tasks. First, there is storage of visual materials in LTM and the re-
trieval of them (in imagistic form). Second, there is deployment of various
operations on representations in the visual buffer. Of course, many cogni-
tive tasks deploy both of these sorts of resources.
Concerning storage and retrieval from LTM, consider the following task
described by Kosslyn. Suppose you are asked how many windows there
were in the living room of your previous residence. If you are like most
subjects, you probably have this information stored only in visual form. You
need to generate from LTM an image of your former living room, and then
count the number of windows. Few people have a purely propositional
memory of the windows in their former living room. Only visually stored
information enables them to answer the question. So there is heavy reliance
on LTM's use of the visual code. But part of the task requires scanning the
generated image while one counts the windows. This includes an operation
on representations in the visual buffer.
Visual materials seem to be well retained in LTM, which accounts for the
use of imagery as a mnemonic device. This is most vividly illustrated by the
historically popular 'method of loci', the origin and nature of which are re-
lated by Kosslyn:
The Ancient Greek Simonides, the story goes, was performing his role as
bard and orator at a banquet when he was called outside to receive a mes-
sage. At just that time, the ceiling collapsed, mangling all the dinner guests
beyond recognition. When asked to recount who the hapless victims were,
Simonides discovered that he could mentally picture the table and "walk
around it," seeing all those seated with his mind's eye. Simonides reportedly
found it quite easy to name all those in attendance by employing this tech-
nique, and went on to develop it as a general way of improving one's mem-
ory. The product of Simonides' and his methodological descendants' efforts is
usually called the "method of loci." The method of loci involves first select-
ing a series of familiar places, among which one can imagine walking in se-
quence Then, when one is trying to memorize a list, one imagines walking
from place to place and "leaving" an image of each item in the list at each
successive location. When later recalling the list, one then walks by the loci,
and "sees" what is present at each one.32
Internal Codes 265
Let us turn now to uses of imagery in the visual buffer alone, without (sig-
nificant) reliance on LTM. In this category are the "mental rotation" tasks
explored by Shepard and colleagues. Given the task of deciding whether
certain line drawings depict congruent three-dimensional objects (see Fig-
ure 12.1), virtually all subjects imagine the rotation of one of the depicted
objects, compare the two imagined objects, and then offer an answer. Sig-
nificantly, subjects' performances on this task are extremely accurate. In
Shepard's initial experiments nearly 97 percent of the 12,800 recorded re-
sponses of the eight sabjects were correct, with error rates for individual
subjects ranging from 0.5 percent to 5.7 percent.33
Shepard also emphasizes the significant record of image-assisted creativ-
ity in the annals of science.% The theory of relativity is said to have had its
inception when the young Albert Einstein performed his epochal Gedanken
experiment of imaging himself traveling along with a wave of light at 186,-
000 miles per second. James Watson's account of how he and Francis Crick
deciphered the double-helical structure of DNA contains still more specific
references to mental performances involving imagery, such as his sudden,
crucial realization that under an appropriate rotation in space an adenine-
thymine pair had the same shape as a guanine-cytosine pair. And the con-
joint inventions of the self-starting, reversible induction motor and the
polyphase system of electrical distribution burst upon the Hungarian-born
inventor Nicola Tesla in imagistic form. Tesla had an almost hallucinatory
vision of the rotating magnetic field that would be induced by a circle of
electromagnets, each energized by the same alternating current but succes-
sively shifted in phase.
These major creative feats are not necessary to establish the significance
of image-assisted thinking in science. It is the common practice of average
scientists both to draw diagrams and to think in terms of imaged diagrams.
Most scientists would be greatly handicapped by any stricture against dia-
grammatic imagery in their scientific thought processes. Outside science,
examples of image-assisted tasks are also plentiful as blackberries. One such .
example is chess. Chess players typically find it convenient to picture the
chessboard configuration that would be generated by a move sequence they
are contemplating. This helps them determine favorable or unfavorable
prospects of such a sequence.
In all of these cases, it is quite clear that imaging helps. It enables one to
get information one could not get otherwise, either at all, or with equal fa-
cility. Not all uses of imagery, though, are helpful or accurate. A familiar
case is the untrustworthiness of imagery as a test of possibility. The fact that
one cannot imagine, that is, picture, a chiliagon does not demonstrate the
impossibility of a chiliagon.
Other cases of unhelpful, or merely pseudohelpful, imaging are presented
by ~ ~ l ~ s hImagine
y n . ~holding in your two hands, and then simultaneously
266 Internal Codes
dropping, a large and a small object, or two identically shaped objects of

different weights. Which object in your image hits the ground first? Imagine
a transparent yellow filter and a transparent blue filter side by side. Now
imagine slowly superimposing the two filters. What color do you see in your
image through the superimposed filters? Imagine a glass half full of sugar
and another nearly full of water. Imagine the sugar being poured into the
glass with water in it. Examine your image to see the extent to which the
resulting height of water rises (if at all).
As these cases are designed to illustrate, the answers you give from your
imaging will only be as good as your knowledge of physical laws. Clearly,
what you imagine in these cases will be guided by your beliefs about physi-
cal principles: principles of gravity, of color mixture, and of chemical solu-
tions involving sugar and water. Pylyshyn stresses that, in these cases, it is
this knowledge-presumably propositional, or 'symbolic' knowledge-that
would guide your imagination. Imagery does not naturally obey physical
laws, just in virtue of its intrinsic medium or mechanism. If you get correct
answers to the foregoing questions, it is only because you have (possibly
tacit) knowledge of the relevant physical laws. Pylyshyn contends that all
answers or solutions provided by imagination are of this kind. Imagery can
simulate perception because we have had a lot of perceptual encounters
with the world, and therefore know how things look when subjected to cer-
tain changes or transformations. If this is right, imagery per se contributes
little or nothing to our cognitive success. Its apparent contribution is really
attributable to tacit, symbolic knowledge.
Pylyshyn uses such considerations to argue against any analog mechanism
as a piece of "functional architecture," on the grounds that functional ar-
chitecture must be cognitively impenetrable, that is, nonalterable by propo-
sitional content states.% The cognitive impenetrability criterion is
problematic and cannot be explored here in detail. But it is not clear to me
why, even if imagistic operations need to be selected or even guided by
propositional beliefs, this would undermine their epistemic significance.
There may still be kinds of tasks that cannot be accomplished without imag-
istic processes, even if these need guidance by propositional knowledge.
However, Pylyshyn's challenge instructively dramatizes the questions:
What, precisely, are the contributions that imagery can make to problem
solving or question answering? Do these contributions (if any) accrue from
intrinsic features of imagery-related mechanisms, as opposed to background
knowledge? If so, which intrinsic features? What generalizations can be
made about the kinds of problems or tasks for which imagery has epistemic
utility? ,
.
Starting with visual coding of information in LTM, one possibility is that

LTM is better with visually coded information than with propositionally
coded information. LTM might, for example, be able to hold visually coded
Internal Codes 267
information longer; or LTM might be able to preserve it more accurately. .
In this connection we should note a study by Ralph Haber in which subjects

were exposed to 2,560 slides of randomly chosen natural scenes, each slide
being exposed for 10 seconds.=' Recognition performance approached 90
percent one hour after the original exposure. Haber remarks that the results
suggest that recognition of pictures is "essentially perfect." Furthermore,
work by Mary Potter indicates that 10 seconds of exposure is more than
subjects need. Potter reports that subjects' performance in the Haber para-
digm asymptotes at an exposure interval of about 2 seconds per slide.38
Even if LTM were not better with visually coded material than sense-
neutral material, visually coded material could still be an epistemic asset.
As long as material happens to have only a visual encoding, utilization of
that material can enhance total cognitive performance. (This assumes, at
any rate, that retention of such material is sufficiently reliable.) But if recall
of visual material is better than that of nonsensory materials, cognizers
should prefer visual encoding and actively promote it. The historical suc-
cess of the method of loci suggests that this may well be true (at least for
some individuals), though the matter needs to be addressed in a systematic,
experimental fashion.
Results of such a prospective investigation might bear on another sort of
epistemological application. Suppose there is a conflict between two of
your apparent memories, one of them a propositional recollection and one
of them an image. Which ostensible recollection should be given greater
credence? This has direct bearing on the selection of a right system of
J-rules.
Let us turn next to the deployment of imagery in STM (or the visual
buffer). What are the scope and limits of its fruitful uses? Why does it seem
to be so useful, say, for the chess player? And how would this extend to
other tasks? One possibility is that the qwlntity of information graspable in
imagistic form exceeds what can be held in STM in nonimagistic form. Per-
haps imagery's way of representing information about size, shape, orienta- .
tion, and location interdependently gives it special advantages on this score.
And this possibility raises new questions: What differences are there in the
way visual coding 'chunks' or 'unitizes' information about space? Is this
mode of chunking what confers special advantages? How does the rate of
fading of imagistic information compare with that of other information in
short-term, or active, memory? Is imagistic information easier to retain than
other kinds of codes? (Is this what makes people think of imagery as espe-
cially vivid?)
Another tempting explanation of imagery's utility is the cognizer's ability
to project changes in certain parts of a contemplated scene or scenario, and
then 'read off the results of these changes for the system in question. Just
how accurate this kind of operation is and what its accuracy depends upon
268 Internal Codes
needs to be determined, as does the extent, if any, to which it depends on

the sort of tacit knowledge Pylyshyn postulates.
In all of these areas we should bear in mind the three dimensions of epis-
temic value I have stressed throughout: reliability, power, and speed. In
each area we should ask: How reliable is the use of visual coding? How
powerful is it? In particular, can it help generate correct answers to ques-
tions that could not be generated without its deployment? Finally, how can
visual coding increase the speed of problem solving? Similar questions
should of course be asked about other sense-linked codes: the acoustic code,
the motor code, and the like.
12.4. M a t h t i c s and Sensuous Imagination

Most of the foregoing discussion has focused on the virtues of visual coding.
Imagery seems highly useful for many cognitive tasks, even if the basis of
this utility needs further investigation. Historically, though, imagery has
had a checkered reputation. It has often been viewed with suspicion or
downright hostility. I cannot leave the topic of imagery without appraising
the merits of this opposition.
One source of concern has centered on mathematics. Perceptual and im-
agistic thinking was once assigned a pivotal position in mathematics, espe-
cially geometry. Kant, for example, viewed geometry as knowable by means
of quasi-perceptual apprehension, or intuition. But the last century has
largely rejected Kant's philosophy of mathematics, and with it the reliance
on intuition, or sensuous imagination. This rejection was hastened by the
growth of formal approaches to mathematics, approaches that were partly
sparked by discovery of the non-Euclidean geometries. The alternatives to
Euclidean geometry persuaded people that the senses and sensory imagina-
tion cannot certify controverisal axioms or postulates. Development of for-
mal accounts of proof also persuaded people that knowledge of
mathematics rests on systems of language, not perceptual or quasi-percep-
tual intuitions. The tools of mathematical thought should therefore be con-
fined to logic and abstract reasoning, not imagery or sensuous intuition.
A second source of worry is the apparent seductiveness of sense-linked
thinking. People of science may bemoan the fact that the mind is often un-
moved by discursive argumentation or statistical facts, but easily swayed by
dramatic, sensorily vivid materials, even when the latter are of questionable
probity. This danger is scouted by the 'epistemo-psychologists' Richard
Nisbett and Lee Ross, who berate our cognitive system's susceptibility to
"concrete," "vivid" data and its relative weakness at processing "pa!lid," or
"abstract," material^.^' This leads them to a pessimistic view of our native
cognitive system, at least that portion of the system which responds to
vivid, sensory stimuli. This might be taken as a warning, or even a proscrip-
tion against the use of sense-coded information.
Internal Codes 269
These cautionary statements must be assessed by primary epistemology.

Just what is the proper use of sense-linked materials? In this section I ex-
plore this question in connection with mathematics, and in section 12.5 1
examine the contentions of Nisbett and Ross.
Can sense-linked thinking play any role in the cognition of mathematics?
Doubtless there have been sound reasons for challenging the role once as-
signed to sensuous intuition in mathematics. Such intuition turned out to be
deceptive in so many areas of the discipline that claims of 'self-evidentness'
allegedly vouchsafed by sensory intuitions are properly greeted with skep-
ticism. Here is an illustration of the point.
It used to be thought that (sensuous) intuition forced us to acknowledge
that no curve could fail to have a tangent at any of its points. It was be-
lieved that a curve must possess an exact slope, or tangent, if not at every
point, at least at an overwhelming majority of them. But this is not so. Karl
Theodor Weierstrass invented a curve that lacks a precise slope or tangent
at any point. Intuition would tell us that there cannot be such a curve. But,
as elsewhere, intuition fails us.
Examples of this sort properly fuel doubts about the epistemic role of
sensory imagination. These doubts are easily carried too far, however.
There are many alternative roles that imagery, or sense-linked thinking,
might play in the epistemology of mathematics. These different roles must
be considered, lest conclusions be drawn too hastily.
The push toward greater formalization of mathematics in the past cen-
tury seems to exclude imagistic thinking from the discipline, since formal-
ization is naturally associated with 'abstract' thought or with linguistic
forms of representation. Indeed, the general emphasis on proof in mathe-
matics seems to be congenial only to propositional and logical thinking, not
sensory imagination or imagistic representation. Let me therefore pause to
dispel some confusions on this topic.
In confronting this topic we should not forget the distinction between
primary and secondary epistemology. In the context of mathematics sec-
ondary epistemology would be concerned with the study of proper methods
for establishing mathematical truths, including various methods of proof.
The secondary epistemology of mathematics would be concerned, for ex-
ample, with the propriety of nonconstructive versus constructive proof
procedures. But this would still leave many questions for primah episte-
mology about the proper mental processes for arriving at, and understand-
ing, proofs.
Setting aside disputes over the status of axioms and controversies over the
proper notion of 'proof, let us assume that a proof of a mathematical propo-
sition is a necessary component of anyone's being justified in believing it.
Still, such a proof can only be an instrument or method of justification. To
be justified in believing the proposition, a person must also employ certain
mental processes suitably directed at such a proof. The mere existence of
270 Internal Codes
the proof does not suffice for the believer's being justified. He must at least
believe that such a proof exists. But mere belief is not enough either. He
must believe this justifiably. Such justification might come from someone
else's authoritative say-so, in which case the relevant mental processes are
those that infer the existence of the proof from this testimony plus other
background beliefs. Suppose, however, that his justified belief arises from
firsthand study of the proof. Then some mental operations must be em-
ployed that make him justified in believing that this is a proof of the propo-
sition in question.
The typical case is one where the cognizer reads the proof, and this ob-
viously involves perception. So perceptual processes are among those typi-
cally involved in the acquisition of mathematical knowledge. (This needn't
undermine the familiar claim that mathematics is a priori, since this notion
might be defined in terms of the possibility of nonperceptual knowledge.)
What other mental processes would typically be involved? Clearly, it is im-
portant to remember previous steps of the proof, and to remember the in-
ference rules. These memories might involve visually coded representations
of those steps and rules. One would also have to recognize, of each step, that
it is indeed an application of relevant rules to previous theorems. This sort
of recognition could also involve imagistic representations. In short, a suit-
able-that is, justification-conferring-set of mental operations might well
include imagistic thinking (at least for some people on some occasions).
Nothing in the notion of 'following', or 'understanding', a proof excludes
imagistic representation from a role in such a process.
It is worth remarking in this connection that visually imaged spatial lay-
outs may be used as analogies for an indefinite assortment of problems, in-
cluding logical problems. That such imagistic strategies are quite common
has been hypothesized by Janellen Huttenlocher, among others.* Hutten-
locher presented subjects with tasks of the following sort:
Arthur is taller than Bill.
Tom is shorter than Bill.
Is Arthur taller than Tom?
The subjects in Huttenlocher's experiments describe themselves as pro-
ceeding as follows.
First, they say, they arrange the two items given them in the first premise.
Sometimes Ss [subjects] describe this array as horizontal and sometimes as
vertical. For vertical arrays, Ss claim that they start "building" at the top
and work towards the bottom. For horizontal arrays they claim to start at the
left and work towards the right . . . The goals are to identify the third item
and tsdetermine its position with respect to the other two items.41
Huttenlocher's hypothesis that subjects solve these sorts of logic problems
with spatial imagery is controversial. (It has been challenged, for example,
Internal Codes 271
by H. H. lark.)^^ But it illustrates how mental operations in 'checking' the

correctness of a proof might employ spatial imagery, even though the con-
tent of the proof has nothing whatever to do with spatial matters.
This discussion shows how imagery could figure in the justification stage
of the epistemology of mathematics. But my conception of epistemology
comprehends not only the justification stage of inquiry, but all aspects of
problem solving (see Chapter 6). On this conception the epistemology of
mathematics should include not merely the process of doxastic 'decision-
of deciding which statements of mathematics to believe, or which alleged
proofs to accept as proof-but all facets of the construction of mathemati-
cal ideas and 'products'. Primary epistemology, in particular, should study
the components that may (properly) go into any of the following cognitive
processes or activities: forming or grasping mathematical concepts; invent-
ing mathematical conjectures; and constructing (or trying to construct)
proofs of conjectured theorems.
Let us consider how sense-linked processes-or even perception itself-
might enter these activities. First of all, a number of proposed accounts of
the basic subject matter of mathematics either require or emphasize sense-
linked cognition in the formation or grasp of mathematical concepts. David
Hilbert maintained, for example, that the subject matter of mathematics is
expressions: either physical marks or imageable marks. He therefore held
that the content of mathematics (at least in the simplest cases) is anschau-
lich-a German word that means both visual and intuitive.43
Among current writers, several vaguely physicalistic theories of mathe-
matics are propounded, each implying that the grasp of mathematical con-
cepts is, or can be, closely affiliated with perceptual and imagistic
representations. This seems to be the view, for example, of Michael Resnik
(although details of his view are difficult to pinpoint).44Resnik holds that
mathematics is about abstract patterns. We begin by experiencing physical
objects as patterned, an experience that involves visual, auditory, or other
sensory modes. Mathematical discourse is constructed by abstracting from
concrete patterns. Nonetheless, thinking about abstract entities is fre-
quently accompanied by retention of the visual, auditory, and other sensory
images that led to the abstract conceptions.
Philip Kitcher's view has affinities with ~esnik's.~'Kitcher holds that
mathematics is about certain activities or operations, such as collecting,
segregating, and correlating. Rudimentary arithmetic truths, he says, are true
in virtue of "collective" operations, which can be learned by perceptual
experience. Thus, the child learns that if one performs the collective opera-
tion called "making two," then performs on different objects the collective
operation called "making three," then performs the collective operation of
combining, the total operation is an operation of "making five." Arithmetic,
he says, is about "permanent possibilities of manipulation." Acts of col-
272 Internal Codes
lecting need not, however, be physical. We can colle'ct objects in thought

without moving them about. Kitcher goes on to characterize mathematics
as some sort of idealization of various operations. Still, the general view is
congenial to the idea that the content of mathematics is grasped in the first
instance by acts of perception and spatial imagination.
Penelope Maddy has also taken a position congenial to this outlook. She
holds that certain sets exist in space and time, and that numerical truths
about such sets can be acquired by perception.4s
Turn next to the invention of mathematical conjectures. This too can be
assisted by spatial imagination. Even if the true subject matter of mathe-
matics is completely abstract, as platonists contend, or purely formal, as
formalists contend, a cognizer may best be able to construct and assess
plausible hypotheses about this subject matter by thinking about possible
models or instantiations. This may involve spatial imagery. Although geom-
etry and topology may not be about actual physical objects or relations
among objects, one of the best ways to think about geometrical and topolog-
ical hypotheses, and assess their initial plausibility, is by imagining possible
models of the conjectures in question. The same point holds for set theory,
group theory, and other branches of mathematics.
Even if the proper subject matter of mathematics consists of abstractions
from concrete objects and events, it may assist a cognizer in grasping these
abstractions-and envisaging new ones-to imagine instances from which
the abstractions are derived. There is plenty of anecdotal evidence that
learners grasp abstract ideas better if they are given concrete exemplars,
such as diagrams and illustrations, of the ideas. Perhaps the mind performs
abstractions more easily on sense-derived representations. Take the basic
idea of an ordering, for example. We first encounter orderings through spa-
tial and temporal experiences. We see objects lined up in a row; we hear a
series of sounds in temporal succession. The general notion of an order, so
fundamental to mathematics, is plausibly abstracted from these sensory r e p
resentations. The process of framing new mathematical conjectures may
similarly involve sense-linked representations of various kinds of patterns.
Perhaps this is what the famous mathematician G. H. Hardy had in mind by
his remark, "A mathematician, like a painter or a poet, is a maker of
patterns."47
Finally, how might imagistic thinking find its way into attempts to con-
struct proofs of conjectured theorems? In trying to construct a proof it helps
to represent the general structure, outline, or contours of the envisaged
proof. Many cognizers may find that such a representation consists in an
imaged-spatialstructure. Obviously, this is not a logically necessary compo-
nent, nor perhaps even a psychologically necessary component, of proof dis-
covery. But anything that contributes to a cognizer's problem-solving
power is epistemologically relevant, on my view. And imagery could well
Internal Codes 273
be a noteworthy contributor to the power of solving proof-construction

problems (such as, 'Is there a proof of T? or 'What is a proof of T?'). While
this point does not affect any of the ontological, or foundational, questions
in mathematics, it is certainly germane to the primary epistemology of
mathematics.
12.5. Vividness, Pallidness, and Norm

Let us now examine Nisbett and Ross's allegations of a "disproportionate"
impact of vivid, or sensorily coded, representations. They start by remind-
ing us of the powerful impact of vivid writing, how social movements have
been influenced by books like Harriet Beecher Stowe's Uncle Tom's Cabin
and Upton Sinclair's The Jungle. Although there were probably more au-
thoritative, better reasoned, and more factual writings on the same topics
that these books addressed, their vivid, concrete, and absorbing writing had
more impact. Nisbett and Ross contend that, in general, vivid information
exerts a disproportionate impact on belief as compared with pallid and ab-
stract propositions of substantially greater probative and evidential value."
People have a tendency to "weight" vivid information more heavily, and
this, they imply, is counternormative.
Note that Nisbett and Ross's "vividness" construct cannot be equated
with sense-linked codes. They themselves analyze vividness into several
components, including (a) emotional interest, (b) concreteness and imagery
provokingness, and (c) proximity in a sensory, temporal, or spatial way.
Only component (b)and part of (c) are directly germane to the focus of this
chapter. Still, the overlap with my present concerns is large, and their dis-
cussion raises important questions for primary epistemology.
Nisbett and Ross cite several experimental studies to illustrate the com-
parative impact of vivid information. The following two studies give the
flavor of the point. W. C. Thompson, R. M. Reyes, and G. H. Bower asked
subjects to read testimony from a drunk driving trial.49The vividness of the
testimony was manipulated so that half read pallid prosecution testimony
and vivid defense testimony while the other half read vivid prosecution tes-
timony and pallid defense testimony. For example, a pallid version of prose-
cution evidence stated that the defendant staggered against a table,
knocking a bowl to the floor. In the vivid version the action was described
as knocking "a bowl of guacamole dip" to the floor and "splattering guaca-
mole all over the white shag carpet." Subjects were then asked to indicate
their judgment of the defendant's guilt, both immediately after reading the
testimony and the following day. The immediate judgment of guilt did not
show any effect of the vividness manipulation, but the delayed judgments
did (at least when the defendant was portrayed as having good character).
Subjects exposed to vivid prosecution testimony shifted toward guilty ver-
274 Internal Codes
dicts, and subjects exposed to vivid defense testimony shifted toward not-
guilty verdicts. (However, delayed judgment did not show any effect of the
vividness manipulation for a bad character defendant.)
In an experiment by R. Hamill, T. D. Wilson, and R. E. Nisbett, subjects
were given a description of a single welfare case." The description painted
a vivid picture of social pathology: an obese and irresponsible woman who
had been on welfare for many years, lived with a succession of 'husbands', in
a home full of dirty and delapidated plastic furniture, with cockroaches
walking about in the daylight, and so on. In a second set of conditions this
description was omitted, and subjects were given statistics showing that the
median stay on welfare was two years, and that only 10 percent of recipi-
ents remained on welfare rolls for four years or longer. The pallid statistical
information, however, had no effect on subjects' opinions about welfare re-
cipients; but the vivid description of one welfare family prompted subjects
to express more unfavorable attitudes toward recipients. So a questionably
informative case history had greater effect on inferences than dull statistics.
Anecdotal evidence supporting the vividness-impact hypothesis includes
the following two items (originally given by Nisbett, E. Borgida, R. Cran-
dall, and H. ~eed.)'l (1) Mastectomies performed on Mrs. Ford and Mrs.
Rockefeller in the fall of 1974 produced a flood of visits to cancer detection
clinics, whereas widely disseminated statistics about the lifetime risk for
breast cancer had never produced a comparable impact. (2) While all physi-
cians are aware of the statistical evidence linking cigarette smoking to
cancer, only physicians who diagnose and treat lung cancer victims are
quite unlikely to smoke; and radiologists have the very lowest rate of smok-
ing. So informational vividness seems to influence even sophisticated people
who have been exposed to the most probative data.
Why does vividness have this (putatively) large impact? One factor cited
by Nisbett and Ross is that vivid information is more likely to be stored and
retrieved than pallid information.
There is good evidence that informational concreteness and imaginability

promote recognition and recall. A number of investigators have shown that
memory of pictures is astonishingly good and is markedly better than mem-
ory of either words or sentences . . . This superiority of picture retention is
shown at every retention interval from a few minutes to several months. In
addition, Paivio . . . has shown that recognition and recall of concrete words
. . . are substantially better than of abstract words . . .
Imagery also enables better retention of arbitrarily yoked concepts. Bower
. . . showed that simple paired-associate learning is greatly facilitated when
subjects are given instructions to link the stimulus word to the response word
using images. For example, if subjects are required to learn the stimulus-re-
sponse pair "dog-bicycle," their learning is aided greatly by creating an
image linking the two terms (for example, a mental picture of a dog riding a
bicycle).52
Internal Codes 275
Nisbett and Ross cite other possible explanatory factors, beyond good re-
tention. For example, vivid materials may convey more information; or
they may recruit connected information by means of memory links; or they
may remain longer in (conscious)thought.
Before proceeding to reflect on these possible causes of the impact of
sense-linked materials, I should note that many of Nisbett and Ross's con-
jectures about vividness have been challenged on empirical grounds.
Doubts about their empirical support are expressed by Shelley Taylor and
Suzanne Thompson in a review of the l i t e r a t ~ r e Eight
.~ studies, they re-
port, have directly tested the concreteness hypothesis, manipulating vivid-
ness through concrete and colorful language. Seven found no differences in
attitudes as a function of the appeal. Four studies employed photographs as
manipulations of vividness and found no differences between the pic-
ture/no picture conditions. Thirteen studies compared the impact of video-
taped information with identical or similar information presented through
other media (such as oral or written presentations). Of these, six found no
differences in persuasive impact of various media; five found video presen-
tations to be more effective, but only under limited conditions; one found
print to have more impact than either television or audio presentations; and
one found radio to be a more persuasive medium than television.
Concerning imagery, Taylor and Thompson indicate that actual studies
contradict expectations that imageable vivid material will influence judg-
ments more. The one area where the so-called vividness effect is supported
by empirical studies concerns case histories. Case history information does
seem to have a persuasive impact on judgments, when contrasted with
base-rate or abstract information. However, the extent to which this finding
speaks to the vividness effect is questionable. One possibility, they say, is
that subjects are underutilizing base-rate or other statistical information,
which they do not understand well, rather than overusing case history in-
formation. Perhaps subjects can readily discern the causal relevance of case
history information to the judgments they make, but are less able to see the
causal relevance when information is presented statistically. However this
empirical debate is settled, these last points are extremely important for the
nornuztive assessment of the use of imaged materials, to which I now
proceed.
Nisbett and Ross entitle their chapter on vividness "Assigning Weights to
Data: The 'Vividness' Criterion." The implication is that cognizers fre-
quently have both vivid and pallid data equally before their minds, that
they understand them equally, but just assign greater weight to the former.
However, this is an implausible description even of the cases they describe,
and an inaccurate account of even (some of) their own diagnoses. In several
of the studies they cite, subjects are given either vivid or pallid information,
but not both. Thus, they are not preferring one kind of information over the
other. They are just operating separately either with the vivid information
276 Internal Codes
or with the pallid. Now it may well be (as Taylor and Thompson suggest)
that the subjects tend to underutilize the pallid, statistical information. But
this does not show that they are doing anything wrong in deploying their
vivid, concrete, sensorily imageable information. Consider the women who,
upon learning of the well-publicized mastectomies of Mrs. Ford and Mrs.
Rockefeller, visit cancer detection clinics. Surely this is not an obvious cog-
nitive error on their part. Maybe their earlier failure to take similar action
upon reading statistical data on breast cancer was irrational. So why is posi-
tive responsiveness 'to the case histories of Mrs. Ford and Mrs. Rockefeller
counternormative?
Suppose sensorily coded data indeed have more influence over inference
and action than nonsensorily coded data. Does this betoken any epistemic
flaw in the handling of sensorily coded data? That is not obvious. As Nisbett
and Ross themselves emphasize, one factor responsible for this situation is
that sensorily coded data are better retained and recalled. But, surely, there
is nothing counternormative about accurate and efficient recall of genuinely
learned facts! There cannot be any epistemic objection to the memory sys-
tem for being good at remembering sensorily coded information.
Where, then, can the fault lie? It could lie, as Nisbett and Ross's chapter
title intimates, in the handling of the total 'clump' of information available.
But how is that faulty? If the pallid statistical information is not well re-
membered, it seems perfectly proper to rely only on the other information,
which is remembered. The executive portion of the mind, we assume, can
only operate on activated materials. It is blameless if it fails to draw con-
clusions from materials once encountered but no longer available. One
could say that there is a flaw in the memory system: it ought to recall statis-
tical information better. However, even if we sympathize with this wish,
how would it reflect ill on the accurate retention and retrieval of imaged
information?
A further point-also suggested by Taylor and Thompson-should be
stressed. Perhaps statistical information just is not well understood. This
could influence either how well it is remembered (that is, poorly), or how
much it is used when it is remembered. Recall (from Chapter 10) that
strength of retention seems to be influenced by depth of processing. Poorly
comprehended material will not be processed deeply, because one does not
know what connections to establish between it and previous items in.LTM.
This possibility would accommodate the fact that, as Taylor and Thompson
note, many studies do not show appreciable inferiority of printed material.
Perhaps it is only the complexity of statistical materials that make them
p~orly~remembered, not the fact that they are not sensorily coded.
Suppose that statistical information is recalled but poorly understood,
and the executive system fails to draw permissible conclusions from it. Is
this counternormative?It might be argued, on the one hand, that the system
ought to understand statistical data better. Perhaps, but on the other hand,
Internal Codes 277
is there anything wrong with the refusal to draw conclusions from informa-
tion which, as a matter of fact, is not well understood? I should think not.
To draw conclusions in this situation amounts to random guessing or grop-
ing in the dark. That is not likely to promote reliability.
What about the suggestion, then, that the system behaves counternorma-
tively if it fails to understand statistical data? My reaction-to be elabo-
rated upon in later chapters-is that it is unreasonable to require this of the
system's basic architecture. Just because some species of information would
be cognitively useful, it does not follow that a system is poorly designed if it
is not equipped with an innate propensity to grasp that information. (If the
information is in principle ungraspable by the system's architecture, that is
another matter. This is not in question here.) Granted, statistical informa-
tion is something it is good for human beings to learn to understand and uti-
lize. (This might even be a principle of secondary epistemology.) Failure to
understand statistics at the outset, however, is not obviously a flaw in the
system's fundamental architecture.
This topic, however, transcends the issue of sense-linked thinking. Is
there evidence that something is wrong with our cognitive system's use of
sensory codes? For example, is its use highly unreliable, or does it contrib-
ute to unreliability? As I have emphasized, the recall strength of sensorily
coded materials is hardly a defect. The most that Nisbett and Ross are enti-
tled to claim, I think, is that (1)sensory information is likely to concern indi-
vidual, concrete cases, and (2) the system has a tendency to draw
unwarranted conclusions from small samples (for example, isolated case
histories). However, there is no necessary connection between sensorily
coded information and sample size. Imagery can, in principle, carry infor-
mation about large numbers of cases. For that matter, it can encode and
store information about statistical tendencies: it can help represent and re-
member diagrams or charts that present statistical patterns. Moreover, if
the system does tend to make improper inferences from small samples, that
is not a problem with sensory coding per se. It is far from clear, then, that
operations governing sensory codes are deficient or defective.
One moral of this discussion-encountered in previous chapters as
well-is that a very complex bundle of operations needs to be surveyed and
appraised in order to construct even a small piece of a right system of
J-rules. It is because of this complexity that I have eschewed any attempt to
present and defend any particular J-rule system. The reader may be able to
extract from my discussions any number of candidate operations that might
(or might not) be permitted by a system of J-rules. A proper appraisal of
such a system, though, must weigh the consequences of the system as a
whole. This contrasts with the (self-confessed) propensity of Nisbett and
Ross to underline the dangers and costs of certain cognitive operations (in
the present case, deployment of sense-linked materials), not to weigh their
overall performance in the context of a complete cognitive system.
chapter 13
Deductive Reasoning
13.1. Logicality, Rationality, and Intelligence

The last two chapters concerned aspects of mental representation. I now
return to a more traditional preoccupation of epistemology: the formation
of belief (or partial belief) as it occurs in reasoning or inference. This chap-
ter tackles deductive reasoning.
We saw in Chapter 5 that the question of deductive norms is complicated
by the fact that people sometimes have false beliefs. Given this fact, it is not
a generally correct policy to infer any and every deductive consequence of
what you antecedently believe. If you suddenly recognize that some propo-
sition p follows from your prior beliefs, you might reasonably conclude, not
that p is worthy of belief, but that some previously believed proposition or
propositions should be rejected. In this chapter I want to avoid this kind of
complication. Therefore, rather than discuss deductive reasoning in the
context of an entire belief corpus, I shall isolate the topic of deduction by
focusing on judgments about the deductive validity of arguments. This is
convenient for another reason: most of the psychological literature on rea-
soning focuses on argument validity.
What is the epistemological significance of psychological research on de-
ductive reasoning? What can normative epistemology learn from psychol-
ogy? Doesn't pure logic distinguish good deductive reasoning from bad, that
is, valid deductive reasoning from invalid? And isn't this all that epistemol-
ogy needs?
I made a start on answering these questions in Chapter 5. Even setting
aside the problems of whether and when to make deductions from prior be-
liefs-a question to which pure logic offers no answer-logic alone yields
no theory of justified belief. Although logic gives principles for deciding
whether an argument is valid or invalid, it gives no principles for deciding
whether a person who correctly ascribes validity (or invalidity) to a given
argument makes that ascription justifiably. The theory presented in Chap-
ter 5 says that whether a cognizer's belief is justified depends on the basic
Deductive Reasoning 2 79
psychological processes he uses to form this belief. The psychology of rea-

soning is relevant to justifiedness, then, because it inquires into the basic
cognitive processes that can occur in reasoning.
The epistemological significance of the psychology of reasoning is also
linked to the evaluative dimension of rationality. Epistemology is presum-
ably interested in such issues as: (1)Are human beings rational? (2) If they
aren't perfectly rational, how rational are they? What are their flaws, or
deficiencies, in matters of rationality? (3) How can human rationality be
improved, or ameliorated? It is evident that the psychology of reasoning
should be relevant to these questions, since whether, or to what degree,
humans are rational presumably depends on their (contingent) psychologi-
cal traits.
There is also another reason why psychology is relevant. The very stan-
dards for rationality cannot be specified without recourse to psychological
facts. Plausible standards of rationality cannot be introduced by reflection
on logic alone. They need to appeal to basic psychological capacities of
human beings. Let me elaborate this point.
On one popular approach to epistemology the field should proceed by
advancing certain ideal models, which encapsulate what a perfect cognizer
would be like. It would then evaluate human cognitive performance by ref-
erence to such an idealized agent. The strategy of idealization is familiar
from science, where ideal models are often introduced. The classical gas
laws only hold for 'ideal' gases; the laws of classical economics only govern
'perfect' competition; and models of grammatical competence idealize
away defects of performance. These idealization strategies, of course, are
directed at explanatory and predictive ends, not evaluative or normative
pursuits. But idealization may well serve normative undertakings as well.
This strategy was formulated, for example, by Carnap:
Rational credence is to be understood as the credence function of a com-
pletely rational person X; this is, of course, not any real person, but an imagi-
nary, idealized
How would this approach run for logical, or deductive, rationality?It might
seem easy, at first blush, to say how it would run. If one tries to keep the
approach plausible, however, it quickly encounters difficulties and
ambiguities.
The first possible ideal model is logical omniscience. A being is logically
omniscient, we might say, if and only if (a) it believes of every logical truth
that it is a logical truth, and (b) it believes nothing to be a logical truth that
is not a logical truth. One question about this conception is the scope of the
term 'logical'. Does logic include second-order as well as first-order theory?
Does it include recursive function theory, proof theory, model theory, and
all kinds of deviant and intensional logics? Obviously, many of these
280 Deductive Reasoning
branches of logic are indistinguishable from mathematics. But, on this ideal

conception of rationality, a cognizer would be irrational--or at least less
than completely rational-if he were ignorant of any facts of these disci-
plines. Yet that seems counterintuitive in the extreme. Why should failure
to know (all the truths of) certain, highly specialized branches of mathemat-
ics make one irrational? Clearly, failure to know all the truths of topology or
group theory would not make one irrational, or even less than fully rational.
(Lots of these truths, of course, have not yet been discovered by anyone.)
Why should failure to know all truths of recursive function theory make one
irrational, or even less than fully rational?
Even if the scope of logic is restricted to classical first-order theory, the
logical omniscience model seems ill-motivated. Consider a being who does
not believe all truths of (first-order) logic simply because he has not had
much interest in logic, and therefore has not devoted much attention to de-
termining its set of truths. Such a being might still have ideal logical capaci-
ties, even though he is not logically omniscient. Certainly he should not be
deemed irrational, or less than completely rational, just because he has not
exercised those capacities. A better conception of ideal logicality, then,
would be one that specified the sorts of logical tasks that an ideally logical
being could execute.
However, there are several different kinds of logical tasks that can be
specified. So if we seek to define an ideally logical being (ILB) in terms of
some set of logic tasks it is capable of performing, we need to choose the
tasks in question. To simplify matters, restrict attention to semantic tasks,
specifically those concerning implication, or argument validity. Even with
this restriction, tasks may be divided into two classes: Wh-tasks and Y/N-
tasks. Wh-tasks pose questions of the form: 'What conclusions follow from
this set of premises?' Y/N-tasks pose questions of the form: 'Does this (spe-
cified) conclusion follow from these premises-yes or no?
Starting with Wh-tasks, two levels of performance suggest themselves as
possible candidates for marks of an ILB. (1)An ILB might be defined as one
who, for any set of premises, would generate and believe all and only the
conclusions that follow from that set. (2) More weakly, an ILB might be de-
fined as a being who, for any set of premises, would generate and believe
only conclusions that follow from that set. In other words, on the first of
these conceptions an ILB must have the capacity to provide a complete, or
exhaustive, list of correct implications; whereas on the second conception
an ILB need only have the capacity to produce a sound list, not necessarily
a complete one. The latter would demonstrate total reliability in the pro-
d u c t i o ~of implications, but not total power. Indeed, the list of correct con-
clusions could be as short as you please (even null?).
Reference to Y/N-tasks would provide a different characterization of an
ILB. An ILB could be defined as a creature who, for any queried argument
Deductioe Reasoning 281
<p, q>-where p is the premise set and q the conclusion-would correctly

determine whether <p, q> is valid and would believe that answer.
Needless to say, one could select an even stronger conception of an ILB
by requiring capacities to perform all of these tasks (at the highest level of
performance). But let us look at the individual tasks, since some of these
alone may be too strong for the purposes at hand.
How adequate are any of these conceptions of an ILB as a model of
rationality? The model that requires an ILB to generate and believe all im-
plications of a queried premise set obviously requires belief in infinitely
many propositions, since any premise set has infinitely many consequences.
On the approach to belief I have advocated, this immediately disbars
human beings. But this, like previous proposals, is very counterintuitive.
Should mere finitude of belief capacity imply that a being is irrational, or
even incompletely rational? To the extent that I have definite intuitions
about 'rationality', this requirement seems far too strong.
The second Wh-task model, however, does not have this drawback. It
merely requires an ILB to construct (and believe) a sound list of implica-
tions, which can be as short as you please. This model, however, might well
be regarded as too weak. What if a certain creature could only identify im-
plications of the form (p or r) for any premise p? Is this sufficient for perfect
rationality? I think not. This demonstrates only a low level of logical
proficiency.
Turn next to the Y/N-task model. This criterion implies that a being is
rational only if it can decide the validity or invalidity of any argument in
first-order logic-which, of course, includes general quantification theory.
But general quantification theory is not decidable, as Alonzo Church
proved. There is not (and cannot be) any effective, algorithmic procedure
for deciding validity or invalidity. There is a mechanical procedure that
will unfailingly show any valid argument valid: if an argument is valid, the
routine that shows validity will terminate. If the argument is not valid,
though, the routine can go on and on without any decision.
Is it reasonable to characterize rationality in a manner that transcends
computational possibility? This is very dubious. If there is no possible pro-
cedure that would ensure the execution of a specified set of tasks, how can
anyone be called 'irrational' who lacks the capacity to execute those tasks?
Such a degree of idealization is surely excessive.
We could lower the requirements still further. We could just require of
aq ILB determination of consistency. Or we could require determination of
validity for truth-functional arguments only, where effective decision pro-
cedures do exist (for example, using truth tables).
Suppose there exist effective procedures for solving certain kinds of logic
problems but a cognizer fails to use any such procedure. Does it follow that
he is irrational, or imperfectly rational? Here I wish to inject a new element
into the discussion: speed. Suppose that the only effective procedure for
solving a certain class of problems would be very slow for a certain subclass
of those problems, so that use of the procedure for that class of cases would
not yield an answer in the being's lifetime. Is a being who uses a noneffec-
tive procedure instead-a procedure not guaranteed to produce a correct
answer, but which might produce one in a timely fashion-is such a being
irrational? Christopher Cherniak considers a cognizer who wishes to test a
system of propositional beliefs (beliefs involving no quantificational struc-
ture) for consistency, and his whole belief corpus contains (only) 138beliefs.
The truth table method provides an effective procedure for testing this cor-
pus for consistency. But if each line of the truth table for the conjunction of
these statements can be checked in the time a light ray takes to traverse the
diameter of a proton-an appropriate cycle time for an ideal computer-a
truth-tabular consistency test would still take more time than the estimated
twenty billion years from the dawn of the universe to the
As this example illustrates, the theoretical adequacy of a method is hardly
equivalent to practical adequacy or efficiency. Should we say, nonetheless,
that creatures who use noneffective procedures when effective procedures
exist are imperfectly rational? Again, that seems wrong. Switching to the
evaluative term 'intelligent', I feel even greater confidence in saying that
fast but noneffective procedures may be more intelligent than effective but
laborious procedures. (Notice that 'slow' is one approximate synonym of
'unintelligent'.)
The general topic of speed of computation is currently studied under the
title of 'computational complexity'.3 Complexity theory studies the practi-
cal feasibility or unfeasibility of algorithms. An algorithm's feasibility is
evaluated in terms of whether its execution time grows as a polynomial
function of the size of input instances of the problem. If it does, the algo-
rithm is generally treated as computationally feasible; if it increases faster,
usually as an exponential function, the algorithm is regarded as computa-
tionally intractable. Such intractability turns out to a large extent to be in-
dependent of how the problem is represented and of the computer model
involved.
The problem of determining whether a truth-functional argument is
valid belongs to a set of problems called ' ~ ~ - c o m ~ l e tWhile
e ' . ~ NP-com-
plete problems have not been proven inherently to require exponential
time, they are strongly conjectured to do so; and hence are thought to be
computationally intractable. So complexity theory suggests that even some
comparatively simple reasoning tasks cannot feasibly be performed in ways
guaranteed to be correct. Thus, it seems implausible to say that human
beings are 'irrational' or 'unintelligent' just because they fail to possess, or
utilize, algorithms for such tasks.
Reflections on the speed desideratum, and on the sketched results of
Deductive Reasoning 283
complexity theory, suggest the difficulty of giving an unambiguous charac-

terization of an ILB. Even if we could agree on some such characterization,
though, I question whether it is a sensible requirement for 'perfect rational-
ity', or 'perfect intelligence', that one should be an ILB. The point I now
have in mind is one I hinted at earlier. Suppose we could identify a set of
optimal logical operations. Does it follow that someone who fails to possess
or deploy these operations (in appropriate task situations) is unintelligent
(or imperfectly intelligent)? This ignores an alternative, and much more
plausible, diagnosis: that he lacks a perfect education or training in logic.
An oft-voiced criticism of IQ tests is that they really test acculturation or
acquired knowledge, rather than intelligence. This makes it clear that, in-
tuitively, intelligence consists in a 'raw' ability to learn or perform other
cognitive feats. It is to be distinguished from acquired skills and informa-
tion. If a person lacks optimal logical operations, that need not be ascribed
to suboptimal intelligence, any more than a failure to deploy optimal opera-
tions for solving topology problems betrays suboptimal intelligence. In both
cases we might well say that optimal operations are things to be learned,
not things we expect to find in a creature's cognitive architecture.
However, it is certainly plausible to hold that some, at least minimal, log-
ical abilities are required for learning, and for general problem solving. So
unless a creature has at least those minimal abilities, as part of its cognitive
architecture, it would indeed count as irrational or unintelligent. But it is
not obvious that there is some unique minimal set of logical abilities that is
requisite for learning and problem solving. A number of distinct sets of abil-
ities might each suffice.
There are two ways one might proceed at this point. One might explore
the minimal logical abilities necessary for rationality by constructing vari-
ous sets of such abilities, implementing them on computers, and seeing how
well they can learn, soIve probIems, and so on. This is the route taken by
(some) practitioners of artificial intelligence. Another route is to examine
the fundamental logical endowments of human beings, on the assumption ,
that human architecture includes only rudimentary logical abilities. These

abilities enable people to learn and solve problems, but do not generally
confer the highest attainable logical skill without formal instruction in logi-
cal concepts and techniques. Nonetheless, since these abilities suffice for
what we ordinarily call rationality or (moderate) intelligence, we can iden-
tify a baseline for these evaluative predicates by exploring the native logical
endowments of human beings. And this exploration belongs to the psychol-
ogy of reasoning.
The distinction I have been drawing between basic logical abilities and
acquired logical abilities might be paralleled by a distinction between basic
rationality and acquired rationality, or between basic intelligence and ac-
quired intellectual skill. We might then say that it is the job of primary
epistemology to identify the components of basic rationality or basic intel-

ligence; whereas the identification of the methods and techniques appropri-
ate to acquired rationality or acquired logical skill would fall to secondary
epistemology. Since my goal is to construct a primary epistemology, it is the
contours of basic rationality or basic intelligence that will be explored. In
this investigation the psychology of reasoning has a natural place.
13.2. Basic Reasoning Abilities

What are people's 'basic', 'natural', or 'intuitive' logical abilities? An opti-
mistic view is that people are constitutionally good logicians. People's natu-
ral reasoning proclivities are not only 'precursors' of formal logic, but carry
people quite far on logical tasks even without formal instruction. No doubt,
formal training can augment people's repertoire of deductive strategies (for
example, the reductio ad absurdurn technique); it can show them how to
deal with complex problems systematically; and it can increase their speed
in solving logic problems. But, according to this optimistic view, native
logic abilities are fundamentally quite sound and moderately powerful.
Formal training only polishes these abilities, makes them more systematic,
and explains underlying theoretical principles.
A pessimistic view, by contrast, would hold that people are constitution-
ally so illogical that they can only become logical by being completely 're-
programmed'. Even more pessimistic, it might be held that we are natively
such bad logicians that even formal training cannot eradicate our funda-
mental logical flaws. I do not know of anyone who holds this last view, but
something rather like it has been held about peqple as intuitive statisticians
or probability assessors. Not only do people have a very poor intuitive grasp
of statistical principles, it has been held, but even those highly trained in
statistics persist in certain blunders (see Chapter 14).A similarly bleak view
of human logical endowments is conceivable as well.
It is hard to hold a completely bleak view of human educability in logic.
At a minimum, large numbers of people are capable of learning to 'do'
logic, in the sense that they can pass formal logic courses, which involves
solving many problems and answering many questions correctly. But do
they really understand what they are doing? Many a logic instructor feels
that a large proportion of students can only do problems mechanically,
without true understanding. And if our constitutional logical endowments
were so strong, shouldn't formal logic be easy and perspicuous to everyone?
I want to shift the question now from understanding to justification.
When people study formal logic, they presumably come to believe in the
correctness of the procedures they follow. But are these beliefs justified?
According to my theory of justifiedness, justification involves (at least) pri-
mary justifiedness, that is, the use of approved basic processes. Does our
constitutional endowment provide suitable basic processes?
To be more concrete, suppose someone has heard promulgated certain

logical principles, and he adopts those principles. Suppose these principles,
or 'methods', are formally correct. Then even if he uses those methods to
determine the validity of a given argument, his belief in the argument's va-
lidity may not be justified. In particular, his belief is not justified unless he
has acquired the methods appropriately, for example, by justifiably believ-
ing them to be correct (see Chapter 5). But in order for a belief in a
method's correctness to be justified, it must have been arrived at (or subse-
quently reinforced) by an approved set of basic cognitive processes. Justifi-
cation must ultimately rest on basic processes.
Three possibilities now unfold. First, primary justifiedness vis-A-vis logi-
cal principles may arise from the application of some sort of native deduc-
tive processes. Second, primary justifiedness vis-A-vis logical principles may
arise from application of nondeductive processes. Third, people may not be
justified vis-8-vis these principles at all. They may lack any permissible set
of basic operations to qualify for justification. Let me examine these possi-
bilities briefly.
Here is one way a person might come to believe in the soundness of
modus ponens (the first possibility). He first constructs a truth table and
notes that whenever two sentences of the forms P and (If P then Q) are both
true, then the sentence of the form Q is also true. He then argues as follows:
(1)In every possible situation in which both P and (If P then Q) are true, Q
is true too. (2) If, in every possible situation in which both P and (If P then
Q) are both true, Q is true too, then the inference schema <P; If P then Q;
therefore Q> is sound. Therefore (3),the inference schema <P, If P then Q;
therefore Q> is sound. Obviously, this argument is an application of modus
ponens. But if this is part of a natural, or basic, inferential routine, it could
well be approved by a right system of J-rules. Some might see a danger of
circularity here. That seems to be the worry of Lewis Carroll in his famous
dialogue "What Achilles Said to the ~ortoise."~ My theory of justifiedness,
however, gives no grounds for such a concern.
The second possibility is the conferral of justifiedness by a nondeductive
process. For example, someone hears the soundness of modus ponens pro-
pounded by a certain authoritative figure, such as the teacher of his logic
course. His reliance on this personage might be a specimen of good induc-
tive reasoning. Perhaps he has independently verified, on previous occa-
sions, the particular speaker's credibility, or the credibility of any lecturer
employed by the school or university he attends. Then his belief in the
principle's correctness might be justified.
But what about authorities themselves? How are their beliefs in logical
procedures justified? If we do not have native deductive processes to do the
job, are there suitable nondeductive processes that do not appeal to the au-
thority of others? In earlier eras epistemologists turned to 'intuitions' about
logic to salvage the situation. But nobody has ever made it clear what kind
of cognitive process an 'intuition' is. Nor is it clear what would enable in-
tuitions to confer justifiedness. On my theory intuitions would have to be
very reliable. But what is the nature of an intuition that might give it relia-
bility? Lacking answers to these questions, appeal to intuitions is not at all
appealing.
The third possibility is that there are no basic processes by which belief in
logical principles can be arrived at justifiably. This possibility, of couFse,
would be savored by the skeptic.
So far I have mainly been conceiving logical endowments as inference
rules, procedures, or transformational routines. But there is another general
category in which our logical endowments might be described. This is the
set of logical concepts, or operators, that are native parts of our cognitive
equipment (assuming that some concepts must be native). Does our native
conceptual repertoire feature the concepts of implication and consistency?
Does it have the standard repertoire of operators familiar from formal logic:
'not', 'if-then', 'if and only if, 'or', 'there exist some', and the like?
Most of the research on the psychology of reasoning has concentrated on
the repertoire of inference procedures. But the category of logical concepts
is just as fundamental. It is often suggested, of course, that people are not
familiar with the material conditional. But it is also questionable whether
they have an antecedent hold on the logician's concept of validity or logical
implication. One possible explanation of why logic is difficult (for some stu-
dents) is that the concepts of logic do not have precise analogues in our na-
tive system of thought. Learning logic, on this scenario, is not just a matter
of translating new notation into the language of thought. Formal logic ben-
efits people because it introduces a new array of concepts and displays pro-
totypes of argumentation not already contained in thought.
In the section that follows, I shall examine some selected work in the psy-
chology of reasoning that bears on the issues I have just broached, especially
the question of justifiedness. Very little of this, however, will deal with na-
tive logical concepts, since less research has been done on this notoriously
difficult topic.' In fact, the entire topic of reasoning is very difficult, and the
research I shall sample will not manage to answer (to everyone's satisfac-
tion) all the questions I have raised. Nonetheless, this research gives illu-
minating hints at the sorts of answers that would be of interest.
13.3. Mental Logic versus Mental Models

There are two main competing theories in the current psychology of de-
ductive. reasoning: the mental logic theory and the mental models theory.
Roughly, the mental logic theory seeks to explain people's performance on
logic tasks by positing mental inference rules that carry out mental deriva-
tions. This approach assumes that people possess internal deduction sche-
Deductiue Reasoning 287
mata, such as mental analogues of the ones found in Gentzen- or Jaskowski-

style natural deduction systems. They might include rules like modus
ponens, conjunction introduction, universal instantiation, and so on, that
are familiar from introductory texts in predicate logic. Theories of this type
have been proposed by P. N. Johnson-Laird, D. N. Osherson, M. D. S.
Braine, and Lance ~ i ~ sThe
. ' mental models approach claims that deduc-
tive reasoning is a matter not of mental proofs, but of mental models.
Roughly, people allegedly construct a kind of inner 'semantics' of an argu-
ment's premises, and attempt to manipulate the constructed model or
models to assess the argument's validity. The chief proponent of this ap-
proach is Johnson-Laird, who has now renounced his own earlier endorse-
ment of the mental logic doctrine.
The best recent development of the mental logic theory is Rips's article
"Cognitive Processes in Propositional Reasoning." This will serve as our
prototype of this approach. Rips's article outlines a computer model of pro-
positional reasoning, intended to simulate the processes of untrained sub-
jects in deciding the validity of arguments. The model is nicknamed
"ANDS," for "A Natural Deduction System." Although it incorporates
many of the sorts of rules used in natural deduction systems, it also features
routines that control the steps which are taken in trying to construct a
proof. Natural deduction inference rules only say what inferences are per-
mitted from what previous steps. They do not instruct, guide, or control a
person's application of such rules for the task at hand. But consider the situ-
ation of a person who is asked whether a specified conclusion logically fol-
lows from a given set of premises, for example, whether the conclusion
below follows from the premise:
If there is not both an M and a P on the blackboard, then there is an R.
(Therefore) If there is no M, then there is an R.
Even if a person in some sense 'possesses' a list of inference rules, these do
not say when they should be applied in trying to construct a proof of a con- .
clusion from some premises. Some sort of control system is needed. Such a
system will guide the agent in a goal-directed manner, toward constructing
a proof of the specified conclusion from the specified premise. Nothing in a
standard natural-deduction system corresponds to such a goal-directed sys-
tem. Only such goal-directedness can enable an agent to solve predesig-
nated problems, either in logic or in other areas.
Reasoning in ANDS means constructing a proof in working memory. The
two chief components of these p r o o f ~ a l l e d"treesw-are the assertion
tree and the subgoal tree. An assertion tree encodes a natural-deduction
proof (or part of a proof). It records the logical steps leading from the prem-
ises to the conclusion of an argument. The subgoal tree may be thought of as
the reverse pathway from the conclusion to the premises. Typically, ANDS
proves its theorems from "outside in," working alternately backward from
the conclusion (in the subgoal tree) and forward from the premises (in the
assertion tree). The subgoal tree has no obvious counterpart in formal logic
proofs. Subgoals keep the proof procedure aimed in the direction of the ar-
gument's conclusion rather than allowing it to produce implications at ran-
dom from the premises. (This basic idea is due to A. Newel1 and H. A.
~imon.)'
ANDS' inference routines control its proof endeavors by placing new as-
sertions and subgoals in memory-these are the actions instigated by the
routines. Each routine consists of a set of conditions that must be checked
against the memory trees to determine whether the routine applies. If all
the conditions are met, then a series of actions are carried out that modify
the trees in specific ways. The routines are tested for applicability in a fixed
order, and the first routine whose conditions are fulfilled is then carried out.
Here are two examples of ANDS' inference routines, formulated in a
condition-action format. (This format resembles the format of production
systems, which I will explore in Chapter 17.) Corresponding to ordinary
modus ponens, ANDS has two inference routines, R1 and R2:
R1 (modus ponens, backward version)
Conditions: 1. Current subgoal = q
2. Assertion tree contains IF p, q
Actions: 1. Set up subgoal to deduce p
2. If Subgoal 1is achieved, add q to assertion tree.
R2 (modus ponens, forward version)
Conditions: 1. Assertion tree contains proposition x = IF p, q
2. x has not been used by H1 or R2
3. Assertion tree contains proposition p
Actions: 1. Add q to assertion tree.
Forward and backward routines occur in several prior artificial intelligence

theorem provers,, such as D. Hewitt's PLANNER, and D. V. McDermott
and G. J. Sussman's CONNIVER.'
Notice that backward rules like R1 are not passive implementations of
logical principles; they incorporate heuristics that specify the sorts of proof
strategies which are likely to pay off. This illustrates why logic, narrowly
considered, does not suffice for intelligence, which embodies the capacity
for goal-directed activity. Logic per se does not provide goal-oriented
problem-solving strategies even in logic tasks! (Actually, many logic text-
books do include practical suggestions for finding proofs, for example,
Quine's techniques of the "fell swoop" and the "full swap" in Methods of
Logic.'' But these are not strictly part of a natural deduction system or an
axiomatic system.)
To make ANDS more realistic, Rips adds the assumption that his fourteen
inference routines are unavailable on some proportion of trials (where una-

vailability may be due to failure to retrieve a rule, to recognize the rule as
applicable, or to apply it properly). He assigns a probability to each routine
to indicate its (degree of) availability. He then assumes that subjects, asked
to judge the validity or invalidity of an argument, will attempt to construct
a mental derivation. If they successfully deduce the conclusion, they will
judge the argument to be valid. Otherwise, they will either guess at the an-
swer or simply declare the argument invalid. Other factors-such as mis-
perceiving or misunderstanding part of the argument--could keep subjects
from finding a correct proof. But Rips assumes, for simplicity, that these
sources of error are negligible compared to rule availability.
To make precise predictions of subjects' responses to various arguments,
Rips formulates an equation that embodies two main elements: the proba-
bility that a subject will construct a proof-and therefore answer 'valid'-
and the probability that he won't construct a proof but will guess the an-
swer 'valid'. After assigning values to the pertinent parameterethe
availability probabilities of each of the rules and the guessing parameter-
the model makes specific predictions for particular problems.
In one experiment subjects were given 104 arguments to assess for valid-
ity. Among the valid arguments in this experiment, subjects correctly
judged them valid on about 51 percent of trials, with scores for individual
arguments varying from about 17 to 92 percent. (The arguments clearly
spanned a wide range of difficulty for the subjects.) For the invalid argu-
ments, the percentage of (incorrect)'valid' responses was about 23 percent.
These observed percentages came out very close to the percentages pre-
dicted by the model for each of the problems. In addition to this sort of test
of the model, Rips provides other empirical evidence in the form of (a) pro-
tocols of a subject while working on a problem, and (b) memory for proofs.
Granted that the evidence cited in Rips's article provides substantial
support for the theory, there are problems and questions that can be raised.
Of substantial interest, at least for my purposes, is the fact that Rips's model
posits no fallacious or incorrect inference propensities. All rules embodied
in the fourteen routines are sound. (Rips indicates that the routines are not
intended to be exhaustive; but he does not suggest that any fallacious rou-
tines might be added.) True, the model allows for erroneous judgments; but
these can only arise from guesses that are offered when no proof is found.
There is plenty of other ostensible evidence, however, of fallacious infer-
ence propensities, at least for some people. Furthermore, there are certain
kinds of sound inference rules-for example, modus tollenjthat seem to
be poorly grasped by many subjects. (Interestingly, Rips omits any modus
tollens rule from his routines.) One would like to know why modus tollens is
not so well grasped by human beings as other inference principles. Are
there some deeper-lying facts about mental logic that explain this? If so,
290 Deductiue Reasoning
they are wholly neglected by the present approach. I will return to this
issue in section 13.4.
We come now to the mental models theory. According to this approach,
reasoning is not a matter of applying abstract logical rules at all; it is a mat-
ter of creating and manipulating internal analogues. The internal analogues
'simulate' what is asserted in the premises and conclusion. By manipulating
these simulations, allegedly, one arrives at a judgment of validity or inval-
idity, or at a conclusion that appears to follow from the specified premises.
This approach has something in common with other developments in cog-
nitive science that stress inner representations of concrete instances, or ex-
emplars, such as the prototype theory of concepts.
The best developed version of the mental models approach to reasoning
is presented by Philip Johnson-Laird, in Mental ~ o d e k . As
' ~ he explains it,
the models people use to reason "are more likely to resemble a perception
or conception of the events (from a God's-eye view) than a string of symbols
directly corresponding to the linguistic form of the premises . . . The heart
of the process is interpreting premises as mental models that take general
knowledge into account, and searching for counterexamples to conclusions
by trying to construct alternative models of the premises."12
How, in particular, would this be done?13Suppose you are given a pair of
syllogistic premises and are asked to produce a conclusion that follows from
these premises. (This is the standard task that Johnson-Laird has used in his
experiments, rather than the task of deciding the validity of a full argument,
with conclusion specified.) For example, you may be given the following:
(1) Some of the artists are not beekeepers.
All of the clerks are beekeepers.
Therefore, ???
According to the theory, the first thing to do is to interpret the individual
premises, and you do this by configuring a set of mental exemplars that cor-
respond to particular artists, beekeepers, and clerks. The theory stipulates
that the model for the first premise is the one in (2), and the model of the
second premise is that of (3).
(2) artist
artist
(artist) = beekeeper
beekeeper
(3) clerk = beekeeper
clerk = beekeeper
(beekeeper)
(beekeeper)
In these diagrams the exemplars are indicated by words, though the words
themselves have no significance other than to differentiate exemplars of dif-
ferent types. To indicate identity among the exemplars, you place an equal-
ity (or identity) sign between them. So the equal sign between clerks and
beekeepers in (3) represents the fact that those clerks also are beekeepers.
In (2) the artists above the line are those that are definitely not beekeepers.
However, the parenthesized artist below the line symbolizes the possibility
that there may be an artist who is a beekeeper. In (3)both of the clerk ex-
emplars are beekeepers and there may or may not be beekeepers who are
not clerks.
After you form these representations of the premises, you have to com-
bine them into a unified model of the problem as a whole. One such model
might be:
(4) artist
artist
(artist) = beekeeper = clerk
beekeeper = clerk
(beekeeper)
(beekeeper)
In order to generate a conclusion to the syllogism, you then try to 'read'
from this model the relationships that hold between the 'outer' terms: 'art-
ists' and 'clerks'. Two conclusions that suggest themselves on the basis of the
unified model is 'Some artists are not clerks' and 'Some clerks are not art-
ists'. However, to be sure that these candidate conclusions really follow, you
need to check that there are no alternative models of the premises that
make these conclusions false.
Johnson-Laird claims that his theory can successfully predict the diffi-
culty people have in producing correct conclusions.
The major source of difficulty should be the construction and evaluation of
alternative models. This process places an additional load on working mem-
ory-the memory system that is used for holding the mental models while
they are manipulated. The greater the number of models to be considered,
the harder it should be to construct and to evaluate them. Hence, the task
should be harder when it is necessary to construct two models, and harder
still when it is necessary to construct three models.14
These predictions are borne out by experiments that Johnson-Laird did
with various colleagues. In one experiment, for example, twenty students at
Teachers College, Columbia University were given all sixty-four possible
pairs of syllogistic premises. For the premise pairs requiring one model to
be constructed to yield the correct conclusion, 92 percent of the correct
(valid) conclusions were drawn. Of the premise pairs requiring two models,
46 percent of the correct conclusions were drawn. And of the premise pairs
requiring three models, 27 percent of the correct conclusions were drawn.''
Johnson-Laird also believe-uite plausibly-that a good theory of rea-
soning should explain children's ability at reasoning. With this in mind, he

did an experiment on syllogistic reasoning involving two groups of intelli-
gent children: one group of nine- to ten-year-olds and another group of
eleven- to twelve-year-olds. Unlike most adults, these children could only
(in the main) draw valid conclusions that depended on constructing one
mental model. The older group responded correctly on 58 percent of the
one-model problems-there were eight such problems and a child had to
get four of them right to perform reliably better than chance. Not one of the
older subjects drew a correct conclusion to the two-model problem or to the
six three-model problems. The younger group responded correctly to 41
percent of the one-model problems, three subjects (16 percent) responded
correctly to the two-model problem, and not a single subject responded
correctly to the three-model problems. This seems to conform to Johnson-
Laird's predictions about relative difficulty, and perhaps supports the idea
of locating difficulty in memory load (since children may have deficits com-
pared to adults in that respect).
How plausible is the mental models approach? In "Mental Muddles"
Rips makes some telling criticisms. The central question is whether mental
models can replace mental inference rules. Mental modelers get rid of tra-
ditional inference rules, Rips charges, only at the expense of positing some
rules of their own. Procedures are necessary to specify how the models
should be used to represent the premises, how these premise models should
be combined, how one combined model should be transformed into an-
other, and how potential conclusions should be evaluated for consistency
with a combined model. The rules that make use of elements in a model
have to respect properties that seem like those of logical syntax. The equal
sign, for example, has to be interpreted in a way that enforces the idea that
the two tokens refer to the same object. So manipulation of mental models
is not fundamentally different from manipulation of mental propositions. It
seems, then, that the rules that operate on the models must be inference
rules. If so, the contrast Johnson-Laird sets up between mental logic and
mental models collapses.
It might be replied that some of this criticism applies specifically to
Johnson-Laird's version of the mental models approach, but may notgener-
alize to all variants of this approach. The underlying theme of the mental
models approach is intuitively very attractive, even if Johnson-Laird's par-
ticular version seems problematic. Still, the sort of challenge Rips poses
would seem to be germane to any specimen of the mental models approach.
Rips also raises some technical problems about Johnson-Laird's models,
which cast doubt on his principles for counting the number of models re-
quired for a given premise pair. But I will not go into this matter. It should
also be noted that Johnson-Lairdonly applies his leitmotif to sy2bgistic rea-
soning. It is not clear how it would apply to sentential logic or to prehcate

logic generally,
13.4. Justif;cation and Natural Logic

Given the research of Rips, Johnson-Laird, and many others in the field,
what are the grounds for either optimism or pessimism about our native log-
ical endowments? What appears to be the extent of human ability to have
wtijied beliefs about logical subject matter, even without formal training
in logic? What precisely are the operations that can yield such justifiedness?
Even a first pass at these questions will require a critical assessment of the
research I have summarized in the previous section, and related work as
well.
I will start with Rips's approach and then proceed (more briefly) to John-
son-Laird's. On the surface Rips's model-if it is correct--offers substantial
ground for optimism. It depicts the human cognizer as possessing a set of
inference routines that all embody sound inference rules, such rules as
modus ponens, DeMorgan's law, disjunctive syllogism, and-elimination,
and-introduction, or-introduction, law of excluded middle, and the like.
Though Rips does not suggest that these routines suffice-even in princi-
ple--to solve all (propositional) logic problems, the soundness of the rules
apparently guarantees perfect reliability. This would be just the sort of
characteristic that can confer justifiedness.
But there are several questions and doubts to be raised. First, do the infer-
ence routines really guarantee reliability? How can that be so, since the
model's predictions came quite close to the subjects' observed perform-
ances, and yet these performances involved rather high error rates? How is
this possible, and what are the implications for justifiedness?
Second, is a Rips-style model really correct? Do people have only sound
inference routines? How is that compatible with other evidence of psychol-
ogists (and logic teachers) that untrained subjects are prone to commit vari-
ous sorts of fallacies and mistakes? What do these fallacies and mistakes
imply about the prospects for justifiedness?
Third, Rips's model would be grounds for optimism about our constitu-
tional logical endowments only if the model really depicts people's native
processes. But it is not clear that it does. (Nor does Rips explicitly claim that
it does.)
Concerning the third issue, notice that even if subjects do use the sorts of
routines featured in Rips's model, these routines are not obviously 'native'
or 'basic'. They could be acquired 'methods' that subjects have evolved or
picked up. To be sure, the shbjects in Rips's experiments had had no formal
training in logic (that is, no logic coursework). Still, the experiments were
done on college students (typically at good undergraduate institutions, like

the University of Chicago), and any college student has many layers of ac-
culturation imbedded in his or her thinking. Examples of proof procedures
would have been encountered in mathematics courses, even in high school
geometry courses. And smatterings of logic inevitably appear in many
books and lectures outside of so-called logic courses. So while Rips's em-
pirical data-including the protocol data and memory-for-proofs data-
may indeed support his model, they leave it entirely open whether the op-
erations in question are natural inferences. But only the properties of
natural inference concern us here. Although this point is extremely impor-
tant for our purposes, let us waive it in the remainder of the discussion,
turning to the other issues concerning Rips's theory and results.
First, what are we to make of the appreciable number of errors that sub-
jects in Rips's experiment made? How is that compatible with the soundness
of the routines he postulates? In the experiment to which I alluded in sec-
tion 13.3, the subjects correctly judged valid arguments to be valid on only
about 51 percent of the trials; so their responses on valid arguments were
wrong about 49 percent of the time. On problems involving invalid argu-
ments, their responses were wrong about 23 percent of the time. These are
not very high truth ratios, and certainly not perfect ratios.
The main explanation is that soundness of the routines only guarantees
that whenever an ostensible proof is constructed (of the conclusion from the
premises), the argument is genuinely valid. So whenever a subject judges an
argument valid because of a proof he has constructed, the judgment will be
correct. But there are also occasions when a subject fails to construct a
proof, either because the argument is invalid, and hence there is no proof of
the conclusion from the premises, or because his routines have not suc-
ceeded in devising a proof despite the argument's validity. With a proof
unavailable, however, it seems that the subject must guess at the argument's
validity. This is where error possibilities creep in.
An initial point to be made, then, is that subjects performing these sorts
of tasks do not always use the same sequences of operations. So we should
not assume that the overall performance of the subjects as a group, or even
the overall performance of any single subject, represents the truth ratio of a
single process (or sequence of operations). Furthermore, it is compatible
with the prevalence of erroneous responses that certain operation sequences
are perfectly reliable; errors that occur may be wholly assignable to other
operation sequences. In particular, any operation sequence that consists of
(a) using one of the inference routines in ANDS, (b) producing thereby a
proof of the target conclusion from the target premises, and (c) inferring
therefrom a belief in the argument's validity, is a perfectly reliable opera-
tion sequence or process. This is just the sort of process that would presum-
ably be licensed by a right J-rule system. By contrast, any operation
Deductive Rearoning 295
sequence whose final stage incorporates a mere guess presumably would not
be very reliable and would not be licensed by a right J-rule system.
By 'guess', I here mean a process that fixes a belief. There is another sense
of 'guess' that simply involves the choice of a verbal response, unaccompa-
nied by a genuine belief. Many of the subjects' responses in Rips's experi-
ment may have been guesses in this second sense, because subjects were
forced to respond either 'valid' or 'invalid'. But since the present discussion
focuses on processes that yield justified or unjustified beliefs, I am only con-
cerned with guesses in the sense that triggers beliefs. (Notice that if many of
the responses were just guesses in the second sense-mere verbal responses
not registering genuine belief-then the number of subjects' false beliefs
was not as high as the data suggest.)
It might be argued that when a subject tries to construct a proof, fails,
and infers that the argument is invalid, this last inference is not a mere
guess. It is an inductive inference, perhaps of an 'explanatory' variety. It
might run: 'I tried a variety of ways of constructing a proof, but none of
them worked. The best explanation of this failure is that there is no proof.
Therefore, the argument is not valid'. This is indeed one way a subject
might reason. It is not obvious, though, whether this would be an approved
process of reasoning. What it shows, however, is that even in Rips's model
there might be justified beliefs in propositions concerning logical subject
matter that are justified by inductive rather than deductive inference pro-
cesses. Actually, it seems to be a consequence of Rips's model that the only
way in which a person can justifiably believe in an argument's invalidity is
by this sort of inductive inference. The ANDS model offers no way of prov-
ing invalidity.
I can now summarize the implications of Rips's model as follows. If this
model is right-and right about our basic reasoning processes-there
clearly are some operation sequences used to form beliefs in argument va-
lidity that would be licensed by a right J-rule system. Specifically, where a
person constructs a proof by means of any sequence of (officially designated) ,
routines incorporated in Rips's model, and infers validity of the target ar-
gument from this proof, then this belief is presumably justified. But it leaves
open the possibility that people have no available operations for judging
argument invalidity that would be permitted by a right J-rule system. Or, if
there are such operations, they must have a significant inductive compo-
nent. This last possibility is especially interesting. Could it be that for some
judgments of logic, people have to use inductive processes to arrive at justi-
fied beliefs? Shades of John Stuart Mill!
We come next to the plausibility of a Rips-style model. How, in particu-
lar, does it square with (apparent) evidence of fallacies and mistakes of vari-
ous kinds? Consider fallacies that are sufficientlyprevalent to have received
labels, such as 'denying the antecedent' and 'affirming the consequent'.
296 Deductive Rearoning
Some data on this matter are found in a study by Rips himself and a col-
league.16 Subjects were asked to evaluate eight types of "conditional syllo-
gisms," shown in Table 13.1. Subjects were asked to judge whether the
conclusion was "always" true, "sometimes" true, or "never" true, given the
premises. As the figure shows, subjects could apply modus ponens quite
successfully. But there was considerable evidence of tendencies to 'deny the
Table 13.1 Percent of total responses for eight

types of conditional arguments.
Argument Always Sometimes Never
8. P > Q
-Q 57" 39 4
. :. -P
Source: L. J. Rips and S. L. Marcus, "Supposition and
the Analysis of Conditional Sentences," in M. A. Just
and P. A. Carpenter, eds., Cognitive Processes in
Comprehension (Hillsdale, N .J.: Erlbaum, 1977).
a. The correct response.
antecedent' and 'affirm the consequent'. In problem 4, 21 percent mis-

takenly said that the conclusion would always be true; and in problem 5,23
percent erroneously said that its conclusion would always be true. These
cases invite the query: What processes were those subjects undergoing that
led them to their incorrect answers? How are those processes to be ex-
plained on Rips's approach? Perhaps answers are to be sought in more
'primitive' operations, which could explain why subjects go wrong in these
ways, as well as why they go right in many other cases.
Notice that subjects also had considerable trouble with modus tollens, a
rather elementary inference form (from the vantage point of an expert). In
problem 8, 43 percent of the subjects failed to see that the conclusion would
always be true, given those premises. Of course, failure to possess modus
tollens as an inferential rule, or routine, is not a fallacy However, it again
invites the conjecture: perhaps there are more primitive operations that
underlie the fact that modus ponens, say, is so easily grasped whereas modus
tollens is often ignored.
Attempts have been made to explain some of the fallacious propensities
in terms of 'misunderstandings' of the conditional. Experiments by J. E.
Taplin and H. Staudenmayer suggest that many subjects interpret the con-
ditional as being what logicians would call the 'bi~onditional'.~'If so, they
would not really be making a fallacious inference. But the failure to apply
modus tollens cannot be explained away in this fashion, since it is a legiti-
mate principle even if the conditional is interpreted as a biconditional.
In addition to the fallacies mentioned above, fallacies are frequently
made on classical syllogisms. The following invalid syllogisms, for example,
are often accepted as valid:
Some A's are B's.
Some B's are C's.
Therefore, some A's are C's.
No A's are B's.
No B's are C's.
Therefore, no A's are C's.
To account for this pattern of errors, R. S. Woodworth and S. B. Sells pro-
posed the "atmosphere hypothesis," which stated that the logical terms-
'some', 'all', 'no', and 'notY-createan atmosphere that predisposes subjects
to accept conclusions with the same terms.18 For example, subjects are pre-
disposed to accept a positive conclusion to positive premises and a negative
conclusion to negative premises. When the premises are mixed, subjects
prefer a negative conclusion. Although the atmosphere hypothesis captures
some of the trends in the data on syllogistic reasoning, it does not represent
the whoie story. It certainly raises doubts, though, about the notion that
people's inference routines embody only sound inference rules.
Another well-known case of logical errors is found in Wason's "selection
Figure 13.1
task."lQIn the central experiment from this research, four cards were shown
to subjects (see Figure 13.1).Subjects were told that a letter appeared on
one side of each card and a number on the other. They were asked to con-
sider the following rule, which referred only to these four cards: "If a card
has a vowel on one side, then it has an even number on the other side." The
task was to decide exactly which cards had to be turned over to determine
the correctness or incorrectness of the rule.
Forty-six percent of the subjects elected to turn over the E and the 4,
which is a wrong combination of choices. The E card does have to be turned
over, but the 4 card does not, since neither a vowel nor a consonant on the
other side would falsify the rule. Only 4 percent elected to turn over the E
and the 7, which is the correct answer. (An odd number behind the E or a
vowel behind the 7 would falsify the rule.) Another 33 percent of the sub-
jects elected to turn over the E only. The remaining 17 percent made other
incorrect choices.
There are two main types of error in this task. First, subjects often turn
over the 4, which is not necessary. Second, there is almost total failure to
appreciate the necessity of turning over the 7. Admittedly, not all of these
mistakes need be attributed to application of unsound inference routines;
they might be attributed to omission of sound routines. But they again cause
one to wonder whether all the manipulations subjects are performing can
be accounted for by the kind of routine repertoire that Rips postulates.
Let us turn briefly now to Johnson-Laird. It is hard to compare his theory
to that of Rips for several reasons. First, it deals primarily with syllogistic
reasoning, whereas Rips's model deals with propositional reasoning. Sec-
ond, the tasks Johnson-Laird gave his subjects were Wh-tasks ('What con-
clusions follows from these premises?), not Y/N tasks ('Is this argument
valid?). Third, Johnson-Laird does not really present any routines by which
subjects are alleged (a) to construct 'models', or 'interpretations', of the
premises and contemplated conclusion, and (b) to check whether they have
considered all relevant models. His description of the operation sequences
that subjects undergo is therefore quite sketchy.
Given this sketchiness, my treatment of the theory must be a bit specula-
tive. It seems, though, as if nothing corresponding to deduction-framed
routines are part of the process envisaged by the theory. In particular, first,
nothing of that sort seems to be at work when subjects attempt to construct
Deductiue Reasoning 299
interpretations of the premises. Second, nothing of that sort seems to be at .

work when subjects decide whether they have constructed all the alterna-
tive interpretations that are possible for the premises in question. So far,
then, there is nothing corresponding to the soundness of Rips's inference
routines to assure high reliability. Of course, processes need not be sound (in
the technical sense) in order to be sufficiently reliable for justificational
purposes. So we cannot conclude that Johnson-Laird's approach, if correct,
would preclude justifiedness. We need to know more details about the pro-
cesses it postulates.
This much seems clear, however. Johnson-Laird's picture of the processes
used in deductive reasoning tasks depicts them as very similar to processes
used in other kinds of cognitive tasks, such as natural language understand-
ing. In all such tasks there are operations of construction and rearrange-
ment of mental 'tokens', which represent relevant contents of the task. (Of
course, the sharing of this feature does not imply that the operations are
otherwise exactly alike. But Johnson-Laird's main theme is this similarity.)
Thus, on Johnson-Laird's approach, the processes involved in deductive
tasks are not at all m i generis.
Given the many questions I have raised, no definite conclusions can rea-
sonably be drawn at this point about the justificational potential of the basic
processes we use in deductive reasoning tasks. Of course, this inconclusive
state of affairs is no cause for embarrassment in cognitive psychology. The
questions are, obviously, intrinsically very difficult. What is clear is that pri-
mary epistemology needs answers to these questions, and that, at least in
principle, they can be answered by cognitive psychology.
13.5. The Dichotomy in Epistemology versus the Unity of Cognition

In this section I draw some implications from our discussion for a long-
standing dichotomy in the history of epistemology. A sharp demarcation has ,
traditionally been drawn between 'truths of reason' and 'matters of fact'.

Allegedly, the epistemologies of these two realms are significantly different.
In particular, the epistemology of logic and mathematics is said to differ
sharply from the epistemology of 'empirical' subject matters. I want to
challenge this sharp dichotomy from the vantage point of this chapter's ma-
terial. Quine, of course, has famously challenged the same dichotomy,20but
the grounds of my challenge are quite different from Quine's.
A number of labels are used-and sometimes uncritically assimilated--in
addressing this dichotomy. These labels come in pairs: 'a priori' and 'a post-
eriori'; 'empirical' and 'nonempirical'; 'necessary' and 'contingent'; and 'an-
alytic' and 'synthetic'. Several of these distinctions are not realIy
episterrwlogical distinctions. The distinction between the necessary and the
contingent is metaphysical; and the distinction between the analytic and

the synthetic is semantic. However, for the sake of argument (at least), I
shall grant the legitimacy of these distinctions. I will even concede, for the
sake of argument, that statements (or facts) of logic differ from 'empirical'
statements on these metaphysical or semantic dimensions. But even if it is
conceded that truths of logic are necessary or analytic truths, it is an open
question what follows from this about the epistemology of logic.
Actually, I do not wish to deny all epistemological differences between
questions of logic and questions of (what is normally called) 'empirical sci-
ence'. After all, I too have presupposed some such distinction in claiming
that epistemology should invoke the help of empirical sciences. What
would be the significance of this claim if there were no contrast between
the empirical sciences and other disciplines? It may be, however, that the
differences between disciplines are mainly differences in their secondary
epistemologies, in the methods appropriate to each. At the level of primary
epistemology, they may still have much in common. This weaker thesis is
the one I will defend. Even this weaker position, however, parts company
with the traditional style of dichotomizing epistemology.
One of the traditional claims is that truths of reason and matters of fact
differ on the dimension of certainty: unlike matters of fact, truths of reason
can allegedly be known with certainty. Quine challenges this principle of
differentiation by first equating certainty with immunity from revision and
then denying that any statements are so immune. I agree with the certainty
challenge, but not because I equate certainty with immunity from revision.
(I would also reject the equation of 'analytic statement' with 'statement im-
mune from revision'; but that is another matter.)
As an epistemic (rather than a doxastic) notion, certainty is best under-
stood as the highest possible level of justifiedness or warrant. The theory of
justifiedness, as I have formulated it, does not admit leveks, or grades, of jus-
tifiedness. But the spirit of the theory lends itself to such an extension. A
belief can be regarded as certain if and only if it is caused by perfectly reli-
able-infallible-cognitive processes (or infallible processes together with
infallible methods). Let us ask, then, whether logic belongs in a separate
epistemic category because it is knowable with certainty.
The suggestion that truths of logic are so knowable could derive plausibil-
ity from the idea that there are mechanical methods for determining truths
of logic, since mechanical methods are infallible. But, of course, Church's
Theorem shows that there are not mechanical methods for determining in-
validity in general quantification theory. Moreover, as we have seen (in sec-
tion 13.,1),complexity theory suggests that many other problems in logic are
not computationally tractable, that is, are not solvable by mechanical pro-
cedures in a human being's lifetime.
Another point is that although mechanical methods may be 'intrinsically'
infallible, in the sense that if you use a mechanical procedure or algorithm

correctly you cannot go wrong, there is still the cognitive task of using it
correctly. This involves the use of cognitive processes, and it is unlikely that
any of these are infallible. You can misperceive marks and misremember
what has occurred in previous steps. You can also mistake a putative algo-
rithm for a genuine one; so your belief that you are proceeding, or have
proceeded, algorithmically-by means of a method that yields correct re-
sults-is liable to error. According to my theory, the status of certainty is
only attainable by the use of infallible cognitive processes (and perhaps, if
methods are necessary, infallible methods as well). Yet infallible cognitive
processes are no more available for the subject matter of logic than for other
subject matters.
This point is not new. Hume wrote: "In all demonstrative sciences the
rules are certain and infallible; but when we apply them, our fallible and
uncertain faculties are very apt to depart from them, and fall into error."21
Similarly, Descartes noted the problem of memory in connection with rea-
~ o n i n ~But
. ' ~ while the point is old, it is nonetheless germane.
Can one distinguish the epistemic status of logic from that of other sub-
ject matters by appealing to the greater certainty, or reliability, of logic-
linked processes-even if they are not perfectly reliable? My discussions in
the preceding sections make this very dubious. First, there is the evidence
of certain fallacious inference propensities. Second, at least Johnson-Laird's
picture of the processes deployed in logic tasks do not make them appear
notably reliable. They are not obviously more reliable than, say, perceptual
processes.
Let us turn now to the distinction between the a priori and the a poster-
iori. This distinction has traditionally been applied to two different kinds of
objects: truths and (pieces of) knowledge. Philosophers sometimes speak of
'a priori truths', and they sometimes speak of things being 'known a priori'.
It is generally conceded, though, that knowledge is the primary locus of a
priority. Truths are dubbed a priori only in a derivative sense: when it is
possible to know them a priori.
The standard explication of a priom' knowledge is knowledge that does
not involve, or depend upon, perception, or 'experience'. However, other
baggage is often added to the term 'a priori'. In particular, it is often taken
to entail certainty. If so, we have already had grounds for denying that logi-
cal truths are (generally) knowable a priori. But let us unload this excess
baggage (certainty) and focus on the alleged irrelevswe of perception to
logic.
Here is an argument for disputing the general knowability of logic with-
out perception. Note first that in actual practice logicians learn new logical
truths by reading the proofs others have constructed, and constructing their
own proofs with pencil and paper. These activities involve perception. The
a priorist would contend, however, that perceptual reliance on inscriptions

is in principle dispensable. One could, in principle, frame all the relevant
representations in the mind. But is this true? Do human cognitive capacities
enable one to construct, or follow, a complex proof in the mind with suffi-
cient reliability to qualify for knowledge? That is doubtful. It is universally
acknowledged that short-term, or working, memory is limited in capacity.
When this capacity is exceeded, error rates increase sharply. It seems likely,
therefore, that many phases (even single steps) of a truly complex proof will
involve more material than can reliably be maintained in working memory.
Therefore, a logical truth whose proof involves that degree of complexity
can only be known with the help of external inscriptions, hence, only with
the help of perception.23
Would it be too hasty to conclude that truths of logic are not generally
knowable a priori? Could one distinguish the realms of the a priori and a
posteriori by saying that propositions known a posteriori are, or can be,
known by perceiving the known facts themselves, not just inscriptions re-
lated to them? And could one say that, by contrast, a priori knowledge only
utilizes perception as an 'aidy-that there is no perception of the target
truths themse~ves?~~ This obviously will not work. Some facts known a post-
eriori--for example, facts involving electrons and nomological facts gen-
erally--cannot themselves be perceived. So this approach would not draw
the line where it is intended to be drawn.
Could one say that truths of reason are sui generis inasmuch as they coukl
be known by beings (such as God) without any perceptual faculties at all?
This criterion also would not draw the demarcation properly. Perhaps God
could know contingent facts without perception-by some sort of nonsen-
suous intuition or clairvoyance.
The prospect of distinguishing the two realms in terms of perceptual
knowability seems dim. There is no need to deny, though, that simple logi-
cal truths may be humanly knowable without any perception, and this trait
may indeed contrast with contingent truths about external matters. Even
granting this point, though, can it sustain the a priorist's hope? No. The tra-
ditional hope of the a priorist is to draw an epistemological dichotomy be-
tween the entire realm of logic and mathematics, on the one hand, and the
realm of contingent fact, on the other. This hope cannot be sustained by the
nonperceptual knowability of a small subset of logical truths.
Notice also that perception is only one of many cognitive faculties. Even
if it is granted that the epistemology of logic does not encourage the same
role for perception as does the epistemology of contingent matters, there is
extensiv-e sharing of other cognitive operations (or types of operations.) This
sharing raises the question, Why harp so persistently on differences con-
cerning the role of perception while downplaying the shared utilization of
other faculties?
It is important in this connection to bear in mind the central place of

problem solving, or question answering, in my account of epistemology.
The epistemology of any 'realm' X is a matter not just of how one is justified
in believing propositions concerning X, but of how one finds answers-and
proofs of answers--to questions concerning X. In the realm of logic, for ex-
ample, it is not just a matter of how one can be justified in believing that
something is a proof of a specified conclusion from specified premises; it is
also a matter of how one goes about searching for such a proof (without
knowing beforehand whether there is one). Similarly, the epistemology of
logic consists not just of what it takes to know that a specified conclusion
follows from specified premises, but also of how one goes about constructing
propositions that answer the question, Which conclusions follow from those
premises?
Let us examine features that the epistemology of logic shares with the
epistemology of 'empirical' subject matters. Although I have stressed the
proof aspects of Rips's theory, which may be distinctive of logical or mathe-
matical subject matter, even Rips's model has significant features that
would be shared by processes used on empirical topics. Recall first the goal-
directed routines and the control system that specifies an order in which
routines are to be tested for applicability. Many of the posited routines,
especially the 'backward' routines, are heuristics that constitute strategies
likely to pay off. Precisely analogous heuristics can be expected for empiri-
cal domains as well. Of course, these would not be proof-construction
heuristics; but they would be heuristics for finding evidence that bears on
the question at hand.
The very existence of counterpart heuristics in different domains strongly
suggests the possibility that the specific heuristics Rips postulates are not
native. They may be acquired in the way other domain-specific heuristics
are acquired. This topic of the acquisition of heuristics (or 'condition-ac-
tion' pairs) will be treated in Chapter 17, where I will discuss 'production
systems' generally and the acquisition of new productions in particular. It is
reasonable to expect that while the details of heuristics differ from domain
to domain, the underlying process of heuristic acquisition (and heuristic re-
tention) is the same across domains. If this is correct, then both the psychol-
ogy of logic and the psychology of empirical subject matters have roots in
the same (second-order) processes of heuristic learning.
Let us look next at the role of memory in logic problem solving. Recall
that Rips assumes that the postulated routines are not always available and
do not have equal availability probabilities. Unavailability of a routine may
be due to several sources, he says. For example, one may simply fail to re-
trieve it from memory. Or one may fail to recognize its applicability. These
same sorts of cognitive obstacles doubtless crop up in empirical domains as
well. A cognizer is always facing the problem of trying to retrieve (from
LTM) the most relevant heuristics, rules of thumb, or procedures. Presum-
ably, the factors that facilitate or inhibit retrieval in the logic domain are
the same as those in the other domains. For instance, if one has recently
used a given heuristic to solve an analogous problem, one may be more
likely to retrieve that heuristic in the present task. This role of analogy will
pera ate in all domains of inquiry, logic included.
Finally, let us turn to the task of constructing possible answers to logic
questions. This is a nontrivial task when the question is, What conclusion (or
2onclusions) follows from these premises? In Johnson-Laird's experiments
le found that one major factor affecting syllogistic inference (previously
lnmentioned here) is the figure of the premises. This "figural effect" is a
~ i a stowards certain forms of conclusion. In particular, people are much
nore prone to infer an A-C conclusion from A-B and B-C premises, or to
nfer a C-A conclusion from B-A and C-B premises, than to draw inferences
n other figures.25Johnson-Laird conjectures that this is due to basic pro-
:esses of working memory, that working memory operates on a "first in,
irst out" basis. It is easier, for example, to recall a list of digits in the order
n which they were presented-the first digit recalled first, the second digit
ecalled second, and so on-than in the opposite order. If this is right, then
ypothesis formation in logic is a special case of much more general cogni-
ive operations.
We saw in section 13.4 that the main theme of Johnson-Laird's theory-
onstruction and manipulation of models-posits processes that are used for
11 sorts of cognitive tasks. If correct, this would certainly confirm the point
eing stressed in this section. Needless to say, the entire topic deserves
lller treatment. The exact extent of shred cognitive operations in logic
rsks and in empirical tasks requires more experimental and theoretical
ladework. I have illustrated the idea, however, that the cognitive ingre-
ients of problem solving, at a deep level of analysis, are fundamentally sim-
ar across different domains. Hence, the classical attempt to dichotomize
?istemology is fundamentally misguided, another example of the potential
npact of psychological work on epistemological theory.
chapter 14
Probability Judgments
14.1. Whut Can Psychologists Tell Us?

In the last decade or so, psychological research on probability judgments
has led to the claim that procedures commonly used in arriving at these
judgments are counternormative. Several writers formulate the matter by
suggesting that people's intuitive processes of probability judgment are 'ir-
rational'. This chapter addresses the findings of this research and their eval-
uative implications.
The most important and influential research in this area has been done by
Amos Tversky and Daniel Kahneman, who are themselves rather cautious
in phrasing their conclusions. However, their language has intimations of
counternormativeness. They speak of the "biases" of certain cognitive pro-
cesses, of cognitive "illusions," of large and systematic "errors," and of
"violations" of rules or laws (of statistics or probability theory) that they os-
tensibly regard as normatively correct.' Kahneman and Tversky also say
that "for anyone who would wish to view man as a reasonable intuitive stat-
istician, [their] results are discouraging."2 Expositors of this and related re-
search, such as Richard Nisbett and Lee Ross, make similar or stronger
claims. They write of human failure to employ normatively correct in-
ferential strategies;3 of "departures from normative standards of infer-
e n ~ e " and
; ~ of "profound systematic, and fundamental errors in judgments
and inference^."^ Nisbett and E. Borgida write of "bleak implications for
human rationality."' P. Slovic, B. Fischoff, and S. Lichtenstein assert that
"people systematically violate principles of rational decision-making when
judging probabilities, making predictions, or otherwise attempting to cope
with probabilistic taskseH7Furthermore, in commentaries on L. Jonathan
Cohen's paper "Can Human Irrationality Be Experimentally Demon-
s t r a t e d ~none
~ of these researchers dispute Cohen's assertion that they
view their research results as signs of human irrationality. Apparently, that
interpretation is close to the mark.
306 Probability Judgments
What can psychologists do, or tell us, that concerns rationality or other
dimensions of intellectual appraisal? I have pursued this question, of course,
through much of this book. At this point it may be helpful to approach it
from a slightly different angle, more directly dovetailing with the work of
people like Tversky and Kahneman.
First of all, the psychologist can arrange experiments in which subjects
answer certain questions, and he can report these responses. Second, the
psychologist can often say what proportion of these responses are true or
false; and presumably he can say whether the responses are consistent or co-
herent. It is not obvious that this second sort of issue is always within the
psychologist's sphere of competence. In perceptual experiments the psy-
chologist tells us the true properties of the presented stimuli; he is pre-
sumed to be authoritative (enough) on that matter. In other kinds of
experimental tasks, though, the truth or falsity of the subjects' responses
may be more controversial. For example, whether specific probability judg-
ments are true or false is entangled in controversial issues about the inter-
pretation of probability sentences. So what a psychologist might say on this
score could be disputed by others.
Third, the psychologist can offer hypotheses about the cognitive pro-
cesses that underlie the observed responses. That is, he can theorize about
the procedures, mechanisms, strategies, or heuristics that must be leading
the subjects to respond as they do. Of course, any proffered explanatory hy-
pothesis is open to empirical dispute. How well does it explain the evi-
dence? What competing model might explain the evidence as well, or
better?
Fourth, there is the question of the normative status of the subjects' re-
sponses; or the normative status of their using certain procedures, and fail-
ing to use others, in arriving at their responses. This question, it seems to
me, falls outside the domain of psychology, narrowly construed. It is not the
job of empirical science to make normative judgments (whether ethical,
aesthetic, or epistemological). When it comes to epistemic normative judg-
ments, this is the task of epistemology.
However, if certain epistemic rules or principles are agreed upon, then
the psychologist may be in a position to tell us whether or not human beings
conform with these rules or violate them, and with what frequency or in
what circumstances. But if there is any dispute about which normative
principles are correct, this dispute goes outside the realm of psychology
per se.
Of course, I have been contending that principles of primary epistemol-
ogy are partly determined by psychological facts. So on my view the psy-
chologist can help epistemology ascertain correct normative principles.
And the very sorts of experiments in question may help do this. But this
flows from the criteria of rightness (at least J-rule rightness) that I have de-
Probability Judgments 307
fended, and the defense of such criteria falls outside psychology altogether.
In the present context, then, there are three points I wish to stress about
the psychologist's role. First, it is certainly possible that the psychologist
might tell us things about human cognitive performances which imply (to-
gether with correct normative principles) that people's beliefs or probabil-
ity estimates are frequently, even characteristically, unjustified, irrational,
or otherwise counternormative. (Here I disagree sharply with Cohen.) But
second, the psychologist is not necessarily authoritative on what the correct
normative principles are. So if he makes any claims about such principles,
those are open to dispute. Third, since I have argued that epistemic rules, or
norms, should specify acceptable (and unacceptable) cognitive processes,
the psychologist can help arrive at such rules by trying to uncover what
sorts of cognitive processes are available to the human organism.
14.2. Empirical Findings

Turning now to specific empirical findings and tendered explanatory hy-
potheses in the area of probability judgments, it seems there are two main
classes of results and explanations. First, people persistently show ignorance
of facts of probability theory, or neglect these facts in the course of various
cognitive tasks, even when they are highly relevant. It is not just untrained
subjects who do this; so do people with considerable formal training in
probability and statistics. Second, it is hypothesized that people use a differ-
ent set of intuitive strategies, or heuristics, to make probability judgments,
rather than the probability calculus. The one most commonly alleged is the
'representativeness' heuristic. Let us examine these results in detail.
Concerning ignorance or misconceptions of probability and statistics,
Tversky and Kahneman cite the following phenomena.g (1)People are often
insensitive to prior probabilities. When asked to estimate the probability
that Steve is a librarian rather than a farmer, they tend to rely on the simi-
larity of Steve's personality profile to the stereotypes of farmers and librari-
ans, while ignoring 'base-rate' information, that is, information about the
number of farmers and librarians in the population. (2) People tend to be
insensitive to sample size. They do not realize that a smaller sample is more
likely to stray farther from a population statistic than a larger sample. (3)
People seem to expect that events generated by a random process will rep-
resent that randomness. For example, people regard the sequence of coin
tosses H-T-H-T-T-H to be more likely than the sequence H-H-H-T-T-T and
more likely than H-H-H-H-T-H, even though all three are equiprobable. (4)
People are insensitive to the phenomenon of regression toward the mean,
despite its ubiquity. Suppose, for example, that a large group of children
have been examined on two equivalent versions of an aptitude test. If one
selects ten children from among those who did best on one of the two ver-
sions, he will usually find their performance on the second version to be

somewhat disappointing. Conversely, if one selects ten children from
among those who did worst on one version, they will be found, on the aver-
age, to do somewhat better on the other version. But people do not expect
such regression in many contexts where it is bound to occur; and when they
recognize regression, they often invent spurious causal explanations for it,
claim Kahneman and Tversky.
In "Extensional versus Intuitive Reasoning" Tversky and Kahneman
study people's insensitivity to a basic law of probability theory: the con-
junction rule. This rule states that a conjunctive event A and B can never be
more probable than either of its conjuncts. This is because the set of possi-
bilities associated with the conjunction must always be included within the
set of possibilities associated with each conjunct. The probability that it will
be windy and rainy tomorrow cannot be greater than the probability that it
will be windy, nor greater than the probability that it will be rainy. How-
ever, people's intuitive judgments of probability often violate the conjunc-
tion rule. They frequently make groups of probability estimates in which
the probability of a conjunction is rated higher than the probability of one
of its conjuncts.
Here is a partial description of a typical experiment in which violations
of the conjunction rule were commonly found. Subjects were given a per-
sonality sketch of a fictitious individual, Linda, followed by a set of possible
occupations and avocations she might pursue.
Linda is thirty-one years old, single, outspoken, and very bright. She majored
in philosophy. As a student, she was deeply concerned with issues of discrim-
ination and social justice, and also participated in antinuclear demonstra-
tions.
Linda is a teacher in elementary school.
Linda works in a bookstore and takes Yoga classes.
Linda is active in the feminist movement.
Linda is a psychiatric social worker.
Linda is a member of the League of Women Voters.
Linda is a bank teller.
Linda is an insurance salesperson.
Linda is a bank teller and is active in the feminist movement.
Asked to rank the probabilities of Linda's current activities, including being

a bank teller (T), being active in the feminist movement (F),and their con-
junction (T and F), the great majority of subjects ranked the conjunction T
and F.as more probable than T, one of its conjuncts.
Tversky and Kahneman did three types of experimental tests of the con-
junction rule: indirect tests, direct-subtle tests, and direct-transparen t tests.
In the indirect tests one group of subjects evaluated the probability of the
conjunction and another group evaluated the probability of its constituents.

No subject compared a conjunction to its constituents. In the direct-subtle
tests subjects did make this comparison, but the inclusion relation between
the events was not emphasized. In the direct-transparent tests the format
highlighted the relation between the conjunction and its constituents. In all
three kinds of tests, even the transparent tests, violation of the conjunction
rule was very high. Using the Linda description with undergraduates at the
University of British Columbia, 85 percent of respondents in a transparent
test indicated that T and F was more probable than T.
To probe the depth of this phenomenon, Tversky and Kahneman asked
subjects to indicate which of the following two arguments they found more
convincing (hoping they would recognize the validity of the conjunction
rule when it was expressly formulated):
Argument 1. Linda is more likely to be a bank teller than she is to be a fem-
inist bank teller, because every feminist bank teller is a bank teller, but some
women bank tellers are not feminists, and Linda could be one of them.
Argument 2. Linda is more likely to be a feminist bank teller than she is
likely to be a bank teller, because she resembles an active feminist more than
she resembles a bank teller.
The majority of subjects (65 percent) chose the resemblance argument, (2),
over the extensional argument, (1).
Modest success in loosening the grip of the conjunction fallacy was
achieved by asking subjects to choose whether to bet on T or on T and F.
The subjects were given Linda's description with the following instruction:
If you could win ten dollars by betting on an event, which of the following
would you choose to bet on? (Check one.)
'Only' 56 percent of the responses in this task violated the conjunction rule.
Many of the respondents in the foregoing studies were statistically naive.
Does statistical education eradicate the fallacy? The answer is not entirely
clear. In one study indirect and direct-subtle tests were given not only to
naive respondents, but to informed and sophisticated respondents, the in-
formed group being first-year graduate students or professional students
who had had one or more courses in statistics, and the sophisticated group
being doctoral students in the decision science program at Stanford Univer-
sity, who had taken advanced courses in probability, statistics, and decision
theory. The sophisticated and informed groups did not perform (much) bet-
ter than the naive group. Eighty-nine percent of the naive subjects, 90 per-
cent of the informed subjects, and 85 percent of the sophisticated subjects
violated the conjunction rule.
In another study, however, statistical education did seem to have an ap-
preciable effect. Sixty-four graduate students of social sciences at the Uni-
versity of California, Berkeley and at Stanford University, all with credit for
several statistics courses, were given a transparent version of the Linda
problem. For the first time in this series of studies, less than half of the re-
spondents committed the fallacy-'only' 36 percent. But even this seems
fairly high, given the subjects' statistical sophistication.
The hypothesis proposed by Tversky and Kahneman to explain the find-
ings I have summarized is the 'representativeness' heuristic. Both in the
studies where subjects were insensitive to prior probabilities (base rates)
and in the conjunction fallacy studies, subjects' probability estimates were
apparently guided by the degree to which the person in question was repre-
sentative of, or similar to, the classes defined by the queried events or out-
comes. For example, the description of Linda presumably made her
representative of an active feminist and unrepresentative of a bank teller.
Presumably, she was also more representative of a feminist bank teller than
a bank teller. Indeed, in many experiments on the conjunction fallacy, sub-
jects were explicitly asked to rank various outcomes both for representa-
tiveness and for probability. The correlation between the mean ranks of
probability and representativeness were typically extremely high, often
greater than .95. Tversky and Kahneman also propose the representative-
ness heuristic as an explanation of misconceptions of chance. People's
(mis)estimates of coin toss likelihoods, they suggest, is due to their expecta-
tion that coin toss sequences will be representative of the random generat-
ing process.
What, exactly, is 'representativeness'? It turns out to be a rather multi-
faceted relation. In "Judgments of and by Representativeness" Tversky and
Kahneman discuss this relation in some detail.'' They say that representa-
tiveness is a relation between a process or model M and some instance or
event X, associated with that model. The model in question could be of a
person, a fair coin, or the world economy, and the respective outcomes
might be a comment, a sequence of heads and tails, or the present price of
gold. Normally, the sample or act is said to be representative of the popula-
tion or person. But these roles can be reversed, and we can say that the oc-
cupation of librarian is representative of a person.
They then distinguish four types of cases in which representativeness can
be invoked: (1)M is a class and X is a value of a variable defined in this class.
For example, the mean income of college professors might be representa-
tive of that class. (2) M is a class and X is an instance of that class. In this
case, John Updike might be regarded as a representative American writer.
(3) M is a class and X is a subset of M. The set of psychology students might
be moqe representative of the entire student body than the set of astronomy
students. (4) M is a (causal) system and X is a (possible) consequence. For
example, M can be the U.S.economy and X the rate of inflation, or M can
be a person and X an act performed by M (such as divorce, suicide, profes-
sional choice).
Probability Judgments 31 1
In early work Tversky and Kahneman suggested that people rely exclu-
sively on the similarity of a sample to its population parameters to make
probability judgments. Now they make a more moderate hypothesis: pre-
dictions and probability judgments are "highly sensitive" to representative-
ness, though not completely dominated by it. The magnitude of the appeal
to representativeness depends on the nature of the problem, the character-
istics of the design, the sophistication of the respondents, and the presence
of suggestive clues or other demand characteristics."
Given the multifacetedness of the representativeness relation, and the
multiple determinants of whether, or to what degree, people rely on it, it is
not clear that any very specific hypothesis concerning the reliance on repre-
sentativeness is currently proposed. Rather, it is a more general suggestion
that people 'often' rely on representativeness, or something of this sort. But
however frequently this reliance is made, there is the suggestion that any
appeal to representativeness in making probability judgments is counter-
normative. There is also the claim that failure to employ correct probabilis-
tic and statistical rules is counternormative. This normative claim is what I
address next.
14.3. Probability and Rationality

Since the evaluative term 'rational' has been introduced in this context, let
me start with that term. Many theorists view the 'violation' of probabilistic
or statistical rules as a specimen of irrationality. They think of the probabil-
ity calculus, along with logic, as a constituent of rationality. Cognitive ac-
tivity that violates the probability calculus-probabilistic 'incoherencey-is
deemed irrational. Violation of the conjunction rule would be a salient case
in point.
Tversky and Kahneman, 'it should be noted, avoid such an outright claim.
In their conjunction fallacy paper, at any rate, they offer a descriptive defi-
nition of 'fallacy' and eschew its use as an evaluative epithet. They define a
judgment as a 'faIlacyYwhen "most of the people who make it are disposed,
after suitable explanation, to accept the following propositions: (a) They
made a nontrivial error, which they would probably have repeated in simi-
lar problems, (b) the error was conceptual, not merely verbal or technical,
and (c) they should have known the correct answer or a procedure to find
it."12
However, even Tversky and Kahneman make evaluative judgments in
other passages and writings; and other theorists are certainly less circum-
spect or continent.
Now no one should dispute the notion that probability theory and statis-
tics can be extremely valuable intellectual tools and 'methods. Certainly I
do not wish to dispute this. But their connection with 'rationality' is a less
straightforward matter. In discussing this issue I will concentrate on proba-
bility theory, where there is a clear-cut and well-accepted mathematical
theory. Statistics, by contrast, is a large set of diverse methods and tech-

niques, the foundations of which are extremely controversial and problem-
atic. Of course, the application of probability theory to issues in (normative)
epistemology is also very controversial. But at least the mathematical the-
ory is compact and manageable.
In this section there are two main groups of questions I shall raise. (1)Is
probabilistic coherence an epistemic desideratum? Is incoherence a defect?
If so, why? How serious a defect is it? And (2) assuming that probabilistic
incoherence is a defect, is it a defect in rationality? In the next section I will
talk about justifiedness, though not exclusively in connection with coher-
ence and incoherence.
Before turning to these questions, we need to distinguish different ways of
interpreting probability judgments. Two main approaches will concern us
here: objective and subjective approaches. According to the objective ap-
proach, a probability is some sort of objective property of things, for exam-
ple, the chance that a certain type of event will occur. Theorists have
sought to explicate the notion of chance in various ways, especially in terms
of long-run relative frequencies, or propensities. I will not explore these at-
tempts in detail, nor the difficulties they encounter. But on the objectivist
approach, when a person says, for example, that the probability of rain to-
morrow is .75, he makes a statement about the objective chance of rain to-
morrow. Assuming the person is sincere, he expresses a belief (an ungraded
belief) in the proposition that the chance of rain tomorrow is .75. The prop-
osition that he believes is a proposition about chances, and such a proposi-
tion presumably has a truth-value (at least according to the objectivist).
Hence, the person's belief in the probabilistic proposition is either true or
false (and more or less accurate).
According to the subjectivist approach, a probability is just a state of
mind: a degree to which a person believes something, or the firmness of his
conviction. Thus, when a speaker says that the probability of rain tomorrow
is .75, he does not make a statement about the objective chance of rain. The
probability qualifier is not construed as appearing within the proposition he
believes. Rather, it characterizes the strength of the belief. What he be-
lieves-the content of his belief-is that it will rain tomorrow; the proba-
bility qualifier just registers the firmness, or confidence, with which he
believes it.
It is also possible to combine these two approaches. A person might have
subjective probabilities in objective probability propositions. That is, he
might have degrees of confidence in propositions about objective chances.
(This is one possible application of the notion of 'probabilities of
probabilities'.)
In addition to the objective and subjective interpretations, there are the
logical and epistemic interpretations of probability. The logical interpreta-
tion refers to degrees of inductive support, or confirmation, between pairs
Probability Iudgments 313
of propositions. In other words, it applies to arguments construed as se-

quences of premises and conclusions. The existence of such degrees of in-
ductive support is problematic. But, in any case, inductive support does not
figure prominently in the topic that presentIy concerns us. Thus, I will give
the notion of logical probability no further attention.
There are various ways of defining the notion of epistemic probability. In
the context of my theory an epistemic probability is best understood as a
probability that it is rational, or justified, for a cognizer to assign, given his
evidential situation. The concept of epistemic probability thereby invokes
some evaluative epistemic notion-either rationality or justifiedness-that
distinguishes it from the other probability notions considered thus far. How-
ever, such an evaluation must apply either to beliefs in objective probabil-
ity propositions or to subjective probability states. In a sense, then, no new
kind of probability is introduced by the phrase 'epistemic probability'. It is
just an evaluation of one of the two kinds of probability-related doxastic
states already presented: either a belief in a proposition about chance or a
subjective probability state.13 Although we are certainly interested in the
evaluative appraisal of such doxastic states, in so being we are not really in-
voking a further species of probability. So I will dispense with the phrase
'epistemic probability'.
We now have the relevant possible interpretations of probability judg-
ments. Before turning to the normative status of probabilistic incoherence,
though, it is instructive to address the normative status of logical inconsis-
tency, that is, the normative status of having a set of (ungraded) beliefs in
inconsistent propositions.
Is logical consistency an epistemic desideratum? Is inconsistency a de-
fect? If so, why? One oft-cited reason is that a person who believes incon-
sistent propositions would be entitled to infer any proposition whatever,
since anything follows from a contradiction. But it has not so far been es-
tablished that a right system of J-rules (or rationality rules) would entitle a
person to draw any or every logical consequence of what he antecedently
believes, even if he appreciates that it is a consequence. Hence, there need
be no cause for alarm on this score.
A second possible reason for regarding inconsistency as a defect is that
anyone who believes inconsistent propositions thereby believes at least one
falsehood. So, on any verific consequentialist approach to epistemic value,
there is at least some bad-making feature of being inconsistent. (Note that
verific consequentialism might also be needed to make full sense of the first
objection to inconsistency. What is so bad, we may ask, about inferring any
proposition whatever from your antecedent corpus of beliefs? Presumably,
it is bad because addition of randomly chosen propositions to one's belief
corpus is likely to yield lots of false beliefs, which is bad according to verific
consequentialism.)
But how bad is it, after all, to believe one falsehood? Not very, it would
seem. So inconsistency cannot thus far be convicted of being a serious epis-

temic defect. Of course, while inconsistency only guarantees one falsehood,
having an inconsistent set of beliefs will frequently involve much more than
one falsehood. If you have arrived at your inconsistent conclusions by infer-
ence, not only will (at least) one of those conclusions be false, but so, proba-
bly, will some of the premises that led to the conclusion or conclusions. (If
the inference was deductive, then at least one of those premises must be
false.) In this way inconsistency will tend to carry with it some larger quan-
tity of falsehoods, although it cannot be said, in general, how many. Thus,
inconsistency should not be dismissed lightly, though it is hard to say, on
verific grounds, just how nasty an infection it is.
Notice that consistency is not necessarily markedly superior to inconsis-
tency in terms of false-belief avoidance. This can be put as follows. Consis-
tency is, of course, a necessary condition for total avoidance of false belief.
But it is only a necessary condition, not a sufficient one. Having a consistent
set of beliefs does not guarantee total avoidance of false belief. In fact, it
does not guarantee avoidance of false belief at all. A set of consistent beliefs
can all be false.
These points have consequences for what one can and cannot expect to
achieve if one discovers that one's belief corpus is inconsistent and tries to
revise the corpus to rid it of this property. In general, it is not clear which of
an inconsistent set of beliefs should be abandoned. In eliminating some of
these beliefs, then, you cannot be assured you will eliminate false ones. Re-
moval of inconsistency eliminates the guarantee of errors, but it does not
guarantee the elimination of errors.
It is not so obvious, then, that elimination of inconsistency should be a
cardinal epistemic precept. Even if you discover an inconsistency, it may be
all right to keep these beliefs 'on hold', especially if you have no inkling at
the moment which of the beliefs are least defensible. Why not retain them
all, at least for the nonce, until a good reassessment of their relative merits
can be made? Maintenance of an inconsistent set of beliefs need not be irra-
tional, even when you are aware of the inconsistency. Many would argue
that you should at least reduce your degrees of confidence in the inconsis-
tent propositions, and perhaps this is right. But for the present we are deal-
ing with the binary, or ungraded, category of belief.
It is even clearer that inconsistency of beliefs need not be irrational when
this results from failure to notice the inconsistency. Failure to discover in-
consistency is not always a mark of irrationality. Axiomatizations of logical
theory by Frege and by Quine were both inconsistent in ways that did not
reveal themselves immediately to many who used each axiomatization ex-
tensively. Were all the people who failed to notice these inconsistencies ir-
rational? That seems like a misplaced epithet. Or consider mathematicians
who believed, before Russell's paradox, in the axiom of comprehension of
set theory, that for every formulable condition, there is a set whose mem-
bers are just the things fulfilling that condtion. Russell's paradox showed
that this axiom implies contradictions; but were those who believed in the
axiom prior to Russell's discovery of the paradox irrational? Surely not. In-
consistencies are often extremely difficult to detect; and failure to detect
them need not betoken irrationality.
An example from another domain may help solidify this point. The artist
M. C. Escher is famous for drawings that appear, on initial perusal, to de-
pict a consistent scene. But close inspection gradually reveals that there is
no possible coherent interpretation. Is a viewer who fails (even for some
time) to recognize the scene's incoherence irrational? I think not. Doubt-
less, such a viewer fails to notice something of importance, but this hardly
warrants the charge of irrationality.
None of these comments is intended to deny that a right set of epistemic
rules-either at the primary or the secondary level-might require at-
tempts to discover and remove inconsistencies in some contexts, perhaps
many contexts. Shaping your belief corpus with an eye toward consistency
is doubtless an effective way of increasing reliability. If one's perceptual
beliefs are usually true, then forming further beliefs consistent with these is
presumably a helpful way of working toward new beliefs with a high truth
ratio. My point has only been to put the seriousness of inconsistency into a
better, more measured, perspective.
Bearing these points about inconsistency in mind, let us turn now to
questions about probabilistic incoherence. Here we must deal separately
with objectivist and subjectivist interpretations of probability judgments.
On an objectivist interpretation probabilistic incoherence is just a special
case of inconsistency. Suppose you judge that the probability of its being
rainy tomorrow is .60, and that the probability of its being rainy and windy
tomorrow is .75. On the objectivist interpretation this means that you be-
lieve both Q, 'The chance of its being rainy tomorrow is .6OY,and Q", 'The
chance of its being rainy and windy tomorrow is .75.' Since chances obey .
the probability calculus (all objectivists assume this much), these proposi-
tions are inconsistent; they cannot both be true. (At least they are mathe-
matically if not logically inconsistent.) So probabilistic incoherence in this
case is just inconsistency. As we have seen, however, inconsistency is not
necessarily the most flagrant or serious epistemic defect. It is not clearly
epistemically objectionable in every case. Similarly cautious normative
conclusions should be made for probabilistic incoherence.
When we move to the subjectivist interpretation of probability judg-
ments, the matter is somewhat different. Probabilistic incoherence is no
longer a case of believing logically inconsistent (or mathematically incon-
sistent) propositions. In the foregoing example the subjectivist would not
interpret you as believing propositions Q and Q". Rather, he would inter-
316 Probability Jtidgments
pret you as having degrees of confidence in a different pair of propositions,

namely, .60-confidence in 'It will be rainy tomorrow' and .75-confidence in
'It will be rainy and windy tomorrow'. These propositions are not inconsis-
tent (either logically or mathematically).
On the subjectivist approach, then, it cannot be said that incoherence is a
defect because it guarantees at least one false belief. Why, then, is it a de-
fect? The most popular argument for the defectiveness of (subjective) prob-
abilistic incoherence is the Dutch Book argument. On the standard
approach a person's degrees of belief generate quotients at which the cog-
nizer is prepared to make corresponding bets. A Dutch Book is a set of bet-
ting odds that guarantee a net loss no matter which propositions turn out to
be true. The Dutch Book argument says that it is bad to have incoherent
degrees of belief because if you do, it is possible for an opponent to make a
Dutch Book against you.
But this argument has many weaknesses, as has been noted in the litera-
ture. One is not always in a betting situation. And what is wrong with
maintaining incoherent subjective probabilities if there is no opportunity or
occasion for anyone to make a set of bets with you? Furthermore, even if
the context is ripe for betting, maybe it is sensible to leave yourself open to
a Dutch Book if it would be very uneconomical in terms of your time and
effort to guarantee against one. Besides, even if an opponent could make a
Dutch Book, there is no guarantee he would. He might not be clever
enough to do this; and there are many situations in which it would be silly
of him to bet Dutch Book style to ensure a small gain.14
Finally, we can separate degrees of belief from betting odds. It might be
defensible to have incoherent degrees of belief if one were careful not to
engage in sets of bets that would necessarily yield a net loss. This would be
analogous to retaining a set of inconsistent beliefs but carefully avoiding
any temptation to infer random conclusions simply because contradictions
entail anything. Of course, standard dispositional theories of degrees of be-
lief would disallow (on conceptual grounds) the separation of belief degrees
and betting propensities. But such dispositional theories are highly ques-
tionable, as I will argue in Chapter 15.
Wherever the foregoing arguments ultimately lead, let us suppose that it
is a defect to make incoherent probability judgments, either on the objec-
tivist or the subjectivist interpretation. Still, would such a defect be a mark
of irrationality? I am inclined to answer, no. Consider the following: proba-
bility theory is a branch of mathematics, like other branches of mathemat-
ics. Failure to have learned or mastered other branches of mathematics is
not normally taken as a shortcoming in rationality. Nor is the failure to rec-
ognize every concrete application of such branches. By parity of treatment,
it would be wrong to view every deficiency in grasping or applying proba-
bility theory as a specimen of irrationality.
Probability Judgments 31 7
Until the time of Cantor, there was a general misunderstanding of the

concept of infinity. People offered many arguments on the subject that with
the hindsight of our improved grasp of the infinity concept, would be
viewed as fallacious. But it would be misleading to say that these misunder-
standings betrayed irrationality. People just do not come naturally to the
properties of the infinite; and until intellectual giants clarified this concept,
there were many confusions. The same could be said of the concept of prob-
ability. The modern concept of probability did not begin to be clarified
until the seventeenth century, and a full axiomatization of probability the-
ory was not achieved until the twentieth century. So it is hardly surprising if
the naive, untrained person should find it difficult to arrive at an adequate
conception of probability, or chance phenomena, all by himself. Similarly,
consider the fact that many different sorts of geometries are self-consistent.
This fact is not likely to be noticed by the uninitiate. Yet that is hardly
ground for calling people (even partly) 'irrational'.
What about people who have proper training in probability and statis-
tics? One of the interesting findings of Tversky and Kahneman is that even
such people frequently violate probabilistic rules. However, the evidence
here is not without ambiguities. As we saw in the previous section, the re-
sults of one study showed well-trained ('sophisticated') subjects performing
about as badly as untrained ('naive') subjects. But in another experiment,
trained subjects did quite a bit better. Only a third (36 percent) of the
Berkeley and Stanford graduate students who had had several advanced
statistics courses committed the conjunction fallacy when the test was
'transparent'. Moreover, as we will see in the next section, even naive sub-
jects do better when they are given problems that call attention to the class
inclusion structure of the problem.
Even 36 percent may seem like a high proportion for people with exten-
sive formal education of the relevant kind. But it is not a trivial matter to
recognize all the facts or principles you have learned that might be relevant
to a given cognitive task. Recall the examples in Chapter 10 concerning the
Sunday library hours and the relation between mold and bacteria. In hind-
sight it may be easy to say that Melanie and the microbiologists should have
pulled certain facts from LTM and applied them to the case at hand. But
this is only easy with hindsight, when we already start with the target facts
or principles clearly in mind. In the context of problem solving the ac-
cessing of relevant, previously learned principles is not a trivial matter. Fur-
thermore, as argued in Chapter 10, failure to access relevant facts is not
plausibly construed as a defect in rationality, even if it is some sort of cog-
nitive defect.
If these are not instances of irrationality, what on my view would count
as specimens of irrationality? As I have stressed earlier, I have no positive
account of (ir)rationality to offer. However, I am inclined to hold two
31 8 Probability Judgments
things: (a) irrationality is a defect in basic processes, not simply a failure to

have learned or fully mastered intellectual tools; and (b) not every defect in
basic processes is a defect in rationality. In particular, the defect must have
something to do with basic 'reasoning' processes, not just memory retrieval
processes, for example.
Notice that I am not arguing that people are natively rational in matters
of probability. I am only saying that the matter cannot now be settled, be-
cause we do not presently have an acceptable account of rationality and ir-
rationality, especially native rationality and irrationality. Pending a
satisfactory account, we are not in a position to conclude from Tversky and
Kahneman's evidence that people are irrational. Notice in particular that
people are capable of learning probability theory and of applying it cor-
rectly in many applications. The fact that they are misled or confused in a
special class of applications does not obviously make them irrational. Natu-
rally, it would help our theoretical plight if I had an account of rationality
to offer. Unfortunately, I find this a rather misty notion. Although I find
many popular requirements for rationality unacceptable, I have no good
substitutes to offer in their place.
I have been focusing on rationality because that is the term of appraisal
most widely used by other writers on this topic. This is an appropriate
juncture, however, to switch to another term of appraisal, 'justifiedness', for
which I have offered a positive account. The question is, how do people's
beliefs or credences about probability fare on the justifiedness dimension?
14.4. Probability and Justijiedness

In considering justifiedness, we must bear in mind the two distinct compo-
nents of justifiedness posited by my theory: primary justifiedness and sec-
ondary justifiedness. As far as secondary justifiedness (S-justifiedness)goes, it
is certainly possible that respondents who fail to use proper probabilistic
methods when they are applicable should have their resulting beliefs denied
the status 'S-justified'. If so, these beliefs will not be fully justified, since full
justifiedness requires S-justifiedness. However, I have not offered a firm ac-
count of S-justifiedness. There is an unresolved question of whether some-
one unexposed, or not amply exposed, to proper methods, should have his
beliefs denied the status of S-justified. So there is some question of whether
to fault naive subjects who fail to employ proper probabilistic considera-
tions. But the beliefs of amply trained respondents who fail to apply proper
methods certainly seem to merit the charge of secondary unjustifiedness.
Of course, there are no grounds for concluding that people's probabilistic
beliefs are never fully justified. Although the research I have surveyed fo-
cuses on probabilistic errors, it has not been shown that people never apply
probabilistic methods properly. Tversky and Kahneman themselves admit
that the problem tasks in their studies were deliberately constructed to

elicit conjunction errors, and they do not provide an unbiased estimate of
the prevalence of these errors.''
Under what conditions do people succeed in grasping and applying prob-
abilistic methods, and under what conditions do they fail? The root diffi-
culty in understanding and applying probabilistic ideas seems to be that
people do not naturally think of probabilities in terms of classes of possibili-
ties, and therefore do not think in terms of class inclusion, even when it is
relevant. As Tversky and Kahneman express it, people do not generally rep-
resent probabilities "extensionally." This suggestion is supported by several
studies, including some reported by Tversky and Kahneman themselves,
who gave the following health survey problem to statistically naive students
at the University of British Columbia.
A health survey was conducted in a sample of adult males in British Colum-
bia, of all ages and occupations. Please give your best estimate of the follow-
ing values:
What percentage of the men surveyed have had one or more heart
attacks?
What percentage of the men surveyed both are over the age of fifty-five
and have had one or more heart attacks?
This problem produced substantial conjunction errors: 65 percent of the re-
spondents assigned a higher estimate to the second question than to the
first. However, a subtly different version of the health survey problem given
to naive subjects produced very different results. This version read:
A health survey was conducted in a sample of 100 adult males in British Co-
lumbia, of all ages and occupations. Please give your best estimate of the fol-
lowing values:
How many of the 100 participants have had one or more heart attacks?
How many of the 100 participants both are over the age of fifty-five and
have had one or more heart attacks?
In this version only 25 percent of the subjects violated the conjunction rule.
Furthermore, this portion was reduced to a record 11percent when subjects
first estimated the number of participants over fifty-five years of age prior to
estimating the conjunctive category. Evidently, an explicit reference to the
number of individual cases encourages subjects to set up a representation of
the problem so that class inclusion is readily perceived and appreciated.
Other writers have explored the question of when subjects use reasonably
accurate statistical heuristics. R. E. Nisbett, D. H. Krantz, C. Jepson, and Z.
Kunda indicate that reasonably good statistical heuristics are used when (a)
the sample space and the sampling process are clear, and (b) the role of
chance in producing events is clear.'' Similarly, although subjects are often
oblivious to statistical factors like base rates, Fischhoff, Slovic, and Lich-
320 Probability ludgnents
tenstein showed that subjects become sensitive to base rates when they en-
counter successive problems that vary in this critical variable.''
As far as secondary justifiedness goes, people are evidently capable of
using correct probabilistic methods. They often do not use them, even when
they have been trained; but sometimes they do. Hence no universally bleak
conclusion about secondary justifiedness of probabilistic judgments is war-
ranted. Thus far, there are no grounds for the eager skeptic to rejoice.
But how does this conclusion affect primary justifiedness? How adequate
are the basic processes that get deployed in connection with probability
tasks? Are they so poor as to be excluded by all right J-rule systems? Or are
some of them sanctionable by a right J-rule system? Much of the work in
this field has centered on the representativeness heuristic, which bears the
brunt of the attack on native processes in probability judgment. Recall from
section 14.2, however, that Tversky and Kahneman no longer claim that
people rely exclusively on representativeness in making probability judg-
ments. They now say that predictions and probability judgments are
"highly sensitive" to representativeness, not completely dominated by it.
There must, then, be other processes that get deployed in such tasks.
In "Extensional versus Intuitive Reasoning" Tversky and Kahneman
speak of people using a "holistic, singular and intuitive" manner of evaluat-
ing probabilities.18This phrase, though, does not identify any specific cog-
nitive process. In another article they explain their use of 'intuitive' as
follows. A judgment is called 'intuitive' if it is reached "by an informal and
unstructured mode of reasoning, without the use of analytic methods or de-
liberate calculation."19 But this is a largely negative characterization,
which still does not identify the specific processes at work. There is no as-
surance, moreover, that only one such process is used. Maybe a number of
processes can be used, some more reliable than others. In particular, some
reIiable processes may be employed by sophisticated cognizers when they
correctly identify problems as susceptible to probabilistic methods and then
apply those methods appropriately.
But what about the representativeness heuristic? Doesn't the evidence
suggest that it is frequently used? And doesn't this seem to be a basic pro-
cess that would be barred from any right J-rule system? First, a point of ter-
minology. The term 'heuristic' connotes a deliberately chosen shortcut
procedure. In the present case the process is not deliberately chosen and is
not seen as a shortcut. So I will speak instead of a representativeness 'rou-
t i n e ' a n R-routine.
I do not wish to deny that use of the R-routine leads to errors in the class
of cases-to which Tversky and Kahneman draw our attention. Specifically,
when subjects rank the probability of a conjunction as higher than the prob-
ability of a conjunct, this is clearly an error. (At least it is an error if this
ranking expresses a belief in a comparative objective probability statement,
and if each event has a genuine objective probability.) However, it does not
follow from this that the R-routine is not admissible in a right J-rule system.
Perhaps the R-routine is generally quite reliable, though it breeds errors in
this subclass of cases. A case can be made for just this thesis by contemplat-
ing the possibility that the R-routine is a facet of the general use of match-
ing operations, or similarity assessments, in cognition.
Matching operations may well be a fundamental part of cognitive pro-
cessing; and matching of an object X to a model M seems to be the core of
the R-routine, as Tversky and Kahneman view it. Another way of putting
this idea is in terms of similarity assessments. The use of similarity is a ubiq-
uitous facet of human cognition (as noted in Chapter l l ) , and is particularly
stressed in recent work on concepts and categorization. Let us explore this
latter research, as a prelude to the discussion of the R-routine in probability
judgments.
Cognitive psychologists investigate the psychological representations and
processes that underlie the use and comprehension of (natural-language)
words, especially common nouns for ordinary objects like 'bird' or 'chair'. A
popular theory is that people mentally represent the properties in question
by means of 'prototypes', and that they decide how an object should be ca-
tegorized by assessing the similarity of the object to the category's
prototype.
The chief rationale for this theory is the ubiquity of 'typicality' effects.
For example, subjects find it natural to rate the subsets or members of a cat-
egory with respect to how typical each is of the category. Such ratings were
first reported by L. J. Rips, E. J. Shoben, and E. E. Smith and by E. ~osch."
Rosch had subjects rate the typicality of members of a category on a 1-to-7
scale, where 1 meant very typical and 7 very atypical. Subjects were ex-
tremely consistent. In the bird category 'robin' got an average rating of 1.1,
and 'chicken' 3.8; 'football' was rated a very typical sport (1.2) but weight-
lifting was not (4.7); 'carrot' was rated a very typical vegetable (1.1),and
'parsley' a less typical one (3.8).Other typicality effects include these: (a)
categorization is faster with more typical test items; (b) typical members of
a concept are the first ones learned by children; and (c) typical members of
a concept are likely to be named first when subjects are asked to produce all
members of a category.
Typicality effects are to be expected if subjects represent a category in
terms of a prototype, and if categorization proceeds by judging similarity to
prototype. There are also several variants of the prototype theory that
would all invoke similarity assessments, or matching processes, in
categorization.'l
What is important for my purposes is that judgments of similarity, or fea-
ture matching, are fundamental ingredients of the categorization process. If
this general idea is correct, the use of representativeness is an extremely
pervasive facet of human cognition. Whenever objects or events are cate-
gorized as belonging (or not belonging) to this or that concept, representa-
tiveness is employed. Furthermore, as long as cognizers apply this routine
to perceptually detected features of the object, as long as the cognizer's
prototype of the category corresponds closely enough to the prototypes
used by other members of the linguistic community, and as long as the rou-
tine of matching to prototype is executed properly, virtually all the cate-
gorization judgments made in this manner will be true! Objects judged to be
members of the concept Y will really be members of Y-that is, will really
qualify under the lexical label, as it is used by the linguistic community. In
short, this use of the R-routine will be highly reliable.
If the R-routine is so reliable for this large class of cases, the general rou-
tine might be endorsed by a right system of J-rules, despite the fact that the
routine generates errors in some of its applications. For purposes of com-
parison, consider the case of perceptual processes that sometimes breed il-
lusion. (Tversky and Kahneman themselves draw the analogy between
judgmental illusions and perceptual illusions.) Presumably, the processes
responsible for well-known illusions like the moon illusion or the Ames-
room illusion are generally reliable processes. As such, they would presum-
ably be sanctioned by a right J-rule system, despite the fact that they un-
doubtedly breed false judgments in the designated contexts. Thus, even in
the illusion cases, a cognizer's belief is justified, though it is of course false.
Similarly, if the R-routine is simply part of a generally reliable judgmental
process, the beliefs it generates on probabilistic subject matter might still be
justified, even if they are false.
Actually, this needs to be qualified. The most we could conclude is that
these beliefs have primary justifiedness. They could still be denied the status
of secondary justifiedness, because of failure to deploy learned probabilistic
methods. The same point would hold for the perceptual cases. If an adult
believes that the moon hovering over the horizon is genuinely bigger this
evening than at other times, we would not accord this belief the status 'jus-
tified'. We would surely hold that the report of his visual system (his 'visual
belief) should be overridden by his central system, by appealing to facts or
methods he has learned from past experience. Similarly, even if the R-rou-
tine is permitted by a right system of J-rules (at the primary level), someone
who has studied probability theory might be held accountable for failure to
employ correct probabilistic methods to revise his initial, intuitive
judgment.
There is another, more important, reason why the foregoing analysis
might not suffice for justifiedness. We have been supposing that there is a
single process that employs matching, or the R-routine; but this is dubious.
Arguably, there are significant distinctions between the process used in ca-
tegorization and the process used in the probability judgment cases. Al-
though both employ a common component of similarity assessment, or

matching, there are points of difference.
First, in the categorization cases, all (or most) critical features of the tar-
get object are known to the cognizer. If you see a robin flitting about, you
can detect most of the features critical to its being or not being a bird. But
when a subject seeks to judge whether Linda is (probably) a bank teller, or a
feminist and a bank teller, he lacks information about critical features of
Linda. He does not know whether Linda is employed, whether she is em-
ployed in a bank, and so on. So the relationship between the known features
of the object (the features given by the personality sketch) and the target
properties (feminist, bank teller, feminist and bank teller) is quite different.
In categorization cases the cognizer tries to decide whether exemplification
of the known features qualijies the individual for the target property. In the
probability judgment cases it is assumed that the known features do not
qualify the individual for the target property. The only question is whether
they make it likely that this property is instantiated. And if the two uses of
the R-routine are different, we really have two different processes before us.
Clearly, one might be permitted by a right J-rule system while the other
one isn't permitted.
We cannot conclude, then, that the use of representativeness in probabil-
istic tasks does yield justified beliefs. Nevertheless, that possibility is not de-
finitively precluded. Grounds for deep pessimism are therefore premature.
Moreover, since representativeness has not been shown to be the sole pro-
cess available in confronting probability questions, the prospects for justi-
fied beliefs in this domain are brightened still more. Undoubtedly there are
signs of trouble. But the gloomy diagnosis of a (local) skeptic is not yet vin-
dicated, at least as far as justifiedness is concerned.
chapter 15
Acceptance and Uncertainty
15.1. Acceptance and Subjective Probabilities

What kinds of doxastic categories are appropriate for the human mind, and
hence for human epistemology? This is a question that cannot be answered
a plriori; nor should it be answered by uncritical acceptance of the catego-
ries of ordinary language. It should be answered only with the help of cog-
nitive science. Thus, as I suggested at the beginning of Part I1 (in Section
9.1), epistemology needs contributions from cognitive science to select its
descriptive apparatus for doxastic states. This chapter begins by exploring
this topic.
Like many works in epistemology, this work has regularly deployed a bi-
nary notion of belief. A person either believes a proposition, or he doesn't.
Another label for binary belief is acceptance. Other epistemologists, how-
ever, think of belief, or credence, as a graded, or quantitative, notion. This
is commonly expressed in the language 'subjective probabilities', and 'per-
sonal probabilities'. On this view there are varying degrees of confidence
one can have in a proposition, widely represented as points on the interval
from 0 to 1.
The question I now wish to ask is: Are both of these credal-state no-
tions-binary belief and graded belief-psychologically legitimate? If not,
we need to know which one is legitimate; or if so, we need to know what
their connection is. In particular, are they two entirely different psychologi-
cal categories (or continua)?Or do they somehow fit with one another into a
single psychological category?
From an introspective standpoint I find the notion of acceptance, or un-
qualified belief, quite plausible. There seem to be many propositions I just
belieue, without any (salient) doubt or reservation. This would include
propositions about my present address, about the university at which I am
employed, and about the names of the Democratic and Republican candi-
dates for President in the 1984 election.
At the same time, there are many cases in which we are inclined to be-
chapter 15
Acceptance and Uncertainty
15.1. Acceptance and Subjective Probabilities

What kinds of doxastic categories are appropriate for the human mind, and
hence for human epistemology? This is a question that cannot be answered
a plriori; nor should it be answered by uncritical acceptance of the catego-
ries of ordinary language. It should be answered only with the help of cog-
nitive science. Thus, as I suggested at the beginning of Part I1 (in Section
9.1), epistemology needs contributions from cognitive science to select its
descriptive apparatus for doxastic states, This chapter begins by exploring
this topic.
Like many works in epistemology, this work has regularly deployed a bi-
nary notion of belief. A person either believes a proposition, or he doesn't.
Another label for binary belief is acceptance. Other epistemologists, how-
ever, think of belief, or credence, as a graded, or quantitative, notion. This
is commonly expressed in the language 'subjective probabilities', and 'per-
sonal probabilities'. On this view there are varying degrees of confidence
one can have in a proposition, widely represented as points on the interval
from 0 to 1.
The question I now wish to ask is: Are both of these credal-state no-
tions-binary belief and graded belief-psychologically legitimate? If not,
we need to know which one is legitimate; or if so, we need to know what
their connection is. In particular, are they two entirely different psychologi-
cal categories (or continua)?Or do they somehow fit with one another into a
single psychological category?
From an introspective standpoint I find the notion of acceptance, or un-
qualified belief, quite plausible. There seem to be many propositions I just
belkue, without any (salient) doubt or reservation. This would include
propositions about my present address, about the university at which I am
employed, and about the names of the Democratic and Republican candi-
dates for President in the 1984 election.
At the same time, there are many cases in which we are inclined to be-
326 Acceptance and Uncertainty
ity of 99, with respect to each ticket, that the ticket will lose. On the pre-
sent proposal this would be interpreted as acceptance of the proposition
that the ticket will lose, for each such proposition. Yet he also accepts the
proposition that some ticket will win, so the current proposal forces us to
ascribe to him inconsistent acceptances. That seems unfair. Nothing in the
initial description of him warrants the conclusion that he has inconsistent
beliefs.
Thus far I have worried about finding mutually comfortable niches for
both subjective probabilities and acceptance. But now I want to turn to
problems for the subjective probability interpretation of credal uncertainty.
In other words, granted that there is some psychological reality to the no-
tion of tentative, hesitant, or incomplete conviction, is this correctly or best
described in terms of the probability calculus? I wish to question this de-
scriptive assumption, generally associated with (subjective) Bayesianism.
One problem is the empirical evidence against the notion that people
conform their beliefs to the probability calculus. One piece of evidence, in-
tensively discussed in Chapter 14, concerns the principle of probability the-
ory that the probability of a conjunction cannot exceed that of any of its
conjuncts, a rule people readily violate. Similarly, according to standard
models, the probability of any tautology, or necessary truth, is 1; and the
probability of a contradiction, or impossible proposition, is 0. But people
are not always maximally confident in every tautology or maximally confi-
dent in rejecting every contradiction. When they do not recognize a tautol-
ogy or a contradiction, they may have quite different degrees of confidence
in them.
Next, consider whether it is plausible that humans should have a psycho-
logical mechanism that fixes credal states in accord with the probability
calculus. Remember, I am here regarding subjective probability as a de-
scriptive theory, one that is intended to hold for all human beings, whatever
their age or educational background. Subjective credences must therefore
be fixed by native mechanisms. Are there any native mechanisms that
would make credal states behave like probabilities? To assume there are is
to assume there are native mechanisms for making arithmetic calculations
of the kind needed for determining probabilistic relations. But native mech-
anisms do not seem adequate for even the simplest arithmetic. Most arith-
metic needs to be culturally acquired by explicit instruction; it is not
programmed into the organism. Admittedly, there might be mechanisms
that simulate arithmetic computation at a preconscious level, just as visual
mechanisms apparently simulate mathematical operations in determining
distance or in computing the trajectory of an object flying toward you. But
there is no clear evidence of credence-fixing mechanisms that simulate the
probability calculus in this way.
Many theorists of subjective probability might reply that the subjective
probability model is descriptively tenable as long as there is a procedure for
Acceptance and Uncertainty 327
ascribing such states to the organism on the basis of overt choice behavior,
including expressed preferences or indifferences vis-a-vis gambles. A num-
ber of such procedures, of course, are proposed in the literature? This liter-
ature includes a number of 'representation theorems', which prove that if
certain preference axioms are satisfied, then there exists a set of subjective
probabilities that can be imputed to a decision maker. To be more precise,
if these axioms are satisfied, there exists a subjective probability function P
and a desirability function D such that, when 'subjective expected utility'
(SEU) is calculated in terms of them, an item X will be preferred to an item
Y if and only if SEU (X) > SEU (Y).
There are several problems with this approach. First, the mere fact that
subjective probabilities can be imputed to a cognizer does not establish any
psychological reality to those imputed states. Lots of different models of in-
ternal states and processes could be consistent with choice behavior. Just
because a given assignment of desirability and subjective probability is pos-
sible does not show that there is any genuine psychological reality to this
assignment. The epicycles of Ptolemaic astronomers kept their theory con-
sistent with observed planetary trajectories, but that did not establish the
astronomical reality of those postulations.
Second, in the present case, there are empirical doubts about some of the
preference axioms presupposed by the representation theorems. For exam-
ple, Tversky found systematic and predictable intransitivities of prefer-
ence,%nd K. R. MacCrimmon reported violations of most of Leonard
Savage's postulates.6 Moreover, many of the popular approaches to subjec-
tive probability assignments presuppose that choice is determined by
SEU; but this presupposition is empirically quite doubtful. For example,
in one of Tversky and Kahneman's experiments subjects were asked
to choose between gambles A and B (Choice I), and then between C and
D (Choice 11).
Choice I A = $1,000 with probability .5, or $0 with probability .5
B = $400 with probability 1
-
Choice I1 C = $1,000 with probability .l,or $0 with probability .9
D $400 with probability .2, or $0 with probability .8
Nearly all subjects chose B over A, and C over D, and analogous results
were obtained using different payoffs and probabilities. These results pre-
sent grave difficulties for the assumption that choice follows S E U . ~ To the
extent that subjective probability assignments are predicated on standard
decision theory, there are serious grounds for doubts about their empirical
worth.
There is a method for determining subjective probabilities that does not
explicitly presuppose SEU: the Dutch Book method. This determines a sub-
jective probability by taking the highest odds one is willing to offer for a bet
on p-let's say a to b. Then the degree of belief in p is fixed as a/a + b.
For example, if one would bet on p at maximum odds of 3 to 1, then the

assigned degree of belief in p is .75. What justifies this method? The as-
sumption is that if degrees of belief are fixed this way, and bets made ac-
cordingly, then no Dutch Book can be made against one.
This justification, however, is very weak. At most it shows that rational
degrees of belief would reflect maximum betting odds. But we are in-
terested in actual degrees of belief, which may not be rational. A person
may be willing to bet at unrealistically high odds because he likes risky situ-
ations, or unwilling to bet at odds as high as his true confidence level be-
cause he dislikes betting. So his real degree of belief may not coincide with
his maximum betting odds.
Finally, note that the different proposed methods for 'operationalizing'
subjective probability are not equivalent. They would not yield the same
assignments in the same cases8 Which of these nonequivalent operationali-
zations of subjective probabilities should be chosen? Clearly they cannot all
correctly limn the structure of inner mental reality. Once this is recognized,
however, one is forced to wonder whether any of them do.
In conclusion, there are numerous reasons for doubting that subjective
uncertainty can be correctly captured by applications of the probability
notion. There is, of course, a widespread view to the contrary. But perhaps
this merely arises from the allure of a precise mathematical tool-the prob-
ability calculu+which dangles the prospects of a tidy and elegant theory.
This makes it extremely tempting to try to devise a workable application.
But there are scant grounds for thinking that any such theory does capture
psychological reality.
Let me emphasize that what I am challenging here is the use of probabi-
lities as an account of credal uncertainty. I am not challenging the idea that
probabilistic concepts can occur in the propositional contents of people's
credal states. There may well be legitimate notions of objective probability,
or chance, as indicated in Chapter 14. People may learn these notions and
incorporate them into various propositions, some of which become the ob-
jects of belief. A person might accept the propositional content 'It probably
won't rain today', or 'It's more likely to rain today than to snow'. I have no
objection against such 'propositionalized' treatments of uncertainty. But
the aim of subjective Bayesianism, of course, is to replace objective proba-
bilities with probabilities construed as psychological states. This is the theo-
retical orientation I mean to challenge.
15.2. Winner-Take-A11 Networks

In the spirit of psychological realism, what might be a better way to model
uncertainty, and how could we achieve an integrated model of acceptance
as well as uncertainty? Recent work in the 'new connectionist' literature
suggests a promising avenue.
New connectionism is an approach to cognition with two main emphases.

First, it assumes that enough is known about brain structure to offer a seri-
ous constraint on hypotheses about cognition. It is a necessary exercise for
anyone proposing a cognitive theory to provide at least an argument to
show that the brain could work in the way the theory requires.' Second,
new connectionism assumes that an important array of neural processes are
parallel as opposed to sequential; this parallel processing involves the use of
massive numbers of neural units and connections.
One critical argument for this approach appeals to basic properties of
neurons and considerations of computational speed. The basic computation
speed of neurons is a few milliseconds, yet they must account for complex
cognitive behaviors carried out in a few hundred milliseconds. So entire
complex behaviors are carried out in less than a hundred time steps. How-
ever, current artificial intelligence and simulation programs, which are
largely sequential, require millions of time steps. Hence, these sequential
models seem inadequate. lo
Let us examine one way of developing connectionism, the one formu-
lated by Jerome Feldman and Dana Ballard and by Lokendra Shastri and
~e1dman.l'Their approach includes a particular kind of neural network-
the 'winner-take-all' network-which provides, I shall argue, a fruitful way
of conceptualizing acceptance and uncertainty. In this section, then, I first
sketch the general connectionist framework of Feldrnan and colleagues.
Next I present the winner-take-all (WTA) idea. Third, I suggest how WTA
networks may provide a good model for acceptance and uncertainty. In the
following section I shall raise some normative questions about WTA net-
works as belief-king mechanisms.
Feldrnan and colleagues develop a standard model of an information pro-
cessing "unit," based on the current understanding of neurons. A unit may
be used to model anything from a single neuron (or neuron part) to a major
subsystem of neurons. The exact realization of a unit is left intentionauy
vague. Attention is focused on the computational properties of units and
networks of units.
Units communicate with the rest of a network by transmitting a simple
value. A unit transmits the same value to all units to which it is connected.
The output value is closely related to the unit's potential, and is best de-
scribed as a level of activation, such as firing frequency. A unit's potential
reflects the amount of activation the unit has been receiving from other
units. For concreteness, think of a unit as a single neuron, having multiple
dendrites, or dendritic sites, each of which receives inputs from other neu-
rons. This is illustrated in Figure 15.1.
All inputs are weighed and combined in a manner specified by site func-
tions and a potential function, in order to update a unit's potential. The
computational behavior of units can be described in terms of these main
variables:
p = a continuous value called potential

v = an output value (approximately 10 discrete values)
i = a vector of inputs i,, i,, . . ., in
Two functions define the values of potential and output at time t + 1based
on the values at time t:
The arrow notation is borrowed from the assignment statement of pro-

gramming languages. The value at time t + 1, to the left of the arrow, is
determined by the values of the variables at time t, represented by the ex-
pression to the right of the arrow. Time is broken into discrete intervals.
The variable q ranges over a small set of states that a unit can be in; for
purposes of this exposition, though, these states can be ignored. Also, while
the output function relates the output v to several variables, the authors
frequently treat it as just the rounded value of p, the potential. I shall follow
that practice here.
A unit does not treat all inputs uniformly. Units receive inputs via con-
nections, or links, and each incoming link has an associated weight. A unit
weighs each input using the weight on the appropriate link. A weight may
have a negative value, which corresponds to an inhibitory link. This is im-
portant, since inhibition is a basic neural mechanism, and many of the con-
structions that interest us are networks featuring mutual inhibition (or
lateral inhibition), a phenomenon widely found in nature.
As indicated earlier, each unit may have more than one input site, and
incoming links are connected to specific sites. Each site has an associated
site function. These functions carry out local computations based on the
input values at the site, and it is the result of this computation that is pro-
cessed by the functions P and V.
Figure 15.1
A, receives on input of 6 units, then 2 per t i m e step

B, receives an input of 5 units, then 2 per time step
- -
6
5.5
5.5
6
6.5
7.5
9.5
Sot
Figure 15.2
To illustrate these ideas, let a unit be a neuron with dendrites that receive
inputs. Each dendrite can be thought of as an alternative enabling condi- ,
tion: if it receives enough inputs, it activates the neuron. The firing fre-
quency of the neuron might be the maximum of the firing rate at any one of
its dendritic sites. The formula for this unit's potential could be written as
follows:
pt+l+-- pt + Max (i, + i2; iJ + i4; i5 + i6 - i7)
This is shown in Figure 15.1. The minus sign associated with i7 indicates
that it is an inhibitory input.
Drawing on these ideas, I will illustrate a configuration of symmetric
units that mutually inhibit one another. First, consider Figure 15.2, con-
taining two units, A and B, each with two input sites. Suppose the initial
input to site Al is 6, after which that site receives 2 inputs per time step.
Suppose B1 receives an initial input of 5, and then 2 inputs per time step.
Notice that each unit transmits an inhibitory input to its mate, at site 2.
Suppose further that the weights at the sites are as follows. The weight at
site 1 is 1, and the weight at site 2 is -.5. Then at each time step, each unit
changes its potential by adding the external input value and subtracting
half the output value of its mate. So we have the equation:
Finally, assume that the output variable v takes the discrete values: 0, . . .,
9; and that v is always the rounded value of the potential p.
The result, as indicated in the table in Figure 15.2, is that the potential
and output of unit B are gradually reduced to 0, while unit A gradually in-
t P( A ) P( B) P(C) P( D l
1 6 5 6 5
2 6.5 4.5 6.5 4.5
3 7.5 305 7.5 3.5
4 9.5 1.5 9.5 1.5
Sot 0 Sat 0
Figure 15.3
creases its potential and output to a saturation point. There the system
stabilizes.
Another symmetric configuration involves coalitions of units, shown in
Figure 15.3.The idea here is that units A and C form a coalition with mutu-
ally reinforcing connections. This coalition competes with another coalition
comprised of B and D. Competition between the coalitions is determined
by their mutual inhibition. As in the previous example, such mutual inhibi-
tion results in convergence toward a stable state, but here the convergence
is faster: convergence occurs in only four time steps. This system is de-
scribed by the following equations and specifications:
W1 = 1, W2 = .5, W3 = -.5
v = round (p)
pt+= + +
~ pt i, .5(i, - i3)
A, C start at 6; B, D start at 5
A, B, C, D have no external input for t > 1
Competing coalitions of units is a central organizing principle behind many

of Feldman and colleagues' models and is critical to our topic. To illustrate
this principle psychologically, consider two visual interpretations of the
Necker cube, shown in Figure 15.4.At each level of visual processing, there
are mutually contradictory units representing alternative possibilities. The
dashed lines denote the boundaries of coalitions that embody the alterna-
tive interpretations: one interpretation depicts the A-D-H-E face as closer
to the viewer, the other depicts the B-C-G-F face as closer. In the Necker
cube example, though, there may be no convergence on a stable value. In-
stead, there can be the familiar phenomenon of figure reversal, with the two
interpretations oscillating in relative strength.
Of special interest to me are the WTA networks. These exhibit the prop-
erties of convergence and stable states illustrated in Figures 15.2 and 15.3.
The basic idea motivating WTA networks, as Feldman and Ballard explain
it, is that cognitive systems make decisions. This includes behavioral deci-
sions, such as turning left or right and fight-or-flight decisions; but it also
includes cognitive decisions, such as the interpretation of ambiguous words
or images. I am interested in WTA networks because I view them as prom-
ising models of acceptance, or (binary) belief.
The critical property of a WTA network is that only the unit with the
highest potential (among a set of contenders) will have output above zero
after some settling time. An example of a WTA network is one that operates
in one time step for a set of contenders each of which can read the potential
of all the others. Each unit in the network computes its new potential ac-
cording to the rule:
p t--- If p > max (ii), then p; otherwise 0.
\
\ I \ I
Figure 15.4
In other words, each unit sets itself to zero if it knows of a higher input.12
This general idea is illustrated in Figure 15.5. The inhibitory connections
guarantee that each unit stops if it 'sees' a higher value. This is fast and
simple, though probably not terribly realistic. However, there are other
ways in which a WTA network might be realized.
The basic idea of acceptance, as I view it, is that a proposition is accepted
only when all rivals, or contraries, are rejected. This is well captured by the
idea that all rival hypotheses, or interpretations, have their activity levels
reduced to zero. This does not necessarily mean that these competing hy-
potheses cannot later be reactivated. But acceptance marks the emergence
of a decisive winner, not a hypothesis that merely has more strength than its
competitors. At the same time, belief must also be a reasonably stable, or
settled, state. If one's attitude is continually fluctuating, one does not yet
have 3 belief. Roughly this account of belief was endorsed by C. S. Peirce,
who viewed belief as a "calm" and "satisfactory" state, in the nature of a
"habit."13
This account of acceptance carries with it a plausible account of its com-
Winner-Toke- All
E a c h unit stops if i t sees o higher volue
Figure 15.5
plement: uncertainty. Uncertainty is to be viewed as a condition of contin-

uing competition among rival alternatives, whose relative attractiveness-
here interpreted in terms of activity level-is still in flux. It is an unstable
state, which has not yet converged on any one contender. This depiction of
uncertainty is phenomenologically appealing. Uncertainty is irresolution or
indecision: a condition in which competing ideas battle one another, with-
out one clearly besting the others.
Apart from introspective appeal, this approach has the virtue of unifying
the treatment of cognition and choice (or volition). The phenomenon of un-
certainty is as central to the volitional domain as it is to the doxastic. The
difference, of course, is that volition, or choice, is directed at behavioral al-
ternatives, competing plans of action, rather than at hypotheses or proposi-
tions. But here too, I propose, uncertainty consists in the absence of a
clear-cut winner. There is continued oscillation among the set of contend-
ers; or at least multiple contenders are kept 'on hold', while deliberation
continues. As long as such rivalry persists, there is the feeling of hesitancy,
of tentativeness, of 'not having made up one's mind'. Felt uncertainty is the
sense of continuing competition among contenders, whereas acceptance is
the victory of one contender over all others. This parallel treatment of voli-
tional and doxastic decision fits well with the plausible conjecture that the
same types of computing mechanisms subserve decision making in both
domains.
Perhaps the clearest cases of cognitive acceptance occur in perception.
Perceptual systems normally deliver just one interpretation to conscious-
ness, having somehow eliminated all competitors, even when some of these
are quite viable. In the famous Ames room there are two seemingly viable
interpretations of the visual input, but the visual system decisively selects
just one, which portrays the room as normal and the two individuals in it as
dramatically different in size. As it happens, this interpretation is false.
What interests me here, though, is simply the making of a decisive, categor-

ical choice. Ambiguous figures like the Necker cube, which induce inter-
pretive reversals, are comparatively rare.
I believe that the nonperceptual systems also tend to favor categorical
doxastic choice. Although doxastic uncertainty or indecision is more com-
mon here than in the perceptual domain, there is a built-in preference for
finding 'winners'. Thus, contrary to what the standard subjective probabil-
ity model suggests, acceptance is a natural, highly prevalent doxastic state,
not a radical or pathological state, as judged by the normal operating char-
acteristics of the system. This idea was anticipated by Peirce, who viewed
uncertainty, or "doubt," as an "irritation" of which the system seeks to rid
itself by replacing it with belief.
Several other features of the neuropsychological models proposed by
Feldman and colleagues suggest properties of doxastic activity different
from those implied by typical subjective probability approaches. First, al-
though their models do have a place for degrees of confidence-repre-
sented by the magnitude of a unit's output value (for example, its firing
frequencykthere is no assumption that the degrees of confidence among a
set of contenders must add up to a fixed value. This contrasts with the prob-
ability model, in which the sum of all mutually exclusive possibilities must
be 1.14
Second, mutual-inhibition models of the sort I have canvassed tend to de-
press the level of activity of (most) units, which results in rapidly changing
levels of confidence (at least until some stable point is reached). I think this
accounts for the familiar fact that people find it quite difficult to specify
precise degrees of confidence on the basis of introspection, or even to make
comparative assessments of confidence. Typically, the activity level asso-
ciated with competing alternatives changes too quickly. This contrasts with
subjective probability approaches, which usually assume definite degrees of
confidence.
At this point, an objection arises. Surely people are capable of making
precise probability estimates, at least if pressed; and they often settle on
precise odds in placing bets. How are these activities compatible with my
theory? The crucial point here is that judgments of probability need not al-
ways, or even typically, be expressions of subjective confidence. They may
be assessments of objective chance (as indicated at the end of section 15.1).
In making probability estimates of tomorrow's weather, or a patient's pros-
pects for recovery, or the outcome of a horse race, people assess objective
chances, for which known (or believed) frequencies provide relevant clues.
Where probability estimates are anchored to frequencies, rather than to in-
trospected credal intensities, definite numerical values are more easily
formulated.
The proposed account of uncertainty focuses on events accessible to con-
sciousness, or on events occurring in active memory. But what about the

phenomenon of uncertainty in LTM? Does this too consist in nonconverg-
ing, often fluctuating, levels of activation among competing hypotheses?
The proper treatment of uncertainty for LTM may be quite different-a
possibility that current models of subjective probability ignore entirely. In-
deed, I suspect that there is not any corresponding phenomenon of uncer-
tainty in LTM at all. Everything in LTM may be in a 'settled' state, without
ongoing competition. (It is noteworthy, in this regard, that theories of LTM
do not feature a strength variable marking belief intensity. Variables repre-
senting 'strength' are commonly found, but these mark something like asso-
ciative strength between nodes in a network, which concerns the
probability of one node activating another. In other words, the strength
variable concerns likelihood of retrieval, not firmness of conviction.)
I do not mean, to deny the presence in LTM of something functionally
similar to degrees of uncertainty. But what is there, I conjecture, is infor-
mation stored (without reservation!) that is related to manifestations of un-
certainty. First, people may store information about previous entertaining~
of competing hypotheses or past weighings of various strands of evidence.
When the question is raised anew, they retrieve these contents and recycle
the old thoughts, giving rise to renewed uncertainty All this is compatible
with memories being stored as states of binary belief, not as degrees of con-
fidence. Second, people may store information about previous objective
probability estimates. When such beliefs about previous estimates are re-
trieved, they tend to reassign those estimates, unless new information is
encountered.
The first of these proposals may provide a simple explanation of the ob-
served tendency of beliefs to become more confident over time (when no
new evidence is acquired). A person may forget some of the considerations
once raised in favor of a weak hypothesis, which fared relatively poorly in
previous competitions. Indeed, one may fail altogether to retrieve a weak
contender, since it will probably have received less attention in previous .
deliberations. When competition is renewed, it is easier for the formerly as-
cendant hypothesis to win once again, even more quickly and decisively,
since there is less material in mind to compete with it.
I have here sketched a mode1 of acceptance and uncertainty that I regard
as psychologically plausible. Its account of credal uncertainty is better, I
think, than that of standard subjective probability models. Its neat accom-
modation of acceptance, which seems to be a psychologically robust phe-
nomenon, also gives it superiority over those models. Finally, it fits within a
general framework that can make claims for neural realism.
As I have stressed, this theory of credal uncertainty does not account for
all phenomena commonly treated under this rubric. In particular, it gives
no systematic treatment of 'propositionalized' uncertainty, that is, doxastic
attitudes whose propositional content involves notions of objective chance

or likelihood. However, for purposes of primary epistemology, a theory of
credal uncertainty is central; and that is what my (admittedly rough) model
aims to capture.''
15.3. Episternic Evaluation of WTA Networks

The model I have sketched of acceptance and uncertainty is a purely de-
scriptive model, underwriting the appropriateness of the doxastic category
of acceptance. It is now time, however, to turn to evaluative issues. Granted
that human cognition produces categorical (as opposed to graded) beliefs,
how satisfactory are the native mechanisms of belief formation? Proceeding
on the assumption that human cognition widely employs WTA processes,
how good, epistemically, are these processes? Most important, can they
yield justijied beliefs?
To answer this question, of course, we need an assessment of the truth
ratio that WTA processes would tend to generate. We cannot make such an
assessment, however, without more information. First, we need an exact
specification of the WTA process or processes. Does even the slightest dif-
ference in strength between two rival units reduce the weaker rival to zero?
That is what the model sketched in section 15.2 provides. But alternatives
to that model are readily available. There might be some constant d, such
that a given unit's strength is reduced to zero only when the strength of a
rival unit exceeds it by at least d. If d is fairly large, it is hard for a unit to
become a winner. This property might serve to forestall rash beliefs,
thereby increasing the truth ratio. Hence, the existence and the exact value
of such a d-parameter would be critical determinants of the truth ratio of
any WTA network.
Second, we need more information about rivalry, or competition. Con-
sider a lottery example. Suppose there is a 1,000-ticket lottery and S consid-
ers the proposition, 'Ticket no. 1 will lose'. Which propositions are its
'rivals'? (For convenience I speak of propositions, rather than the units that
represent the propositions, as rivals.) If a proposition's rivals are only its
contraries, presumably this proposition is much stronger than any of its
rivals. The chief rival is, 'Ticket no. 1 will win'. This presumably has much
less support than 'Ticket no. 1 will lose', so the latter will be accepted.
However, suppose the rivals of a proposition p include any proposition q
that is negatively relevant to it, meaning any proposition q such that the
probability of p given q is smaller than the prior probability of p.16 Then
'Ticket. no. 2 will lose', 'Ticket no. 3 will lose', and so on, are all rivals of
'Ticket no. 1 will lose'. Since all these rivals are tied in strength, no rival
will be reduced to zero, and none will be accepted.
Many epistemologists would be cheered by this last prospect, since they
deem it inappropriate to accept the hypothesis 'Ticket no. 1 will lose'.

Whether or not this is so, the point is that we cannot determine whether a
WTA mechanism would have this result unless we are given more informa-
tion about how rivalry, or competition, is established. How does the mind-
brain decide which propositions to treat as rivals or competitors, that is,
which ones get represented by mutually inhibiting computing units?
A third question to raise about a WTA process is how it performs when
information is very sparse. Although the only evidence on a given issue may
favor a given hypothesis, this evidential support may nonetheless be (intui-
tively) weak. Does the supported hypothesis still get accepted, because all
rivals have zero strength? Such a mechanism would not be conducive to a
high truth ratio."
Fourth, even were these issues resolved, the truth ratio of a WTA process
alone cannot be assessed. One critical complementary process, or set of
processes, is the process or processes by which initial strengths are assigned
to units. If these processes assign strengths to units in a manner highly sen-
sitive to truth, the WTA process might be justificationally viable. But if no
such truth-sensitive complementary processes are available, it would not be
viable.
This shows that one cannot assess a WTA process for truth ratio all by
itself. But that should come as no surprise. Remember that, within my the-
ory of justifiedness, single rules authorizing single processes are not assumed
to have a determinate truth ratio. I speak only of the truth ratio of an entire
rule-system, authorizing an entire set of cognitive processes. The important
question to ask, then, is whether a WTA process can be complemented by
other processes that are jointly sanctionable by a right rule system, in virtue
of a high enough truth ratio. Thus far, nothing obviously precludes WTA
processes from being licensed by a right J-ruIe system.
There are other materials in the connectionist approach that bear on this
issue and that are worth examining. In particular, there is the idea of mutu-
ally reinforcing coalitions (recall the illustrations in Figures 15.3 and 15.4).1
suggest that this idea may correspond to the epistemologist's conception of
coherence. It may provide support for the notion that coherence-deploying
mechanisms are part of our basic cognitive equipment. These coherence-
deploying mechanisms may also serve to promote a high truth ratio.
By 'coherence' I here mean some sort of positive relationship--not
merely logical compatibility-the details of which would have to be
worked out elsewhere. In outline, it might come to something like this.
Some combinations of events or states of affairs are conceptualized as a
unity because they instantiate previously encountered patterns. For exam-
ple, they match familiar schemas, frames, scripts, or scenarios. To the extent
that a hypothetical event or state can be fitted into such a unity with other
hypothetical events or states, these form a set of mutually reinforcing possi-
bilities: a coalition. Furthermore, the larger and more detailed such a coali-
tion, the greater the amount of reinforcement each hypothesis (unit) re-
ceives from the remaining members. Thus, a comprehensive and detailed
scientific theory would have a high degree of coherence. Of course, no
member of a coalition can raise the strength of others unless it gets some
strength from outside the coalition; presumably, the external sources of ex-
citation are ultimately traceable to perception. Thus, perception as well as
coherence will serve as a major determinant of strength.
How can coherence, or coalition formation, serve as a guide or indicator
of truth? First, since a necessary condition of a hypothesis' being true is its
compatibility with other truths, the absence of coherence may be a good
sign of falsity. 'Isolated' hypotheses cannot benefit from reinforcement by
other coalition members, so they are ultimately rejected. Furthermore, if
coalitions are formed under some previously established pattern*
schemas, frames, scripts, scenarios, and so on-instances of which have
been frequently encountered in the past, there is evidence that this pattern
has a non-negligible prior probability of occurrence. Thus, a coalition will
represent not merely an abstract conjoint possibility, but also an instance of
a type with a non-negligible chance of recurrence. So, on the present con-
ceptualization, coalitions register 'induction' from past observation-in the
form of inductively based schemas, scripts, and so on-as well as the other
ingredients previously cited.''
Although membership in a strong coalition is a useful test of truth, it ob-
viously is not a sufficient indicator of truth. Suppose a rival hypothesis also
belongs to a strong coalition. That ought to raise doubts about the truth of
the target hypothesis. Allowing groups of coalitions to compete may thus be
a useful device for selecting true hypotheses as winners.
Until now I have directed all attention to a single dimension of epistemic
appraisal, justifiedness, and its associated standard, reliability. But what of
the other standards of appraisal that interest us: speed and power? Any
facet of a WTA mechanism that delays acceptance obviously works against
speed. For example, the larger the d-parameter, the more it takes for the
mechanism to reject rivals. On average, it will take longer for the mecha-
nism to select a winner. Similarly, the more the cognitive system seeks and
creates competing coalitions, which might give an initially favored hypoth-
esis a run for its money, the slower the system will be in selecting a winner.
Of course, both of these traits would enhance reliability; but there would be
a loss in speed.
It seems likely that the human system has a speed-sensitive design, not
just a reliability-sensitive design. In language understanding, for example,
the hearer processes sentences very rapidly, normally selecting a unique
grammatical parsing and a unique semantic interpretation for each sen-
tence, even where there are (in principle) several alternative parsings and
readings. (Of course, if there is a special language module, this fact may not
speak to doxastic decisions in other domains. But it is certainly suggestive of
a design that is oriented toward speed considerations.)
The core of a WTA device may also be seen as sensitive to considerations
of power. The fundamental impact of a WTA device is to remove clutter
from active memory. As long as rivals are competing with a given hypoth-
esis, active memory will be cluttered with items requiring processing re-
sources. This makes it more difficult to move on to other questions that
could be answered. Selection of a winner settles an issue and allows the
mind to proceed more quickly to the rest of its agenda, enabling it (in prin-
ciple) to solve a larger number of problems.
What I have said thus far shows that a WTA mechanism is cogenial to the
desiderata of speed and power, but not incompatible with achieving high
reliability. Other epistemologists, however, have not viewed the idea of
'acceptance' so favorably. Here is a kind of objection they might raise
against acceptance, against a process that selects one hypothesis as a winner
while reducing the strengths of rival hypotheses to zero.
As background, notice that many theorists are attracted by what may be
called a pragmatic criterion of evaluation (recall section 5.4). An example of
a pragmatic criterion is one that evaluates cognitive processes by their im-
pact on desire satisfaction. Decision theory stresses that a cognizer's subjec-
tive probabilities are among the factors that determine the choice of
behavior. And behavior, of course, may be more or less successful in obtain-
ing outcomes the agent desires. Now it may be suggested that an optimal
cognitive system is one that so selects behavior as to optimize (the agent's)
desire satisfaction, or payoffs. Is a WTA mechanism optimal according to
this standard?
One problematic feature of a WTA mechanism leaps out in this context.
It is in the nature of a WTA mechanism to reduce the strength of all but one
of a set of competitors to zero. Doesn't this entail a loss of critical informa-
tion, and won't this produce nonpragmatic, or nonprudential, choices? Con-
sider an example. You are driving on a two-lane road, and there is Iittle
traffic in the opposite direction. You are now caught on a hill behind a slow
truck, and you are pressed for time. Should you try passing the truck, even
though you cannot see over the hill? The sparsity of traffic traveling the op-
posite way suggests that the hypothesis 'A car is coming over the hill', is
much weaker in strength than its rival, 'No car is coming over the hill'. It
appears, then, that a WTA mechanism would reject the former hypothesis
and accept the latter. Once these doxastic decisions are plugged into your
behavioral decision matrix, it looks like a decision to pass the truck will be
made. Yet this seems likely to be a nonpragmatic, or nonprudential, choice.
In particular, if the objective probability that a car is coming is, say, .l5,
and if you act this way repeatedly in similar situations, the long-run per-
formance of these mental mechanisms will probably lead to a very bad pay-
off. This kind of case can obviously be generalized. It then begins to look as
if the WTA process of acceptance has very bad properties, at least as judged
by a pragmatic standard.
There are at least two ways of countering this objection, granting (for ar-
gument's sake) the appropriateness of a pragmatic standard. First, an ac-
ceptance process does not necessarily preclude retention of probabilistic
information. As stressed in section 15.2, information about chances may be
accepted and therefore preserved for decision-making purposes. In the
driving example you could accept the proposition 'The chance that a car is
coming in the opposite lane is .IS'. The prevalance of WTA processes, then,
does not entail the suppression of all probabilistic information. Admittedly,
such preservation of information about chances involves belief contents
that invoke an objectiue probability notion, which is philosophically con-
troversial. But if sense cannot be made of objective probabilities, it is not
clear that the prudential criticism can be properly motivated in the first
place. Certainly the form of criticism advanced above relies on objective
probabilities.
Second, there is a possible mechanism, a complementary mechanism,
that would avert the apparent shortcoming of an acceptance process. Sup-
pose there is an allied cognitive mechanism that is sensitive to possible pay-
offs with extreme values: very rewarding or very threatening outcomes. This
mechanism increases the activation level of any contemplated hypothesis h
that is believed to make some extreme payoff likely. That is, h's activation
level is increased if the cognizer believes that some extreme payoff would
be likely if h were true. Such heightened activation would tend to prevent h
from being rejected, would keep it 'in contention7, and hence not ignored
during behavioral choice. In the foregoing example the threat of a fatal ac-
cident heightens the activation level associated with 'A car is coming7.This
keeps it from being rejected, and keeps its negation from being accepted.
Thus, the choice of the 'don't pass' option is promoted. (I do not offer any
specific model of how a choice would then be made. But such models could
readily be constructed.)
I suspect that human beings do have some such mechanism. Indeed, it
may be just the mechanism that accounts for wishful thinking and fearful
thinking. The present analysis shows how such a mechanism can have pru-
dential merit--even if it has attendant liabilities for reliability. If a cognizer
believes hypotheses largely because they would, if true, fulfill his fondest
dreams or deepest fears, these beliefs are unlikely to have a high truth ratio.
But if such hypotheses are kept from being rejected and therefore ignored,
because of their potentially extreme consequences, one's behavioral deci-
sions may be prudentially more effective. Thus, an acceptance mechanism
with built-in sensitivity to extreme values may facilitate prudentially effec-
formance of these mental mechanisms will probably lead to a very bad pay-
off. This kind of case can obviously be generalized. It then begins to look as
if the WTA process of acceptance has very bad properties, at least as judged
by a pragmatic standard.
There are at least two ways of countering this objection, granting (for ar-
gument's sake) the appropriateness of a pragmatic standard. First, an ac-
ceptance process does not necessarily preclude retention of probabilistic
information. As stressed in section 15.2, information about chances may be
accepted and therefore preserved for decision-making purposes. In the
driving example you could accept the proposition 'The chance that a car is
coming in the opposite lane is .ISy.The prevalance of WTA processes, then,
does not entail the suppression of all probabilistic information. Admittedly,
such preservation of information about chances involves belief contents
that invoke an objective probability notion, which is philosophically con-
troversial. But if sense cannot be made of objective probabilities, it is not
clear that the prudential criticism can be properly motivated in the first
place. Certainly the form of criticism advanced above relies on objective
probabilities.
Second, there is a possible mechanism, a complementary mechanism,
that would avert the apparent shortcoming of an acceptance process. Sup-
pose there is an allied cognitive mechanism that is sensitive to possible pay-
offs with extreme values: very rewarding or very threatening outcomes. This
mechanism increases the activation level of any contemplated hypothesis h
that is believed to make some extreme payoff likely. That is, h's activation
level is increased if the cognizer believes that some extreme payoff would
be likely if h were true. Such heightened activation would tend to prevent h
from being rejected, would keep it 'in contention7, and hence not ignored
during behavioral choice. In the foregoing example the threat of a fatal ac-
cident heightens the activation level associated with 'A car is coming7.This
keeps it from being rejected, and keeps its negation from being accepted.
Thus, the choice of the 'don't pass' option is promoted. (I do not offer any
specific model of how a choice would then be made. But such models could
readily be constructed.)
I suspect that human beings do have some such mechanism. Indeed, it
may be just the mechanism that accounts for wishful thinking and fearful
thinking. The present analysis shows how such a mechanism can have pru-
dential merit--even if it has attendant liabilities for reliability. If a cognizer
believes hypotheses largely because they would, if true, fulfill his fondest
dreams or deepest fears, these beliefs are unlikely to have a high truth ratio.
But if such hypotheses are kept from being rejected and therefore ignored,
because of their potentially extreme consequences, one's behavioral deci-
sions may be prudentially more effective. Thus, an acceptance mechanism
with built-in sensitivity to extreme values may facilitate prudentially effec-
chapter 16
Belief Updating
16.1. Psychological Realism

How do human beings respond to new evidence that bears on previous be-
liefs they hold? How do they revise, or update, their beliefs? Are these pro-
cesses of belief updating at all satisfactory from a normative point of view? I
touched on belief revision--or, rather, the lack of revision-in Chapter 10,
in discussing Ross's work on belief perseverance. But that work did not pur-
port to provide a general model of belief updating. In this chapter I present
a general model of updating and examine its normative implications.
The most popular model of updating is Bayesian conditionalization.' Its
popularity is primarily in its normative or prescriptive quality, but it also
has appeal as a descriptive model, perhaps because of the dearth of system-
atic alternatives. However, conditionalization as a descriptive model of up-
dating seems to presuppose that human beings are natively equipped for
performing highly sophisticated probabilistic calculations, or for simulating
such calculations. My discussions in Chapters 14 and 15 raise serious doubts
about this. I therefore want to examine a recent and more plausible model
of belief updating, one proposed by Hillel Einhorn and Robin ~ o ~ a r t h . '
This has the advantage of much greater psychological realism. It can be
taken seriously as a description of basic psychological processes people use
to change their degrees of belief in response to new evidence.
The Einhorn-Hogarth model assumes degrees of belief that can be scaled
on the unit interval, something like subjective probabilities. I raised objec-
tions to subjective probabilities in Chapter 15. However, in the present
chapter I shall waive these objections and go along with the acceptance of
such states, in order to present and weigh the implications of the Einhorn-
Hogarth model. Although that model presupposes (something like) subjec-
tive probabilities, it does not assume manipulation of such probabilities in
accordance with Bayes's Theorem or other laws of the probability calculus.
So it has at least greater realism than standard models, in my view.
It should be emphasized that the nonrealization of conditionalization at
Belief Updating 345
the level of processes is compatible with the use of conditionalization-or

other uses of the probability calculus-at the level of methods. If someone
learns Bayes's Theorem and a rule (such as the Jeffrey rule) of conditionali-
zation, and if that person has, say, (ungraded) beliefs in objective probabili-
ties, he can certainly apply his acquired rule to arrive at new objective
probability estimates. I am not disputing a psychological capacity to learn
and apply probabilistic rules in this fashion. But here, as elsewhere in the
book, I am concerned in the first instance with native processes; and it is
doubtful that our native processes involve anything like Bayesian condi-
tionalization.
16.2. The Anchom'ng-and-Adiustment Model

Einhorn and Hogarth propose that people revise their beliefs by a process
of adjustment from an anchor. One's current attitude toward a given hy-
pothesis provides an anchor, from which adjustments are made upon re-
ceipt of new information. Once the adjustment is accomplished, the new
attitude becomes the anchor for any subsequent information. The crucial
idea in their model is that adjustments work by means of a "contrast," or
"surprise," effect. The stronger one's antecedent belief in h, the greater the
downward adjustment if new negative evidence is received. To borrow a
boxing metaphor, the bigger one's initial confidence, the more it will fall.
Analogously, the weaker one's antecedent belief in h, the greater the up-
ward adjustment if new positive evidence is received. This contrast as-
sumption leads to some interesting predictions, borne out by experimental
data.
To explain the Einhorn-Hogarth model in more detail, I will start with
their model for adjustments upon receipt of negative evidence. This is what
they call their "discounting" model. Let
So = the initial strength of belief in some hypothesis (0<So 6 1);

s(ak)= the strength of the kth piece of negative evidence, ak (0 s s(ak)
s 1);
Sk = the strength of belief in the hypothesis after k pieces of negative evi-
dence (0S Sk S l).
It is assumed that people have some way of assessing whether new evi-
dence is positive or negative and quantifying the strength of that evidence
(assigning some degree of confirmatory or disconfirmatory import to the
evidence). Einhorn and Hogarth make no specific suggestions about this
process, and I too shall simply take it as a given.
To illustrate the process, consider the effect of the first piece of negative
evidence, al, on one's position. Assuming that
346 Belief Updating
wo = an adjustment weight associated with the initial strength of belief,

the resulting strength of belief, S1, after the first piece of negative evidence
is given by
The adjustment weight reflects the importance given to negative evidence

in the discounting process. But the adjustment weight is a function of the
anchor at the given time. (Note that the anchor and adjustment weight have
the same temporal subscript.) The rationale for this is explained as follows.
Imagine that your initial confidence is weak and a strong piece of negative
evidence is revealed. Since your confidence is already low, the new infor-
mation cannot reduce So very much (in absolute terms). But if your initial
confidence in the hypothesis is very high, the reduction of strength by the
same piece of evidence would be larger. This is the basis of the contrast, or
surprise, component of the model.
Reconsider equation (1) and ask what happens when a second piece of
negative evidence is received. The new position is given by
Equation (2a) can be generalized as follows:
Now consider the functional relation between the strength of the anchor
and the adjustment weight (the "adjustment weight function"). It is as-
sumed that bigger anchors have larger adjustment weights. This implies
that the adjustment weight function is a monotonically increasing function
of the size of the anchor. To illustrate this, Einhorn and Hogarth posit a
simple form often found in psychophysical judgments, that is,
In this equation a is a parameter that reflects one's attitude toward discon-

firming evidence. Its value may vary across individuals. To appreciate the
role of a , consider Figure 16.1. Note that the adjustment weight is mono-
tonically increasing with the size of the anchor regardless of the value of a.
When a hypothesis is believed to degree 0, it cannot be adjusted below 0.
When the anchor is 1,the adjustment weight is 1; so a 'certain' belief is ad-
justed solely by the strength of negative evidence. When a > 1, the adjust-
ment weight is less than the anchor. (Since the degree of belief is typically
less than.1, an exponent larger than 1makes the adjustment weight smaller
than does an exponent smaller than I.) As a increases beyond 1, and the ad-
justment weight therefore gets smaller, negative evidence receives corre-
Belief Updating 347
Disconfirm
prone
. Disconfirm
neutral
, Disconfirm
avoiding
Sk-1
Figure 16.1
spondingly less weight in the discounting process. People with an a

greater than 1 are labeled as "disconfirmation avoiding," since they give
relatively less weight to negative evidence. When 0 s a < 1, adjustment
weights are large and initial positions are strongly discounted by negative
information. People with such attitudes are "disconfirmation prone." When
a = 1, a "disconfirmation neutral" attitude is assumed.
The full discount model is now obtained by substituting (4) into (3) to ob-
tain:
The discount model implies that the strength of a belief after the kth piece
of negative evidence is a function of three factors: (1) the size of the anchor,
Sk-l: (2) the strength of negative evidence, s(ak); and (3) an adjustment
weight, which is a function of the anchor and one's attitude toward discon-
firming evidence (a).
A model for adjustments in response to new positive evidence is exactly
parallel to the discounting model. Einhorn and Hogarth call it the "accre-
tion" model. The form of the model is given by
where
rk-l = the adjustment weight for positive evidence
and
s(bk)= the strength of the kth piece of positive evidence, bk (0< s(bk)
s 1).
348 Belief Updating
Equation (6) follows the same general form as the discount model except
that the final position results from an anchoring and upward adjustment
process. The basic assumption in the accretion model is that weak beliefs
are increased more by positive evidence than are strong beliefs. The same
positive evidence 'helps' a weaker position more than a stronger one. As in
the discounting model, Einhorn and Hogarth posit a simple form to capture
the monotonically decreasing relation between rk-l and s ~ - In
~ .particular,
they propose the following relation:
rk- * = (1 - (P 2 0). (7)
The adjustment weight function is shown in Figure 16.2. The parameter P is
interpreted as reflecting one's "attitude toward confirming evidence."
When P > 1,confirming evidence is given a relatively small weight. In fact,
as increases, the weight for confirming evidence approaches 0.When 0 G
/3 < 1, confirming evidence is given a large weight, reflecting a "confirma-
tion prone" attitude. When P = 1, the attitude is labeled as "confirmation
neutral." Obviously, the role of P in the accretion model is directly analo-
gous to the role of a in the discount model. The full accretion model is ob-
tained by substituting (7) into (6):
To develop an updating model for both positive and negative evidence,

Einhorn and Hogarth simply combine the discount model for negative evi-
dence and the accretion model for positive evidence. The resulting mixed
evidence model uses whichever adjustment process is appropriate for the
Confirm
prone
Confirm
neutral
Confirm
avoiding
Figure 16.2
Belief Updating 349
0 1 2 3 k
Figure 16.3
evidence at hand. When the evidence is negative, the discount model is

used; when the evidence is positive, the accretion model is used.
Let us now consider some implications of the model for the kinematics of
belief. First, consider the convergence of beliefs when two people have dif-
ferent initial positions but receive the same evidence. Start with the receipt
of negative evidence. This can be illustrated by Figure 16.3. When k = 0,
the initial positions for persons A and B are shown as SAand SR,respec-
tively. If A and B have the same attitude toward disconfirming evidence,
note what happens after receipt of the first piece of negative evidence.
Since person A has a higher initial position, the decrease in his or her belief
is greater than that for person B. This is because bigger anchors have larger
adjustment weights. Indeed, the slopes of the lines connecting Skover k (for
the same strength of evidence) will decrease at a faster rate for A than for B.
This means that at each successive adjustment, the difference in final posi-
tions gets smaller and will eventually converge.
Although convergence of beliefs occurs when persons have different '
starting points and the same attitude toward disconfirming evidence, diver-
gence can occur when the a's are not equal. Consider person C, whose ini-
tial position is the same as A's but whose a is very high. Since C gives little
weight to negative evidence, Figure 16.3 shows his or her beliefs to be un-
changed (or very slightly changed) over the k pieces. This leads to diver-
gence of beliefs between A and C given the same evidence. However, as k
increases, the two positions will converge toward zero unless C gives zero
weight to negative evidence. Analogous points about convergence and non-
convergence hold for the receipt of positive evidence under the accretion
model.
Moving to the consequences of mixed evidence, a major implication of
the contrast approach is strong recency efiects in belief change. To see why,
350 Belief Updating
consider the strength of belief after receiving one piece each of positive and
negative information. Consider the difference between the belief strength
after receiving them in the positive-negative order, S(+,-), as compared
with the strength after receiving them in the negative-positive order,
S(-,+). Figure 16.4 shows the effects of these two orders at different start-
ing levels, So, and So,. When the initial belief is at So,, the two orders are
shown in the top half of the figure. Compare the effects of the (+,-) and the
(-,+) orders. Note that the slope of the line connecting Sk-Oand Sk-l in
the (-,+) order is less steep than when negative evidence occurs in the
(+,-) order. The reason is that the same negative evidence has a larger
discounting weight after the positive evidence because of the contrast ef-
fect. Similarly, the slope of Skll to Ska2in the (-,+) order is steeper than
the slope for positive evidence from the initial position. These differences
in slopes lead to crossing lines that resemble fish tails. The fish-tail pat-
tern implies recency effects since the final position after the (+,-) order
is lower than for the (-,+) order. The prediction of recency effects does
not depend on one's initial position, as illustrated in the lower half of the
figure.
Another implication of the model concerns the relative effect of present-
ing multiple pieces of evidence simultaneously instead of sequentially. Si-
multaneous presentation of two or more pieces of negative evidence, for
example, results in greater discounting than presentation of these same
pieces one after another. Thus, the contrast or surprise effect is 'diluted' by
sequential processing.
Experiments done by Einhorn and Hogarth, and by J. D. Shanteau, sup-
port these order-effect implications. In two of Einhorn and Hogarth's own
experiments, subjects were given four scenarios, each scenario consisting of
an initial description (the "stem") and four additional pieces of information
Figure 16.4
Belief Updating 351
presented in separate paragraphs. The scenarios concerned a defective

stereo speaker thought to have a bad connection, a baseball player whose
hitting had dramatically improved after a new coaching program, and so
on. In each case the stem provided information concerning a target causal
hypothesis, for example, whether the coaching had caused the player's hit-
ting improvement. At the end of the stem, subjects were asked to rate the
likelihood of the suspected factor being the cause of the outcome, on a rat-
ing scale from 0 to 100. After responding to this question, subjects were
presented with four pieces of additional information regarding the causal
hypothesis. These consisted of two pieces of positive information and two
pieces of negative information. The new information was presented in
either a positive-negative order (+,+,-,-) or a negative-positive order
(-,-,+,+). After this new information was presented, subjects were again
asked to rate the likelihood that X caused Y.
The order in which information was presented had a large recency effect,
as the theory predicts. The mean change in belief is .I55 for the positive-
negative order, while the mean change in belief for the negative-positive
order is .057. This was replicated in a slightly different design, where the
mean change for (+,+,-,-) was .I84 while the change for (-,-,+,+) was
.088.
Einhorn and Hogarth also reanalyze data from an experiment by Shan-
tea^.^ Shanteau's tasks involved probabilistic inference and probabilistic
estimation on the basis of information drawn from urns containing propor-
tions of colored beads. These tasks tested both the recency effect hypothesis
and the dilution effect hypothesis (that sequential processing of the differ-
ent components of a body of evidence will have a smaller effect on beliefs
than simultaneous presentation of that same evidence). The data from
Shanteau's experiment bear out both of these hypotheses, as illustrated in
Figures 16.5 and 16.6. Figure 16.5 shows the prevalance of fish tails, indi-
cating recency effects throughout these data. Figure 16.6 shows that sirnul-
taneous presentation leads to more extreme responses for both positive and
negative evidence.
Among the large set of order-effect data obtained by numerous re-
searchers, there are data indicating primacy as well as recency effects. In
order to account for primacy effects, Einhorn and Hogarth add another ele-
ment to their model: attention. To understand this element, imagine read-
ing a long and complex book. At the outset, attention is high, fatigue is low,
and information is scrutinized carefully. As one continues, however, atten-
tion may begin to wane and information that occurs later in the book re-
ceives less careful consideration. Such a process, in which attention
decreases with serial position, means that evidence occurring early has
more influence than the same evidence would if it occurred late. The pri-
macy effect-the greater impact of early information-may therefore be
explained by "attention decrement." This idea has been developed in detail
352 Belief Updating
ESTIMATION INFERENCE
C
z
W
5
C3
n
3
>.
k
-r
m
a
m
0
LT
a
z
a
W
5
0 1 2 3 4 0 1 2 3 4
k k
Figure 16.5
by N. H. ~ n d e r s o nEinhorn
.~ and Hogarth proceed to incorporate the at-
tention-decrement factor into their theory.
I will not pursue the details of this theoretical refinement. Suffice it to say
that attention decrement works in the opposite direction from the contrast
or surprise factor in that the latter leads to recency effects, whereas atten-
tion decrement leads to primacy effects. The net effect of these opposing
EXP. 1 EXR 2
Simul t.
fq.
4 Neqs. 4 Pos. 4 Negs. 4 Pos.
Figure 16.6
Belief Updating 353
forces can result in primacy, recency, or neither, depending on the exact

values of the relevant parameters.
16.3. lwtijicatwn and Calibratwnal Convergence

Let us now inquire into the normative status of the anchoring-and-adjust-
ment process that Einhorn and Hogarth postulate. Can such a process, if it
indeed exists as a basic psychological process, confer justification? Would it
be permitted by a right system of J-rules?
A complaint might immediately be lodged against the anchoring-and-
adjustment (A-and-A) process: it produces order effects. Surely, the objec-
tion would run, the order in which evidence is presented should make no
difference to the resulting doxastic states. 'No order effects' ought to be a
formal, a priori principle of adequacy for any belief-updating process. But
the A-and-A process clearly violates this principle. It can yield either recency
or primacy effects, and it can produce the dilution effect, in which sequen-
tial presentation of multiple pieces of evidence has less impact on belief
change than simultaneous presentation. Given these 'biases', the A-and-A
process should be ruled out immediately as justificationally unacceptable.
This objection would be right if a formalist principle like 'no order ef-
fects' were appropriate. But I am very dubious about the validity of such a
principle. Various formal principles of rationality have been proposed in
the philosophical and social scientific literature, principles for the rational-
ity of belief and the rationality of choice. But I am unpersuaded of their
(unconditional) correctness. Two such formal principles encountered in
earlier chapters were noninconsistency and probabilistic coherence. In both
cases I gave reasons for rejecting these principles as (categorical) prin-
ciples of rationality or of justification. Another example of a formal prin-
ciple, in the theory of rational choice, is the requirement of transitivity of
preference. It too has been called into question, at least in social choice
theory.5
Clearly, the general approach to justifiedness I favor, verific consequen-
tialism, does not necessarily imply such 'formalist' principles. Principles re-
quiring noninconsistency or probabilistic incoherence do not function as
high-level, categorical principles in my theory. At most they would have to
be derived (as lemmas) from the more fundamental, verific consequentialist,
family of criteria. However, such derivations, as we saw in previous chap-
ters, are very much in doubt. Similarly, I am dubious about the derivability
of the no-order-effects principle. It is not obvious how it would follow from
the verific consequentialist criterion I have defended, in particular, from
the good-calibration variant that applies to degrees of belief. Nor does it
follow from a new variant of this approach that I shall introduce shortly. It
remains to be asked whether the A-and-A process, in particular, is compati-
354 Belief Updating
ble with these calibrationist criteria of justifiedness. This is the topic to

which we should now turn.
Before addressing this question, I should mention parenthetically that
order effects associated with the contrast or surprise process can be miti-
gated by the attention-decrement process. This is an intriguing observation
made by Einhorn and Hogarth. As noted at the end of section 16.2, the con-
trast or surprise factor and the attention-decrement factor are forces work-
ing in opposite directions, the former toward recency effects and the latter
toward primacy effects. If the bias from the contrast or surprise process
were exactly offset by the bias from attention decrement, the net effect
might be zero. Perhaps, then, the contrast or surprise process does not nec-
essarily breed (net) recency effects. However, it would be remarkable if the
two sorts of biases exactly offset one another.
Let us now ignore the no-order-effects principle and ask how the contrast
or surprise process, or more generally, the entire set of A-and-A processes,
would fare under the good-calibration criterion. Recall that good calibra-
tion was defined for a system of J-rules. Whether such a system is well cali-
brated is a function of the partial beliefs, or subjective probabilities, that
would be produced if cognizers used processes that conform with that sys-
tem. Consider the set of partial beliefs that would be so produced, and par-
tition them into separate classes for each degree of belief D (or classes
spanning small ranges of degrees of belief). The system is well calibrated
just in case the proportion of truths within each class is approximately the
same as the degree of belief of that class. For example, given the class of
subjective probabilities equaling .60 that would be produced in a realiza-
tion of the rule system, approximately 60 percent of the propositions as-
signed this subjective probability must be true.
The main point to be made here is that there isn't any reason why an A-
and-A process described by Einhorn and Hogarth could not perform ade-
quately, as judged by the good-calibration criterion. Hence, there isn't any
reason why it could not be approved by a right J-rule system. Note in par-
ticular that the good-calibration criterion only requires that for each
strength of belief D, the truth ratio of beliefs of strength D that would result
from an approved set of processes should be approximately D. I see no rea-
son why an A-and-A process of belief revision could not satisfy this require-
ment, especially if the margin of approximation is fairly loose.
Of course, the generic A-and-A process is not solely responsible for
achieving proximity to good calibration. This depends on at least two other
factors. First, it depends on the processes for assessing evidence: processes
for deciding (a) whether a given piece of evidence bears positively or nega-
tively on the target hypothesis, and (b) how weighty a piece of evidence it
is. Einhorn and Hogarth's model presupposes such decisions but is silent
about their details. Obviously they are critical in determining the calibra-
tional adequacy of belief updating procedures. Second, calibrational ade-
Belief Updating 355
quacy also depends on the values of a and P, parameters of the adjustment

weight functions. I will focus on this second factor.
Recall that different values of a and ,8 represent different attitudes to-
ward disconfirmation and confirmation respectively. Extreme values of a
and ,8 represent either extreme proneness toward disconfirmation or confir-
mation, or extreme avoidance of disconfirmation or confirmation. Any of
these extreme values might militate against good calibration.
Reflecting on this point, the general A-and-A model should be construed,
for our purposes, as picking out not a single process, but rather a family of
belief-updating processes. Members of the family will differ in terms of dif-
ferences in attitudes toward confirmation and disconfirmation. It is unlikely
that all members of this family would perform well as judged by the good-
calibration criterion. But some members of this family might perform ade-
quately. Thus, some A-and-A processes, or contrast or surprise processes,
could be justificationally admissible according to verific consequentialism.
People who use these A-and-A processes can have justified (partial) beliefs.
But partial beliefs updated by extreme confirmation proneness, or by ex-
treme disconfirmation avoidance, will not be justified.
Some comments on Bayesian conditionalization might be apt at this
juncture. There is nothing in my account which suggests that Bayesian con-
di tionalization would not also be a justificationally acceptable process. If
Bayesian conditionalization were a basic psychological process of belief up- ,
dating, it could be sanctioned by a right J-rule system. (Recall that my me-

tatheory allows multiple right J-rule systems.) Of course, the mere use of
Bayesian conditionalization would not by itself ensure good calibration.
Highly inaccurate conditional probabilities could lead to poor calibration,
despite conformity with conditionalization rules. But conditionalization per
se might be justificationally admissible; that is, combined with suitable pro-
cesses of forming conditional subjective probabilities, it might produce
good calibration.
As I have said, however, there is good reason to believe that Bayesian ,
conditionalization is not a native psychological process. And as I have

stressed, it is unlikely that human beings have the native capacity to make
probabilistic calculations of the sort conditionalization requires. Further-
more, it is very doubtful that conditionalization rules would be generally
applicable. conditionalization requires that a cognizer have a prior assess-
ment of the conditional probability of the target hypothesis on the evidence
in question. But use of conditionalization then requires that a cognizer have
a conditional probability assignment for any piece of evidence he might re-
ceive in the future, which is certainly unrealistic. Nobody can anticipate all
possible evidence that might be encountered.
Returning now to A-and-A processes, advocates of the no-order-effects
principle might interject the following point. Granted that some members
of the A-and-A family might fit into a well-calibrated rule system, couldn't a
356 Belief Updating
belief-updating procedure that did not feature the contrast or surprise

mechanism help form an even better calibrated system? In other words,
wouldn't a process with no order effects be even a better component of a
coordinated set of processes, according to calibration consequentialism?
This may be true (though it would have to be established). Even so, two
replies are in order. First, although an abstractly possible process of this sort
might be better, human beings may possess no such process-not, at any
rate, as part of their native cognitive equipment. Second, the criterion of
rightness for J-rule systems is not comparative. Even if one rule system were
better calibrated than another, it would not follow that the less well-
calibrated one is not right. According to my criterion, a J-rule system is
right if it is sufficientlywell calibrated. It cannot be demoted from this sta-
tus by finding a better calibrated system (especially one that licenses pro-
cesses that are just abstractly possible). True enough, one might try to refine
my approach so as to admit 'degrees' of rightness or degrees of justifiedness.
Then superior calibrational properties might be used to rank alternative
sets of processes. But even this would not mean declaring order-effect pro-
cesses wholly nonjustificational.
I have said that some members of the A-and-A family might be justifica-
tionally admissible according to calibration consequentialism. But further
reflection makes this somewhat doubtful. Indeed, no belief updating process
might be admissible by this criterion. This suggests that a revised criterion
of J-rule rightness might be in order.
The difficulty is created by the fact that a person might encounter only
one or two pieces of evidence, perhaps 'small' pieces of evidence, that bear
on hypothesis h. One's subjective probability vis-A-vis h will therefore be
based on very slender grounds. Now consider a large class of such cases.
Even if cognizers respond to such evidence with one of the better members
of the A-and-A family of processes (that is, one with moderate a and /3 pa-
rameters), is it really likely that the resulting truth ratio of their beliefs will
be well calibrated? That is doubtful. Just one or two applications of even a
good belief-updating process may not lead in the direction of truth, espe-
cially if the evidence is quite weak. This raises doubts about whether any
A-and-A process would qualify for justificational approval under calibration
consequentialism. Moreover, this is a potential problem for any belief-up-
dating process! It is not peculiar to the family of A-and-A processes. Yet it
seems unreasonable to preclude all such processes. If repeated use of a pro-
cess would lead to good calibration, or would gradually converge on good
calibration, that ought to suffice to vindicate the process. The process ought
not be (justificationally) prohibited just because one or two applications of it
would not yield good calibration.
One possible way to meet this difficulty is to point out that a string of uses
of a given A-and-A process might be viewed as a single process. N applica-
Belief Updating 357
tions of the A-and-A process is a large conjunctive process, and where N is

large enough, such a process might make a suitable contribution to good
calibration. However, it is not clear that this would work for any fixed N.
Since pieces of evidence can vary widely in relevance, there may be no
fixed number of applications that will do the job.
An alternative solution is to introduce a different variant of the calibra-
tion approach: calibrational convergence. A set of process rules will
be called calibrationally convergent if and only if the more frequently they
are applied, the closer the resulting truth ratio of beliefs to perfect calibra-
tion. (A more precise definition could be given using the mathematical
notion of a limit.) We might then propose that a system of J-rules is right if
it is calibrationally convergent. This requirement is obviously weaker than
our original calibration criterion and might be more attractive for that
reason.
A possible refinement of this approach would introduce speed of conver-
gence as a justificationally pertinent factor. If two sets of rules are both ca-
librationally convergent but the use of one would converge faster than the
other, the former might plausibly be deemed epistemologically preferable.
Although my metatheoretical structure does not explicitly accommodate
degrees of justification or comparative justifiedness, such notions are intui-
tively quite plausible. We might say that (partial) beliefs formed in accord-
ance with the former set of rules are more justified than beliefs formed in
accordance with the latter. For example, beliefs that result from updatings
by an A-and-A process with optimal values of a and /3 may be called more
justified than beliefs that have been updated by an A-and-A process with
suboptimal a and /3 values. But even the latter may be justified to some de-
gree.
I shall not try to develop this theme in detail. However, let me comment
on its kinship with sundry other approaches in the philosophy of science
and the theory of epistemic norms. Many philosophers of science have sug-
gested that even if good scientific method need not bring us to the truth, its
repeated use should at least bring us closer to the truth. C. S. Peirce is fa-
mous for this suggestion, although he went on-very implausibly-to define
truth as the (set of) opinion(s) toward which inquiry leads. My idea of cali-
brational convergence as a test for justifiedness is a cousin of this suggestion.
Another close relative of my proposal is Hans Reichenbach's "pragmatic"
justification of induction6 Reichenbach appealed to the possibility of an in-
ductive method converging toward a true value, or producing gradually
better approximations through repeated use.
Karl Popper's theory of "corroboration" is another case in point. A the-
ory is better corroborated the more often it survives attempts to falsify it.
To be sure, Popper stresses that degree of corroboration is a function not
just of the number of corroborating instances, but also of their severity.'
358 Belief Updating
Still, the core idea of making an epistemic judgment rest on successive ap-
plications of a method is present here too.
The diachronic approach to epistemic evaluation has also been popular-
ized by Imre Lakatos. He proposed that we appraise scientific research
programmes by their historical progressiveness.8 On Larry Laudan's per-
mutation of this idea a research tradition may be evaluated by the "general
progress," or the "rate of progress," of a research tradition in solving prob-
lems over time.'
I do not wish to endorse any of these specific proposals in the philosophy
of science literature. I merely call attention to the fact that the kernel of my
idea-judging a method by its repeated use--has enjoyed significant popu-
larity among philosophers.
An emphasis on diachronic properties of a process or method is also
urged by Robin Hogarth.lo Hogarth argues that the biases of judgmental
processes revealed in "discrete," isolated incidents may be outweighed by
the merits of those processes when they are used "continuously" (repeat-
edly) over time, with feedback.
Although I shall not develop the calibrational convergence approach in
further detail, I regard it as significant for two reasons. First, it brings into
prominence an important way of evaluating judgmental processes and
methods, namely, evaluating their success over time, through repeated ap-
plication. Second, it shows the versatility of the general reliabilist and cali-
brationist approach to epistemic norms. If the criterion of rightness
proposed in Chapter 5 does not strike one as fully correct, the verific ap-
proach still allows for variations in many directions. Calibrational conver-
gence is one such promising direction.
chapter 17
Production Systems and Second-Order

Processes
1 7.1. Anderson's Production System Theory

In this chapter I focus on a general framework for describing cognition: the
production system framework. I choose this framework for two reasons.
First, it provides a very general approach to cognition, and it is worth ask-
ing how our standards of evaluation can be applied to such a general ap-
proach. Second, it is especially useful in discussing a topic introduced in
Part I but heretofore neglected in Part 11, namely, second-order processes.
We saw in Part I that both knowledge and justifiedness depend on how be-
lief-forming processes, or methods, are learned or acquired. But thus far I
have said very little about processes governing such learning or acquisition.
The production system framework provides an excellent sample treatment
of such second-order processes, as I will show in section 17.3.
The basic claim of the production system framework is that human cog-
nition may be conceptualized in terms of a set of condition-action pairs,
called 'productions'. A condition specifies some data patterns, and if ele-
ments matching these patterns are in working memory, then the production
can apply. The action component of a production specifies what to do in
that state. The standard action is to add new data elements to working
memory. A typical production rule, stated informally, might be:
IF person 1 is the father of person 2
and person 2 is the father of person 3
THEN person 1 is the grandfather of person 3.
This production would apply if 'Fred is the father of Bill' and 'Bill is the fa-
ther of Tom' were active in working memory. The system would make the
inference to 'Fred is the grandfather of Tom' and deposit this fact in work-
ing memory. This may be understood as forming a belief.
Production systems can be traced back to the proposals of E. L. post,' but
modern production systems began with A. Newell's work in the early six-
ties, formulated later in Newel1 and H. A. Simon's Hunurn Problem ~ o l u i n ~ . ~
360 Production Systems and Second-Order Processes
From the start, production systems have had an ambiguous status, being in
part programming languages for computer science and in part psychologi-
cal theories. I am interested in them, of course, in the latter guise. I shall
concentrate on perhaps the best developed psychological theory within the
production system tradition: that of John Anderson in The Architecture of
Anderson started his work in this mode in an earlier book, Lan-
guage, Memory, and Thought, where he called his system ACT."^ The
more recent version, which I shall examine, is called "ACT'."
To illustrate the idea of a production system, Anderson provides a hypo-
thetical set of productions, usable for doing addition. Let's see how such a
set of productions might be deployed to solve the following addition prob-
lem:
Here is the hypothetical set of productions.

P1 IF the goal is to do an addition problem
THEN the subgoal is to iterate through the columns of
the problem.
P2 IF the goal is to iterate through the columns of an addition
problem
and the rightmost column has not been processed
THEN the subgoal is to iterate through the rows of that
right-most column
and set the running total to 0.
IF the goal is to iterate through the columns of an
addition problem
and a column has just been processed
and another column is to the left of this column
THEN the subgoal is to iterate through the rows of
this column to the left
and set the running total to the carry.
IF the goal is to iterate through the columns of an addition
problem
and the last column has been processed
and there is a carry
THEN write out the carry
and POP the goal.
P5 IF the goal is to iterate through the columns of an addition
problem
Production Systems and Second-Order Processes 361
and the last column has been processed and there is no

carry
THEN POP the goal.
P6 IF the goal is to iterate through the rows of a column
and the top row has not been processed
THEN the subgoal is to add the digit of the top row to the running
total.
and a row has just been processed
and another row is below it
THEN the subgoal is to add the digit of the lower row to the run-
ning total.
and the last row has been processed
and the running total is a digit
THEN write the digit
and delete the carry
and mark the column as processed
and POP the goal.
and the last row has been processed
and the running total is of the form 'string digit'
THEN write the digit
and set carry to the string
and mark the column as processed
and POP the goal.
P10 IF the goal is to add a digit to another digit
and a sum is the sum of the two digits
THEN the result is the sum
and mark the digit as processed
and POP the goal.
P11 IF the goal is to add a digit to a number
and the number is of the form 'string digit'
and the sum is less than 10
THEN the result is 'string sum'
and POP the goal.
P12 IF the goal is to add a digit to a number
and the number is of the form 'string digit'
and the sum is of the form '1 digit'

and another number sum0 is the sum of 1 plus string
THEN the result is 'sum* digito'
and POP the goal.
The conditions of these productions are given by the I F parts and their
actions by the THEN parts. Figure 17.1 illustrates the flow of control
among the productions in that set. Application of this set of productions is
controlled by the setting of goals. Figure 17.1 basically shows which pro-
ductions respond to which goals and which productions set which goals.
Applied to the addition problem mentioned above, application of the pro-
duction system would run as follows. Production P1, the first to apply, sets a
subgoal to iterate through the columns. Production P2 then changes the
subgoal to add the digits of the rightmost column and sets the running total
Figure 17.1. A representation of the flow of control among the various goals. The
boxes correspond to goal states and the arrows to productions that can change these
states. The goal at the origin of the arrow corresponds to the goal that elicits the
production, and the goal at the terminus of the arrow corresponds to the goal that
is set after the application of the production. Control starts with the top goal.
to 0.Then production P6 sets the new subgoal to add the top digit of the
row (4) to the running total. In terms of Figure 17.1 this sequence of three
productions has moved the system down from the top goal of doing the
problem to the bottom goal of performing a basic addition operation. The
system has four goals stacked, with attention focused on the bottom goal.
At this point production P10 applies, which calculates 4 as the new value
of the running total. In doing this it retrieves from the addition table (in
+
LTM) the fact that 4 0 = 4. Production PI0 also pops the goal of adding
the digit to the running total. 'Popping' a goal means shifting attention from
the current goal to the one above it in the hierarchy. In this situation atten-
tion will return to iterating through the rows of the column. Then P7 ap-
plies, which sets the new subgoal of adding 8 to the running total. PI0
applies again to change the running total to 12, then P7 applies to create
the subgoal of adding 3 to the running total, then P11 calculates the new
running total as 15. At this point the system is back to the goal of iterating
through the rows and has processed the bottom row of the column. Then
production P9 applies, which writes out the 5 in 15, sets the carry to the 1,
and pops back to the goal of iterating through the columns. At this point the
production system has processed one column of the problem. This should
suffice to illustrate the application of these productions.
In Anderson's ACT" theory three memories are postulated: working
memory, declarative memory, and production memory. Working memory
contains the information that the system can currently access, consisting of
information retrieved from long-term declarative memory as well as tem-
porary structures deposited by encoding processes and the actions of pro-
ductions. Working memory is declarative knowledge, permanent or
temporary, that is in an active state. Declarative memory contains declara-
tive knowledge that may not be in an active state. Production memory con-
tains the set of productions, or procedures, available to the system. In a
mature human cognizer, Anderson puts the number of productions some-
where between the tens of thousands and the tens of millions.
The interactions between these memories is illustrated in Figure 17.2.
Most of the processes depicted here involve working memory. Encoding
processes deposit information about the outside world into working mem-
ory; performance processes convert commands in working memory into be-
havior. The storage process can create permanent records in declarative
memory of the contents of working memory and can increase the strength
of existing records in declarative memory. The retrieval process retrieves
information from declarative memory. In the rnatch process data in work-
ing memory are put into correspondence with the conditions of produc-
tions. The execution process deposits the actions of matched productions
into working memory. The whole process of production matching followed
by execution is referred to as production application. Note that the arrow
APPL'CATloN
DECLARATIVE PRODUCTION
MEMORY
WORKING
MEMORY
ENCODING PERFORMANCES
T
OUTSIDE WORUI
Figure 17.2
marked 'application' cycles back into the production memory box, reflect-
ing the fact that new productions are learned from studying the history of
application of existing productions. This involves the idea of procedural
learning, which I shall explore in sections 17.3 and 17.4.
I will not summarize all the basic principles of ACT*, but let me mention
a few salient ones. Since all activities of productions require activation-
only activated data can be matched with production condition+an im-
portant question concerns the sources of activation. What Anderson calls
"source nodes" are the "springs" from which activation flows throughout
the system. Nodes can become source nodes in several ways. First, they can
be created by the perception of objects in the environment. Second, ele-
ments deposited in working memory by a production's action become tran-
sient source nodes. These are "internal" sources. They stay active for a brief
period of time and then cease to be sources..A special goal element, how-
ever, permits one internal source that is not so transient, and this enables
the system to maintain focus on a current goal.
A critical issue concerning the flow of cognitive activity is the selection of
productions that get matched. Given what has been said thus far, it would
seem possible for incompatible productions to match data in working mem-
ory, and for all to get applied. But this is assumed to be psychologically un-
realistic. Rival productions are not allowed to apply simultaneously. What,
then, are the principles for conflict resolution, for choosing the production
that does get applied? ACT0 involves several such principles.
First, ACT0 allows the conditions of a production to be applied when its
condition is matched only partially, not fully. However, a full match is pre-
ferred to a partial match, and among partial matches, more complete
matches are preferred to less complete matches.
Any number of phenomena, Anderson says, point to the fact that produc-
tions can be evoked by partial matches. Faces change but still can be recog-
nized. We can recognize partial lexical patterns, as in Figure 17.3, and this
seems best explained by the concept of partial matching. Many errors de-
scribed by Donald or man' also can be explained by partial matching. For
instance, there is the case of a person who threw a dirty shirt by mistake
into the toilet rather than into the hamper. Apparently a 'hamper-match-
ing' pattern partially matched the toilet.
A second principle for production selection concerns the relative
strength of rival productions. Stronger productions are preferred to weaker
productions. The determinants of production strength will be examined in
section 17.3.
A third principle for production selection is the specificity principle. It
states that when two productions match the same data, preference is given
to the production with the more specific condition. Condition A is more
specific than condition B if A matches in a proper subset of the situations
where condition B would match. This principle handles cases of exceptions
to rules, as in productions P1 and P2 below.
P1 IF the goal is to generate the plural of a noun
THEN say the noun s.
P2 IF the goal is to generate the plural of man
THEN say men.
Production P1 gives the general rule for pluralizing a noun, while P2 gives
the exception for man. The specificity principle enables the exception to
take precedence over the general rule. There are two ways in which one
production can be more specific than another. One possibility is that the
more specific one contains additional clauses in its condition. The second is
that the production contains additional tests about a data structure, as in
the foregoing example.
A fourth principle concerning production selection involves goal domi-
nance. Many productions contain in their conditions tests for the current
Figure 17.3
goal. There can be only one current goal active in the system at a time. Pro-
ductions that refer to the current goal take precedence over productions
that might otherwise apply. A goal-directed production that matches the
current goal applies more rapidly and reliably, but if it does not match the
current goal, it will not apply no matter how well the rest of its condition
matches. If a number of goal-directed productions match the current goal,
the principles of degree of match, strength, and specificity select the one
that will apply.
These are some of the main features of ACT*. More of its features will be
presented in section 17.3. Let me now turn (briefly) to some evaluative re-
flections on the features of ACT* introduced thus far.
17.2. Production Architecture: Reliability, Speed, and Power

Are productions belief-forming processes? It might seem as if they are. Ap-
plication of a production often involves the insertion of new data structures
into working memory. Such an insertion seems to be the formation of a be-
lief (at least in many cases). So a production--or the application of a pro-
duction-seems to be a belief-forming process. Since primary epistemology
is concerned with the reliability of belief-forming processes, it apparently
should be concerned with the reliability of individual productions or groups
of productions.
However; a moment's reflection should indicate that productions are not
(in general) belief-forming 'processes' in my special use of this term. They
are not, in particular, basic cognitive processes. In general, individual pro-
ductions are learned, highly task-specific procedures. They are not part of
the fixed, native architecture of the cognitive system. This is sufficient
grounds for their not being basic processes. They are (in general) what I
have called 'methods' rather than processes. (There are possible exceptions,
Some productions may be original, innate features of the system. These
could count as basic processes. But since ACT* has nothing to say about
such productions, I will ignore them in my discussion.)
The reliability of methods may be relevant to the secondary justifiedness
of beliefs formed by those methods. But method reliability is not relevant to
the primry justifiedness of these beliefs (see section 5.3). In this work, at
least in Part 2, my concern with justifiedness is restricted to primary justi-
fiedness. Hence, since productions are methods rather than processes, we
need not be concerned with the reliability properties of productions.
Are there any features of a production system architecture, then, which
do interest primary epistemology? Yes, two kinds of features. First, a pro-
duction system architecture includes basic processes for using productions
already in the system's repertoire. Second, a production system architecture
has processes for adding productions to the repertoire and selectively in-
creasing or decreasing the strengths of productions (strengths that in turn
affect when the productions are applied). Clearly, all of these processes are
of interest to primary epistemology. In this section I shall reflect on (some
of) the processes in ACT" for using productions.
Processes for using methods are included under the category of 'first-
order' processes. Processes for acquiring new methods, and for deleting or
modifying them in various ways, would fall under the category of 'second-
order' processes (see sections 5.3 and 5.7). Thus far I have only presented
first-order processes of ACT". So in this section my evaluative comments on
production use are confined to such first-order processes. An exposition of
ACT"'s second-order processes will be given in the next section; an evalua-
tion of them will appear in section 17.4.
Let us turn now to the reliability of the processes for using a repertoire of
productions. This includes the processes for selecting which productions to
apply on given occasions as well as the process of application itself. Of
course, the reliability of selection and application processes cannot be com-
pletely isolated from the reliability of the productions on which they oper-
ate. In the interest of simplicity, though, let us assume that all productions
in a person's repertoire are perfectly reliable. How, then, would selection
and application processes affect reliability?
One important point to note here is the partial matching feature of
ACT". Even if all the productions are perfectly reliable, error can creep in
through partial matching. A new belief can be formed in working memory
by application of a production that does not really fulfill a preexisting con-
dition. This amounts roughly to taking a sound inference procedure and
misapplying it.
A partial matching process will typically reduce overall reliability. Does
this mean that any cognitive architecture that realizes ACT0 has a feature
that precludes its approval by a right J-rule system? Not necessarily. This
will depend on two factors: the degree of reliability required for justified-
ness, and the degree of match that constitutes the minimal threshold for
production application. In section 17.1 I stressed the comparative principle
for production selection: ceteris paribus, productions with a higher degree
of match are preferred to productions with a lower degree of match, But
ACT" also includes a cutoff feature: below a certain level of match, a pro-
duction will not be selected at all. Obviously, this cutoff level is important
for reliability.
However, even if the cutoff level of a given architecture allows too much
error, as judged by a rightness criterion with a fixed truth-ratio require-
ment, this does not mean that no belief formed under this architecture is
justified. When productions are selected in virtue of a sufficiently high de-
gree of match-high enough to satisfy the specific rightness criterion-then
beliefs so formed may be justified. In those cases the minimal cutoff point is
not the causally operative degree of match. (Recall my discussion of this
sort of case in Chapter 3.)
Although partial matching is a potential liability from the standpoint of

reliability, it is probably an asset from the standpoints of speed and power.
Apropos of speed, the cognizer does not have to search or wait for all fea-
tures of a pattern to match before applying a production. He can recognize
stimuli when they have been only partly observed. This is presumably a sig-
nificant timesaver. (ACTO'spattern matcher responds faster to more com-
plete matches, when more complete matches are available. But this does
not conflict with the point I am making.)
Apropos of power, the principle of partial matching allows the cognizer
to deal successfully with the commonly found stimuli that only partly
match stored conditions, whether through occlusion, change, deformation,
or the like. Examples of occlusion were mentioned in Chapter 11.Examples
of change and deformation include two that were mentioned in the previ-
ous section: a face that has changed since previous encounters, and de-
formed lexical patterns such as the one in Figure 17.3. Without partial
matching, the cognizer could not correctly answer the questions, 'Who is
that? or 'What does that say? questions he wants to answer in these exam-
ples. Thus, partial matching can enhance power,
Another feature of ACT0 architecture that is highly congenial to ques-
tion-answering power is goal-directed processing. Complicated questions
require complicated operations to obtain (correct) answers. Frequently,
many different facts have to be retrieved (or obtained), and complex algo-
rithms or heuristics need to be applied. All of this takes time and mental or-
ganization. The mind needs to stick to the problem at hand and keep from
getting distracted. It also needs to perform operations in an appropriate se-
quence and to keep track of significant relationships. The goal-directed
structure of the ACT0 architecture facilitates satisfaction of these requi-
sites. The goal structure solves the focus-of-attention problem by allowing
the current goal to be a prime source of activation. And the provision for
hierarchical goals facilitates the necessary seriality in the control of cogni-
tion.
In the arithmetic example given in section 17.1 we saw how cognition
may be controlled by hierarchically ordered goals. This explains how the or-
ganism can engage in planning, a topic that has engrossed many cognitive
scientists. Like the dominant view of planning in cognitive science: this
depicts the topdown aspect of planning: the process starts with high-level
goals and refines them stepwise into achievable actions. But Anderson re-
marks that his ACT' system can also accommodate bottom-up, data-driven
phenomena.
Data-drivenness is accommodated by the fact that perception can serve
as a source node, that is, a source of activation which can redirect attention.
Furthermore, ACT0 is designed to reflect the fact that even planning can
involve bottom-up components. This has been stressed by B. Hayes-Roth
and F. Hayes-Roth in their conception of opportunistic phnning7 In study-
ing how subjects plan a series of errands around town, these researchers
found evidence of the subjects' proceeding unsystematically and multi-
directionally in their planning space, seeking to develop the most promising
aspects of their plan as the occasion permits. Subjects thereby mix low-level
and high-level decision making, instead of proceeding in a strictly hierar-
chical fashion.'Anderson tries to capture such possibilities in ACT*. If the
resulting procedures are indeed psychologically correct, it is an open ques-
tion how they affect the general power of the ACT0 architecture. Clearly,
however, the goal-linked properties of this architecture are importantly
connected to the epistemically significant dimension of question-answering,
or problem-solving, power.
17.3. Procedural Learning

As I remarked earlier, what concerns primary epistemology is not only pro-
cesses that use productions but also those that govern production acquisi-
tion and adjustment of production strength. In this section I give an
exposition of the processes governing acquisition and strength adjustment.
Anderson discusses them under the heading procedural learning. In my
framework they fall into the category of second-order processes, because
they control the acquisition and modification of methods.
Notice that the label 'second-order' is apt to cause confusion. One might
suppose that second-order processes are only relevant to secondary episte-
mology. But this is incorrect. Since a second-order process is a kind of basic
process, its status is exclusively of interest to primary epistemology-in-
cluding the question of primary justifiedness. It is essential to keep the con-
trasts between (a) first-order and second-order processes and (b) primary
and secondary epistemology (or justifiedness) sharply distinct. Secondary
epistemology, including the theory of secondary justifiedness, is concerned
with properties of methods, not with properties of second-order processes.
Let us now turn to the mechanisms of procedural learning built into
ACT*. Anderson first draws a sharp distinction between two kinds of
'knowledge': declarative knowledge and procedural knowledge. This di-
chotomy is reflected in two of the memories depicted in Figure 17.2-de-
clarative memory and production memory-which store the two kinds of
knowledge. The content of procedural knowledge is productions: knowl-
edge about how to do things. The phrase 'how to do things' first brings to
mind motor skills, because these are the most concrete examples of proce-
dural knowledge. But motor skills like riding a bike or typing are not the
prime focus of the ACT0 theory, although they are included within proce-
dural knowledge. In Anderson's theory the focus is on cognitive skills such
as decision making, mathematical problem solving, and language genera-
tion. This focus also suits my main concerns.
In addition to the declarative-procedural dichotomy, Anderson draws a
370 Production S y s t m and Second-Order Processes
clear distinction between the nature of fact-acquiring processes and pro-

duction-acquiring processes. It is not possible, he says, simply to add a pro-
duction in the way that one simply encodes a fact or cognitive unit.
Procedural learning occurs only in executing a skill; one learns by doing.
This is one reason why procedural learning is much more gradual than de-
clarative learning.
Notice that a skill can be embodied in two ways---either declaratively or
procedurally. For example, a declarative encoding of a recipe can be used
to guide cooking behavior with the help of a general recipe-following set of
productions. In this case the declarative information expressed by the rec-
ipe is used as a set of instructions. When, in this fashion, existing produc-
tions use declarative knowledge to guide behavior, there is an
"interpretive" embodiment of a skill. Once this knowledge has been ap-
plied interpretively a number of times, a set of new productions can be
"compiled" that apply the knowledge directly. At this stage, there arises a
"procedural" embodiment of the skill.
One might wonder why new productions are compiled at all, if skills can
be embodied interpretively. Anderson replies that interpretive application
of declarative knowledge has serious costs in time and working-memory
space. The process is slow because interpretation requires retrieving declar-
ative information from LTM and because an interpreter's individual pro-
duction steps are small in order to achieve generality. The interpretive
productions also require that the declarative information be represented in
working memory, and this can place a heavy burden on working-memory
capacity (which is much more restricted than declarative LTM or produc-
tion memory). Subjects' errors and slowness on various tasks can be attrib-
uted to working-memory errors. Students keep repeating the same
information over and over again as they lose critical intermediate results
and have to recompute them. Building up new, task-specific productions
greatly increases efficiency in terms of both time and working-memory de-
mands.
Anderson divides the process of creating new productions into two sub-
processes: composition and proceduralization. Composition takes a se-
quence of productions and collapses them into a single production that has
the effect of the sequence. This speeds up the skill considerably by creating
new operators that embody the sequences of steps used in a particular
problem domain. Proceduralization builds versions of the productions that
no longer require the domain-specific declarative information to be re-
trieved into working memory. Rather, the essential products of these re-
trieval-operations are built into the new productions themselves.
One example Anderson gives concerns dialing a telephone number? At
first, a person might use the following two productions to begin dialing a
telephone number. (In what follows, the abbreviation 'LV' stands for 'local
variable'.)
P1 IF the goal is to dial LVtelephone-number

and LVdigitl is the first digit of LVtelephone-number
THEN dial LVdigitl.
P2 IF the goal is to dial LVtelephone-number
and LVdigitl has just been dialed
and LVdigit2 is after LVdigitl in LVtelephone-number
THEN dial LVdigit2.
The process of composition, operating on these two productions, could cre-
ate a "macroproduction":
Pl&P2 IF the goal is to dial LVtelephone-number
and LVdigitl is the first digit of LVtelephone-number
and LVdigit2 is after LVdigitl
THEN dial LVdigitl and then LVdigit2.
A composed production like Pl&P2 still requires that the information (in
this case, the phone number) be retrieved from declarative memory, held in
working memory, and matched to the second and third clauses in Pl&P2.
Proceduralization eliminates clauses (in the condition of a production) that
require information to be retrieved from declarative memory and held in
working memory. In P1&P2 the second and third condition clauses would
be eliminated. The local variables that would have been bound in matching
these clauses are replaced by the values they are bound to in the special
case. For example, if this and related productions are repeatedly applied in
dialing Mary's telephone number, which is 432-2815, procedularization
would eventually build a macroproduction that dials the full number:
P' IF the goal is to dial Mary's number
THEN dial 4-3-2-2-8-1-5.
In using this production, one no longer has to retrieve Mary's number from
declarative memory, match it to the conditions of a production, and
.
thereby use up working memory space with it. The number is part of the
production itself. This knowledge-specific procedure makes it easier for the
system to simultaneously perform a second concurrent task that does make
working-memory demands.
There are two other processes of building new productions that Anderson
discusses, although he considers them under the heading "tuning" (thereby
grouping them, somewhat oddly, with the processes that affect production
strength). These processes are generalization and discrimination. General-
ization is a process in which production rules become broader in their range
of applicability, and discrimination is one in which the rules becomes nar-
rower. Strictly, this means that new productions, either broader or nar-
rower, are constructed.
Generalization may be illustrated by an example from the game of
bridge. While bridge players often learn rules of play from expert instruc-
tion, the following example assumes that the person has compiled some
special-case production rules from experience. Two such rules might be:
P1 IF I am playing no trump
and my dummy has a long suit
THEN try to establish that suit
and then run that suit.
P2 IF I am playing spades
and my dummy has a long suit
ACT* will create a generalization of these productions:
P3 IF my dummy has a long suit
It is certainly useful to generalize overly specific procedures. But it is
even more necessary to restrict the range of overly general procedures. The
discrimination process tries to restrict the range of application of a produc-
tion-actually, construct a new production-to just the appropriate cir-
cumstances. This requires that the cognizer have examples of (ostensibly)
correct and incorrect applications of the production. The discrimination
process remembers and compares the values of the variables in the correct
and incorrect applications. It chooses a variable for discrimination among
those that have different values in the two applications. Having selected a
variable, it looks for some attribute that the variable has in only one of the
situations. A test is added to the condition of the production for the pres-
ence of this attribute.'
The feature selected for discrimination is determined by comparing the
variable bindings in the successful and unsuccessful production applica-
tions. A variable is selected on which the applications differ, and features
are selected to restrict the bindings. Anderson says that the probability of
choosing the right feature to discriminate upon depends on the similarity of
the successful and unsuccessful situations, The more similar they are, the
fewer the distracting possibilities, and the easier it is to identify the critical
feature. This is the idea of near misses, discussed for example by P. H. Win-
ston.''
A prerequisite for discrimination is that the system have feedback indi-
cating that a particular production has misapplied. In principle, Anderson
says, aproduction application could be characterized as being in one of
three states-known to be incorrect, known to be correct, or correctness
unknown. However, the mechanisms implemented in ACT* do not distin-
guish between the second and third states. If a production applies and there
is no feedback about its success, it is treated as if it is a successful applica-

tion,
There are two basic ways in which ACT* can determine that a produc-
tion application is in error. One is through external feedback, and the other
is through internal computation. In the external feedback situation the
learner may be directly told that his behavior is in error or he may infer this
by comparing his behavior to an external referent (such as a model or a
textbook answer). In the internal computation case the learner must ascer-
tain that an output of the production contradicts prior or subsequently
formed beliefs, or that a goal has failed. In either case the feedback is only
indirectly about the production. It is directly about the action of the pro-
duction that entered into working memory.
It should be emphasized that whenever new productions are created,
whether by generalization, discrimination, or composition, they do not re-
place previous productions. They just coexist with the latter, which are
never replaced. However, although productions are never deleted, even
when they are seen to be incorrect, they can remain or become very weak.
Comparatively weak productions may seldom or never get applied (since
they lose out to stronger rivals). Thus, even though incorrect productions
remain in the repertoire, their existence need not be deleterious.
This brings us to the question of how productions gain or lose strength.
Anderson offers an algebra for strength determination. When first created, a
production has a strength of 1.Each time it applies, its strength increases by
1. However, when a production applies and receives negative feedback, its
strength is reduced by 25 percent. Because a multiplicative adjustment
usually produces a greater change in strength than an additive adjustment,
this 'punishment' has much more impact than a reinforcement. Anderson
adds that although his exact strengthening and weakening values are some-
what arbitrary, the general relationships are important.
Anderson has several further principles that affect strengths of produc-
tions, and I will not mention them all. One additional principle, though,
should be explained. The ACT theory posits a power-law decay principle
for the strength of units in declarative memory, and a similar power-law
decay principle is included for production strength. Thus, productions lose
strength over time through disuse.
1 7.4. Epistemic Assessment of Second-Order Processes

It is time for an epistemic evaluation of ACTe's processes of production ac-
quisition and tuning. First of all, let us recall from Chapter 5 the proposals
for evaluating second-order processes. Focusing on primary justifiedness, I
argued that the theory of primary justifiedness depends crucially on two
kinds of criteria of rightness: a criterion of rightness for first-order processes
and a criterion of rightness for second-order process. Although I gave scant

attention to the rightness of second-order processes, this was discussed in
section 5.7, where I suggested that a rightness criterion for second-order
processes might naturally invoke some notion of second-order reliability, or
metareliability.
Backing up a bit, we may characterize a second-order process as a pro-
cess that controls the acquisition of new methods, or otherwise governs the
repertoire of methods available for use in the cognitive system. Next, a sec-
ond-order process may be called 'metareliable' in case, roughly, the
methods it tends to acquire and preserve have sufficiently good reliability
properties; or if it tends to imprme the reliability properties of the method
repertoire over time.
In section 5.7 two kinds of proposals were made. The first proposal was
that a second-order process is metareliable if and only if the proportion of
reliable methods it tends to acquire is sufficiently high. In other words, the
ratio of reliable methods to all methods it would acquire meets or exceeds
some specified threshold. For example, if the threshold is set at .80, then a
second-order process is deemed metareliable if and only if 80 percent of the
methods it would acquire are reliable. (This presupposes a threshold for fix-
ing a categorical, binary concept of reliability.)
A second kind of proposal concerning metareliability focused on
diachronic improvements in method repertoire. This proposal presupposed
that second-order processes effect modifications, replacements, and dele-
tions of methods. One conception, then, would consider a second-order pro-
cess metareliable in case its operation always yields improvements in
reliability. Modifications of methods must always increase their reliability;
replacements must always substitute more reliable for less reliable methods;
and simple deletions must delete only (comparatively) unreliable methods.
A weaker and more plausible variant would not require all modifications,
replacements, and deletions to effect improvements in reliability. It may be
deemed sufficient that a high enough proportion of them should yield in-
creases in reliability.
To obtain a unified criterion of evaluation, one would have to decide how
to balance, or integrate, appraisals of method acquisition and appraisals of
method refinement (including modification, replacement, and deletion). It
is not obvious how this should be done. At first blush, method acquisition
seems quite important. If a second-order process permits many unreliable
processes to be created, that seems to be a critical defect. But further re-
flection raises doubts on this score. If unreliable methods are promptly mod-
ified, r~placed,or simply deleted, perhaps it is not so bad that unreliable
methods are easily created.
The matter is further complicated by the possibility-introduced here by
Anderson's production system theory-that methods can vary in strength. If
Production Systerns and Second-Order Processes 375
methods vary in strength, as productions allegedly do, it may be harmless

for unreliable methods to remain in a system's repertoire, as long as (a) they
are very weak, and (b) weak methods do not get applied when they conflict
with stronger rivals. The strength variable was not contemplated when my
proposals in Chapter 5 were advanced. Once introduced, it complicates the
conception of metareliability still further.
The last complication is, of course, directly relevant to the learning and
tuning processes within ACT0. ACT0 has no process for deleting produc-
tions. This might look like a serious flaw. But the tuning processes of ACT0
do yield adjustments in strength. In particular, they (ostensibly) favor
strength reductions for unreliable productions. Since strength reduction
lowers the frequency or probability of production application, perhaps
these learning and tuning processes are not so inadequate from the stand-
point of metareliability (or justifiedness).
I am not prepared to propose any unified criterion of J-rightness for sec-
ond-order processes. But as my several proposals indicate, I assume that a
rightness criterion for second-order processes will involve some sort of me-
tareliability standard. Although I am not advancing any single rightness cri-
terion for second-order processes, we can still examine the learning and
tuning processes of ACT0, with an eye to their metareliability, as measured
by any of the standards we have canvassed. Before proceeding to this exami-
nation, though, two comments are in order.
First, we should not assume that second-order processes can be evaluated
in isolation. Like their first-order brethren, second-order processes are in-
terdependent. The effects of any given process partly depend on comple-
mentary processes that are at work. So these processes should be evaluated
collectively: the proper unit of evaluation is really a system, or set, of sec-
ond-order processes. Nonetheless, we can inspect second-order processes
individually to see what contribution they might make to a larger system of
second-order processes.
Second, our interest should not be restricted to metareliability. It should
encompass metapower and metaspeed as well. Justifiedness mainly concerns
reliability. But justifiedness is not the only notion of epistemic importance.
Intelligence, as embodied in power and speed, is also important. Hence, the
capacity of learning and tuning processes to generate and strengthen faster
and more powerful methods is an epistemically positive feature.
With these points in mind, let us now examine the several learning and
tuning processes incorporated in ACT0. First, consider the building of new
productions by composition and compilation. If this has the psychological
properties claimed by Anderson, then it enhances at least the speed and
power of cognitive skills. Speed is increased because the relevant informa-
tion for solving the problem no longer needs to be retrieved from declara-
tive memory nor entered into working memory. Power is increased because
fewer demands are made on working-memory space. This allows that space
to be used for other, concurrent cognitiye tasks, thereby increasing overall
power.
The impact on reliability is somewhat mixed, although in the long run it
appears to be positive. Anderson notes that the transition phase from de-
clarative to procedural embodiment of a skill often breeds the most errors.
But these are eventually corrected. Moreover, he emphasizes that heavy use
of working-memory space produces errors, so the bypassing of working
memory effected by proceduralization should increase reliability.
The fact that new productions can only be compiled when their com-
ponents have already been applied successfully is another feature of the
ACTo processes conducive to metareliability. It means that merely arbi-
trary will not be constructed. Nor will they be constructed
just from verbal instructions, which might possibly be misunderstood or
mistaken.
I turn now to the production-building processes of generalization and dis-
crimination. Both of these processes can sometimes produce incorrect,
meaning unreliable, productions. Overgeneralization is a well documented
phenomenon in the language acquisition literature. It occurs with both
syntactic rules and natural-language concepts. Discrimination can also fail
by introducing inappropriate restrictive conditions, as discussed in section
17.3.
Strength-adjustment processes are ones by which such errors can be miti-
gated. If overgeneralizations or misdiscriminations are detected, that is, if
negative feedback follows their application, then they are reduced in
strength by 25 percent. Also, in many such cases, new productions are in-
troduced to handle (more correctly, it is hoped) the same tasks. With these
rival productions in place, the old mistaken productions tend to lose out and
not get applied. Such disuse leads to gradual decay.
Is it clear, though, that strength-adjustment processes can adequately
compensate for the errors of production-creation processes? This depends
partly on just how meta-unreliable the production-creation processes are,
and partly on the effectiveness of feedback processes.
Mfiththese questions in mind, consider the process of discrimination. Ac-
cording to Anderson's description, the discrimination algorithm randomly
chooses a variable for discrimination from among those that have different
values in successful and unsuccessf+ applications. It then looks for some at-
tribute that the variable has in only one of the situations, a successful situa-
tion, and adds this condition to the production. But it appears that correct
identification of such an attribute is largely a chance affair. True, a discri-
minated production yields the same belief ('action') as the original produc-
tion in the restricted situation; so the discrimination process does not
impose an unreliable production over a reliable one. But it is problematic
whether improvements will be secured very often.
Production Systems and Second-Order Processes 3 77
The matter of feedback is also delicate. Feedback itself is not infallible.

Indeed, what Anderson calls "internal" feedback is especially susceptible to
error. So it cannot be assumed that productions whose applications yield
positive, or non-negative, feedback are ipso facto reliable.
Thus, we cannot in general certify that because a production has survived
the operation of strength-adjustment processes and still remains fairly
strong, it is therefore reliable. However, there may be a greater likelihood
that strong productions are more reliable than weak ones. A production can
only have attained substantial strength by being applied repeatedly without
incurring (much) negative feedback. High strength indicates that a produc-
tion has weathered the test of repeated applications and come through
more or less unscathed. A history of 'failure to fail', of trials without de-
tected error, may be a good indicator of genuine reliability. (The idea here
is analogous to Karl Popper's account of corroboration. On Popper's ac-
count a theory is corroborated to the extent that it survives tests without
falsification.)
This also suggests another idea similar to one mentioned in section 16.3.
Setting aside the rightness criterion for second-order processes, what sort of
J-rules should be contemplated for second-order processes? What I wish to
stress is that a J-rule can permit certain sequences of processes but not other
sequences. For example, in connection with ACTQ,a J-rule system is not
forced to choose between sanctioning single applications of the strengthen-
ing process and refusing to sanction the strengthening process at all. It
could decide to license a sequence of, say, twenty strengthenings of a given
production, but not license shorter sequences of strengthenings.
Suppose it does this. Then if a belief is formed by application of some
production P, and P is applied as a result of having a certain strength T, and
its strength T resulted from (at least) twenty applications on P of the
strengtheningoperation, then the belief would qualify for justifiedness as far
as second-order processes are concerned. By contrast, if the belief is formed
by application of P*, and PQ's strength only resulted from, say, seven
strengthenings, then the belief would not qualify for justifiedness.
Whatever the ultimate upshot of such deliberations, we have here a clear
example of how the learning and tuning processes of ACTQcould be evalu-
ated from the standpoint of primary epistemology or primary epistemics. I
have not, of course, presented a final formulation of a rightness criterion for
second-order processes. This leaves my treatment extremely tentative.
Nonetheless, it should be evident how such a criterion could be applied. I
should also stress, once again, that perhaps neither ACT* nor any other pro-
duction system framework provides a descriptively correct cognitive psy-
chology. And if it does not, then the evaluation of ACTQis not a serious
epistemological goal. Still, ACTQ provides a very instructive example of
how the enterprise of primary epistemics should be conducted, especially
with respect to second-order processes.
Conclusion: Primary Epistemics and
Cognitive Science
I HAVE NOW completed the task of Part 11: to provide some detailed speci-
mens of primary epistemology or primary epistemics. Needless to say, these
chapters do not purport to be a final embodiment of primary epistemics.
They are, as billed in section 9.1, only first steps in the subject. The enter-
prise should of course continue with new developments in a number of di-
rections.
It is a bit misleading to call the present chapters 'first' steps in primary
epistemics. As indicated in the Introduction, there are earlier conceptions
and traditions with which primary epistemics has close affiliations, in-
cluding the psychologistic epistemologies of the seventeenth through the
nineteenth centuries and W. V. Quine's conception of "naturalistic episte-
mology." However, only the accomplishments and prospects of contempo-
rary cognitive science make feasible a serious execution of primary
epistemics. Although the general conception of primary epistemics could
have been advanced in previous centuries (and was indeed adumbrated by
Locke and Hume), comparable opportunities for scientifically grounded
analysis of mental operations were not then available.
What further developments should be expected in primary epistemics?
First and most obvious, developments are to be expected in substantive
cognitive sciencwhanges in theoretical posture as well as refinements in
empirical detail. The 1980s are witnessing new approaches to memory, and
these can be expected to burgeon. Linkages to neural constructs are bound
to be strengthened. Whether such a schema as the production system
framework continues to be fruitful only time will tell. Whether the mind is
importantly modular is equally a matter for future cognitive science to de-
termine. New developments in cognitive theory mean changes in the po-
sited 'iepertoire of fundamental operations, or changes in the precise
properties of these operations. Such changes could require significant revi-
sions in empirical material of the kind discussed in Part 11.
Second, new developments should be encouraged in primary epistemics
Conclusion 379
even with a completely static cognitive science. After all, epistemics un-
dertakes to draw evaluative conclusions about cognitive processes. Even if
cognitive theory is held fixed, considerable work must be done in assessing
the reliability, power, and speed of the components of the cognitive reper-
toire. There is the task of determining what right J-rule systems, if any, can
be devised for this repertoire. Can some (or all) of the processes belonging
to this repertoire be deployed in a way that satisfies a specified truth ratio
criterion? Are there more than one set of processes that can do this? If so,
what are they? Even holding a posited cognitive repertoire constant, an-
swers to these questions are not so easy to give; and revisions in putative an-
swers are naturally to be expected.
Third, developments are needed in the formulation of evaluative criteria.
My own proposals, which have been deliberately flexible, would profit from
some tightening. As an example, let me review my treatment of justified-
ness. I began in Chapter 5 by proposing a family of alternative rightness
criteria for belief-forming processes. My calibrational approach to degrees
of credence, in Chapter 5, was also rather exploratory. Further proposals for
a calibrational approach to degree of credence were advanced in Chapter
16, but these too were quite tentative. Finally, Chapter 17 surveyed alter-
native criteria for second-order processes--all involving metareliability-
but none was definitely endorsed. The criteria for J-rule rightness, then,
clearly warrant further attention.
Looking beyond criteria of justifiedness, more detailed standards of eval-
uation may be proposed for other dimensions of evaluation, such as intelli-
gence or rationality. I have made a few suggestions on these matters, but
more remains to be done. Again, while I have expressed doubts about a
unifid criterion of epistemic appraisal, perhaps other theorists could in-
vent and defend such a measure, integrating such standards as reliability,
power, and speed. This would certainly be within the spirit of the general
enterprise.
Proposals for revising criteria of evaluation can come from different .
sources. One is general reflection on the ordinary use of the relevant evalua-
tive terms. Another might be progress in empirical psychology, which
highlights factors that could be missed in purely abstract reflection. An ex-
ample was given in Chapter 17, in the discussion of rightness criteria for
second-order processes. We had earlier supposed that rightness criteria
would concern only the acquisition and retention of methods. But consider-
ation of ACT* drew our attention to the strength variable. It became plau-
sible to suppose that a satisfactory J-rule criterion for second-order
processes might invoke some factor like strength, or probability of employ-
ment.
Another avenue of development lies in possible applications of primary
epistemics. I have presented the evaluative enterprise as a 'theoretical'
380 Conclusion
rather than a 'regulative' one. But there are some prospects for deliberate
control of cognitive operations, should this prove advisable. Habits in de-
ployment of the cognitive repertoire may be amenable to inculcation and
training. There may be techniques for promoting the use of certain se-
quences or patterns of operations over others. If primary epistemics distin-
guishes superior from inferior processes, it is natural to try to promote the
better over the worse. A challenge is then extended to educational theory to
devise techniques for achieving this end.
There is, then, a substantial agenda for the future of primary epistemics,
and for offshoots from it. Exciting prospects lie ahead for a collaboration
between philosophy and a variety of cognitive disciplines.
In recent years philosophers have laid the groundwork for a working rela-
tionship between cognitive science and the philosophy of mind. The argu-
ment here is that cognitive science is equally important (though in different
ways) to epistemology. However, my conception of epistemology is even
broader than this. At the social level it would intersect with yet other disci-
plines outside of philosophy proper. A full argument for this thesis, how-
ever, must await a future volume.
Notes
Illustration Credits
Author Index
Subject Index
Notes
Introduction
1. Ludwig Wittgenstein, Philosophical Investigations, trans. C . E. M. An-
scombe (New York: Macmillan, 1953), para. 309.
2. W. V. Quine, "Epistemology Naturalized," in Ontological Relativity and
Other Essays (New York: Columbia University Press, 1969), p. 82.
3. Donald Campbell, "Evolutionary Epistemology," in Paul Schilpp, ed., The
Philosophy of Karl Popper, vol. 1 (La Salle, Ill.: Open Court, 1974), p. 413.
4. Two other recent works that treat epistemology in some sort of psycholo-
gistic vein are Gilbert Harman's Thought (Princeton: Princeton University Press,
1973) and Fred Dretske's Knowledge and the Flow of Information (Cambridge,
Mass.: MIT Press, 1981). But the approaches of these authors are quite different
from mine. For example, instead of drawing on psychology to reach epistemolo-
gical conclusions, Harman's main strategy is to use epistemological judgments to
reach psychological conclusions. Specifically, he seeks to decide what inferences
people make by seeing what inferences must be postulated to account for our
judgments of whether or not they 'know' (see Thought, pp. 20-23.) This seems to
me to get the relation between psychology and epistemology backward. Dretske
refines a popular term of cognitive theory-'informationJ-and puts it to episte-
mological use. But his theory assigns no systematic role to empirical psychology;
nor does he conceive of epistemology as an evaluative discipline. '
5. Although it is not clear that Quine intends to neglect the evaluative di-
mension, his actual characterizationsof naturalistic epistemology do not expressly
introduce this dimension. See, for example, "The Nature of Natural Knowledge,"
in Samuel Guttenplan, ed., Mind and Language (Oxford: Clarendon Press, 1975).
6. Otto Neurath, "Protokollsatze," Erkenntnis, 3 (1932):206.
7. Francis Bacon, N m m Organum (1620), bk. 1, sect. 2.
8. John Locke, An Essay concerning Human Understanding (1690), intro-
duction, sect. 4.
9. David Hume, A Treatise of Human Nature (1739), introduction.
10. On Carnap's sharp separation, see The Logical Syntax of language, trans.
Amethe Smeaton (London: Routledge and Kegan Paul, 1937), sect. 72, esp. pp.
278-279. Reichenbach's distinction appears in Experience and Prediction (Chi-
cago: University of Chicago Press, 1938), pp. 6-8.
384 Notes to Pages 7-28
11. Quine, "Epistemology Naturalized," p. 83.

12. Thomas Kuhn, The Structure of Scientific Revolutions, (Chicago: Univer-
sity of Chicago Press, 1962).
13. Philosophical Investigations, for example, para. 486.
14, Richard Rorty, Philosophy and the Mirror of Nature (Princeton: Princeton
University Press, 1979), pp. 170, 174.
1. The Elements of Epistemology

1. Richard Rorty, Philosophy and the Mirror of Nature (Princeton: Princeton
University Press, 1979), p. 308.
2. See especially Karl Popper, The Logic of Scientific Discovery (London:
Hutchinson, 1959) and Conjectures and Refutations (New York: Harper & Row,
1963).
3. See my "What Is Justified Belief?" in George Pappas, ed., Justificationand
Knowledge (Dordrecht: D. Reidel, 1979); reprinted in Hilary Kornblith, ed., Na-
turalizing Epistemology (Cambridge, Mass.: MIT Press, 1985).
4. I draw here on a distinction made by Holly Smith, in her article "Making
Moral Decisions," forthcoming in Nous. Smith talks of two roles, or functions,
that a principle, especially a moral principle, might have: a "theoretical" role and
a "first-person practical" role.
5. One example is Descartes, in his Rules for the Direction of the Mind. In an
earlier work I considered a conception of justification that was linked to action-
guiding principle-what I there called 'doxastic decision principles'. (See "The
Internalist Conception of Justification," in Peter French, Theodore Uehling, Jr.,
and Howard Wettstein, eds., Midwest Studies in Philosophy, vol. 5, Studies in
Epistemology, Minneapolis: University of Minnesota Press, 1980.) In this book,
however, my account of justification does not invoke such decision principles.
6. Notice that appraisal, or assignment of normative status, needn't involve a
judgment of praiseworthiness or blameworthiness. Thus, one might judge a cer-
tain belief to be justified without implying that the cognizer merits praise for it;
and one might judge a belief to be unjustified without implying blame or con-
demnation. Praise and blame are in order where voluntary performances have
been executed. But many cognitive processes culminating in belief are not volun-
tary.
2. Skepticism
1. A classification of types of skepticism is given by George Pappas in "Some
Forms of Epistemological Scepticism," in George Pappas and Marshall Swain,
eds., Essays on Knowledge and Justification (Ithaca: Cornell University Press,
1978). For a historical overview of skepticism, see Richard Popkin, "Skepticism,"
in Paul Edwards, ed., The Encyclopedia of Philosophy, vol. 7 (New York: Macmil-
lan, 1967). A critical survey of skeptical arguments is given by Nicholas Rescher
in Sceptic- (Totowa, N.J.: Rowman and Littlefield, 1980).
2. Sextus Empiricus, Outlines of Pywhonism, trans. R. G. Bury (London:
Heinemann, 1933).
Notes to Pages 30-40 385
3. See Peter Klein, Certainty: A Refutation of Scepticism (Minneapolis: Uni-

versity of Minnesota Press, 1981).
4. See Popkin, "Skepticism," p. 454.
5. Immanuel Kant, Critique of Pure Reason, 2nd ed. (1787).
6. Perhaps Plato's own worry about change, however, was a little different.
7. David Hume, Enquiry into the Human Understanding (1748).
8. Bertrand Russell, The Analysis of Mind (New York: Macmillan, 1921).
9. Outlines of Pyrrhonism, vol. 1, pp. 163-164.
10. Perhaps Nelson Goodman's new riddle of induction can be viewed as a
skeptical argument in this general category. Cf. Fact, Fiction, and Forecast, 2nd
ed. (Indianapolis: Bobbs-Merrill, 1965). For every inductive conclusion, there are
alternative conclusions with seemingly parallel support. Of course, Goodman
presents his riddle not so much as a problem about justification, but rather as a
problem of how to describe our inductive practice. But it still can be viewed as a
skeptical puzzle. As such, it would not quite fit into the 'alternative explanation'
category I expound in the text, but neither does it fit so naturally into the possibil-
ity-of-error category.
11. George Berkeley, A Treatise concerning the Principles of Human Knowl-
edge (1710) and Three Dialogues between Hylas and Philonous (1713).
12. John Stuart Mill, An Examination of Sir William Hamilton's Philosophy
(1865).
13. See Hilary Putnam, Meaning and the Moral Sciences (London: Routledge
and Kegan Paul, 1978) and Reason, Truth and History (Cambridge: Cambridge
University Press, 1981).
14. I first developed a causal approach in "A Causal Theory of Knowing," The
Journal of Philosophy, 64 (1967):355372; reprinted in George Pappas and Mar-
shall Swain, eds., Essays on Knowledge and Justification(Ithaca: Cornell Univer-
sity Press, 1978). Some problems facing this particular theory are discussed in
other selections reprinted in the same collection. There will be more on the causal
approach in Chapter 3.
15. See Gilbert Harman, Thought (Princeton: Princeton University Press,
1973).
16. Sydney Shoemaker, Self-Knowledge and Self-Identity (Ithaca: Cornell
17. Barry Stroud, "Transcendental Arguments," The Journal of Philosophy, 65
(1968):241-256.
18. The most famous sources for these attacks are W. V. Quine, "Two Dogmas
of Empiricism," in From a Logical Point of View (Cambridge, Mass.: Harvard
University Press, 1953) and Word and Object (Cambridge, Mass.: Technology
Press of MIT, 1960).
19. There is also psychological evidence challenging the notion that word
meanings are mentally represented as necessary and sufficient conditions. For a
review of this evidence, see Edward Smith and Douglas Medin, Categories and
Concepts (Cambridge, Mass.: Harvard University Press, 1981).
20. I do not deny that some skeptical arguments might challenge the rational-
ity or justifiedness of even partial beliefs. But the more familiar lines of argument
are directed at knowledge, certainty, or 'acceptance'.
21. It should also be noted that certain questions about knowledge and justifi-
cation can be studied with little or no concern for skepticism. For example, how
should knowledge be analyzed in a way that meets the Gettier problems? This
question about knowledge has little connection to skeptical concerns. Similarly,
there is the question: what is the structure of epistemic justification? Is it foun-
dationalist or coherentist? This problem can also be investigated with little con-
cern for skepticism. It may be assumed that we do have justified beliefs in some
fashion or other; it just remains to be shown how that justification arises. I
have not stressed these points because they will not be central to the continuing
discussion.
3. Knowledge
1. Some people object to using 'cause' in the analysis of knowledge because
that notion is itself philosophically problematic. But we can make progress in un-
derstanding one concept by using others, even if the latter also invite analysis. In
particular, it is useful to analyze epistemological notions in terms of nonepiste-
mological ones, as in the present case. One comment on my view of causation is in
order. I do not presume that causation implies determinism. Hence, while knowl-
edge cannot be attained if beliefs are wholly uncaused, knowledge is possible in a
nondeterministic universe.
2. D. M. Armstrong, Belief, Truth and Knowledge (Cambridge: Cambridge
University Press, 1973). Also see Marshall Swain, Reasons and Knowledge (Ithaca:
Cornell University Press, 1981).
3. I will not reproduce all the details of Armstrong's account, which gets
rather intricate. (See Belief, Truth and Knowledge, pp. 192ff.) But I do not believe
that any of his qualifications ultimately save his approach.
4. Armstrong gives a somewhat similar case, attributed to Ken Waller. (Cf.
Belief, Truth and Knowledge, pp. 178-179.) But my case is a bit different, and is
not (I believe) open to the same rejoinder as he gives to Waller's. For more on the
reliable-indicator versus reliable-process approach--applied, however, to the the-
ory of iustification-see Frederick Schmitt, "Justification as Reliable Indication
or Reliable Process," Philosophical Studies, 40 (1981):409-417.
5. Frank Ramsey, The Foundations of Mathematics and Other Logical Essays
(London: Routledge and Kegan Paul, 1931).
6. These terms are borrowed from Colin McGinn, "The Concept of Knowl-
edge," in Peter French, Theodore Uehling, Jr., and Howard Wettstein, eds., Mid-
west Studies in Philosophy, vol. 9, Causation and Causal Theories (Minneapolis:
University of Minnesota Press, 1984). In earlier drafts of this chapter I used the
phrases 'generic' and 'situation-specific' reliability. But I prefer the 'global' and
'local' terminology.
7. Robert Nozick, Philosophical Explanations (Cambridge, Mass.: Harvard
Universip Press, 1981), chap. 3.
8. Fred Dretske, "Conclusive Reasons," The Australasian Journal of Philoso-
phy, 49 (1971):l-22. Also see his Knowledge and the Flow of Information (Cam-
bridge, Mass.: MIT Press, 1981). It is not clear that Dretske's theory should be
viewed as a specimen of the reliable-processapproach since it places little impor-
tance on the causal production of belief.
9. Actually, the core of the thermometer example is taken from Dretske's

own paper "Conclusive Reasons." Nonetheless, it is easily turned against his ac-
count.
10. Alvin Goldman, "Discrimination and Perceptual Knowledge," The Journal
of Philosophy, 73 (1976):771-791.
11. Edmund Gettier, "Is Justified True Belief Knowledge?' Anulysis, 23
(1963):121-123, showed that even when a true belief is justified, it may not qualify
as knowledge. In particular, his examples were ones in which a true belief in p is
justifiably inferred from a false but justified belief in q. Later writers presented
other sorts of examples in which justified true beliefs fail to be items of knowl-
edge. A review of this literature, and attempted solutions to the analysis of knowl-
edge, is found in Robert Shope, The Analysis of Knowing (Princeton: Princeton
12. Brian Skyrms, "The Explication of 'X knows that p'," The Journal of Phi-
losophy, 64 (1967):373-389.
13. An account of knowledge along these lines was formulated in an unpub-
lished paper of mine, "Inferential Knowledge," first presented at Princeton Uni-
versity in 1974.
14. See Alvin Goldman, "Innate Knowledge," in Stephen Stich, ed., Innate
Ideas (Berkeley: University of California Press, 1975) and "What Is Justified Be-
lief?" in George Pappas, ed., Justifiation and Knowledge (Dordrecht: D. Reidel,
1979).
15. McGinn, "The Concept of Knowledge," p. 536.
16. Ibid., pp. 536-538.
17. McGinn does offer general methodological reasons why a counterfactual
condition is not appropriate. These are interesting but not conclusive, and I will
not enter into a discussion of them here.
18. See McGinn, "The Concept of Knowledge," pp. 533-535.
19. If one adopts a fine-grained theory of individuating process tokens, more
than one token will cause the belief. But on any theory many types will be in-
volved, which is the source of the problem.
20. This point was first made by Frederick Schmitt in comments on Richard
Feldman's "Reliability and Justification," a paper presented at the 1983 Western
Division meeting of the American Philosophical Association.
21. Especially Feldman, in "Reliability and Justification," and also John Pol-
lock, in "Reliability and Justified Belief," C a d i a n Journal of Philosophy, 14
(1984):103-1 14.
22. However, since these algorithms are eventually called 'methods' rather
than 'processes', it is a bit odd to call an acquisition process a second-order pro-
cess. Since nothing hinges on this terminological point, though, I will let the ter-
minology stand.
23. If Humperdink applies this algorithm repeatedly and finds by independent
criteria that it yields correct answers, and if this sustains or supports his continued
use of the algorithm, then his possession and deployment of the algorithm is over-
determined. Moreover, one of the determining processes is a good one. In this
case, by contrast with the case in the text, Humperdink's application of the algo-
rithm can yield knowledge. In the text I assume that the sole cause of his algo-
rithm deployment is his blind faith in Fraud.
!24. For more on the psychology of procedure acquisition, see Chapter 17.
25. Further discussions of this topic appear in section 5.7 and Chapter 17.
26. Cf. Gettier, "Is Justified True Belief Knowledge?'
27. Keith Lehrer and Thomas Paxson, Jr., "Knowledge: Undefeated Justified
True Belief," The Journal of Philosophy, 66 (1969):225-237.
28. Fred Dretske, "Epistemic Operators," The Journal of Philosophy, 67
(1970):1007- 1023, and Nozick, Philosophical Explanations, pp. 204-21 1, 227-
229.
29. "The Concept of Knowledge," pp. 542ff.
30. See William Alston, "Level-Confusions in Epistemology," in Peter
French, Theodore Uehling, Jr., and Howard Wettstein, eds., Midwest Studies in
Philosophy, vol. 5, Studies in Epistemology (Minneapolis: University of Minnesota
Press, 1980).
31. W. V. Quine, "The Nature of Natural Knowledge," in Samuel Guttenplan,
ed., Mind and Language (Oxford: Oxford University Press, 1975), p. 68.
4. Justification:A Rule Frantework

1. I was reminded of this practical context for justification judgments by I. T.
Oakley's paper "Should Epistemology Be Naturalized?' presented at the Univer-
sity of Arizona in 1985. The practice of making "head-world" inferences of the
indicated sort is stressed in Stephen Schiffer, "Truth and the Theory of Content,"
in H. Parret and J. Bouverese, eds., Meaning and Understanding (Berlin: Walter
de Gruyter, 1981) and Hartry Field, "Mental Representation," Erkenntnis, 13
(1978):9-61, though without attention to the relevance of justifiedness in such in-
ferences.
2. I do not mean to beg the question of a possible connection between right-
ness and general acceptance. If there is such a connection, though, it would have
to be established.
3. My full theory really requires J-rule systems that include obligations as
well as permissions (see section 5.6). But I shall concentrate on permissions.
4. Conditions for justified belief, or knowledge, proposed by some writers
would violate this constraint. See, for example, Wilfrid Sellars, "Empiricism and
the Philosophy of Mind," in Science, Perception, and Reality (London: Routledge
and Kegan Paul, 1963), esp. p. 168, and Laurence BonJour, "Can Empirical
Knowledge Have a Foundation?' American Philosophical Quarterly, 15
(1978):1-15. Ernest Sosa properly diagnoses such violations in "The Raft and the
Pyramid," in Peter French, Theodore Uehling, Jr., and Howard Wettstein, eds.,
Midwest Studies in Philosophy, vol. 5, Studies in Epistemology (Minneapolis: Uni-
versity of Minnesota Press, 1980), sect. 8.
5. On the idea of level-confusions, see William Alston, "Level-Confusions in
Epistemology," in Peter French, Theodore Uehling, Jr., and Howard Wettstein,
eds., Midwest Studies in Philosophy, vol. 5, Studies in Epistemology.
6. John Rawls, A Theory of Justice (Cambridge, Mass.: Harvard University
Press, 1971).
7. John Austin, l'he Province of Jurisprudence Determined (1832).

8. H. L. A. Hart, The Concept of Law (Oxford: Oxford University Press,
1961).
9. On this point, see Jules Coleman, "Negative and Positive Positivism," The
Journal of Legal Studies, 11 (1982):139-164.
10. Ludwig Wittgenstein, Philosophical Inuestigations, trans. G. E. M. An-
scombe (New York: Macmillan, 1953).
11. Thomas Kuhn, "Second Thoughts on Paradigms," in The Essential Tension
(Chicago: University of Chicago Press, 1977).
12. Richard Rorty, Philosophy and the Mirror of Nature (Princeton: Princeton
13. Nelson Goodman, Fact, Fiction, and Forecast, 2nd ed. (Indianapolis:
Bobbs-Merrill, 1965), and Rawls, A Theory of Justice, pp. 20-21, 46-51.
14. However, to the extent that a criterion is tested by the intuitive accept-
ability of the criterion's consequences, psychology might be indirectly relevant in
the assessment of certain candidate criteria.
15. Rudolf Carnap made the most systematic attempt to develop an inductive
logic, in Logical Foundations of Probability, 2nd ed. (Chicago: University of Chi-
cago Press, 1962). But few theorists now regard that sort of approach as very
promising.
16. Cf. John Pollock, "Epistemology and Probability," Synthese, 55
(1983):231-252.
17. I know of nobody who formulates precisely this idea. But it has thematic
affinitieswith Descartes's project and with the working practice of some induc-
tive logicians, such as Carnap in Logical Foundations of Probability and "The
Aim of Inductive Logic," in Ernest Nagel, Patrick Suppes, and Alfred Tarski, eds.,
Logic, Methodology and Philosophy of Science, vol. 1 (Stanford: Stanford Univer-
sity Press, 1962).
18. 'Absolutism' can also designate the view that moral rules are exceptionless.
This sense of the term does not interest me here.
19. A recent collection of papers on this topic is Martin Hollis and Steven
Lukes, eds., Rationality and Relativism (Cambridge, Mass.: MIT Press, 1982).
20. Stephen Stich and Richard Nisbett, "Justification and the Psychology of
Human Reasoning," Philosophy of Science, 47 (1980):188-202.
21. Explicit formulation of such rules, or principles, is found in Roderick
Chisholm, Themy of Knowledge, 1st and 2nd eds. (Englewood Cliffs, N.J.: Pren-
tice-Hall, 1966, 1977).
22. Even radical methodological gulfs, such as those separating theologically
and scientifically oriented thinkers, may best be construed in objectivist, truth-
linked terms. For example, the religious thinker supposes that certain texts or al-
leged revelations are the best guides to truth on certain topics, whereas the scien-
tific theorist holds that scientific methodology is verifically optimal on all
subjects.
23. Critical discussions of moral relativism and subjectivism may be found in
the following: G. E. Moore, Ethics (London: Oxford University Press, 1912); Phi-
lippa Foot, "Moral Relativism," The Lindley Lecture, University of Kansas
(1978); Phillip Montague, "Are There Objective and Absolute Moral Standards?"
in Joel Feinberg, ed., Reason and Responsibility, 5th ed. (Belmont, Calif.: Wads-
worth, 1981); and Walter Stace, "Ethical Relativism," in ibid.
For an incisive critique of epistemological cultural relativism, see Hilary Put-
nam's discussion of Rorty's brand of relativism, in "Why Reason Can't Be Natu-
ralized," in Realism and Reason (Cambridge: Cambridge University Press, 1983).
24. See John Pollock, "A Plethora of Epistemological Theories," in George
Pappas, ed., Justijication and Knowledge (Dordrecht: D. Reidel, 1979).
25. Admittedly, I disclaim any general association between judgments of jus-
tifiedness and praise or blame (see Chapter 1, note 6). But we needn't exclude
praise and blame entirely. At any rate, casting the issue in these terms is a useful
device for crystallizing intuitions about the kind of case cited.
26. Philosophy and the Mirror of Nature, p. 170.
27. The distinction between primary and seconcZury justification will be ex-
pounded in section 5.3.
28. See Gilbert Harman, Thought (Princeton: Princeton Upiversity Press,
1973), pp. 27-29.
29. The only exceptions are some proposals on 'secondary' justifiedness.
30. A good exposition of the essentials of foundationalism (plus one kind of
variant on the main theory) can be found in William Alston, "Two Types of
Foundationalism," The Journal of Philosophy, 73 (1976):165-185.
31. See Roderick Firth, "Coherence, Certainty, and Epistemic Priority," The
Journal of Philosophy, 61 (1964):545-557, and Mark Pastin, "Modest Foundation-
alism and Self-warrant," American Philosophical Quarterly, monogr. ser. no. 4
(1975):141-149.
32. Keith Lehrer, Knowledge (Oxford: Oxford University Press, 1974), pp.
78-79.
33. Philosophy and the Mirror of Nature, for example, pp. 163,211-212.
34. This possibility has not been raised in the literature, to my knowledge.
35. I will examine some empirical evidence on this score in section 9.4.
5. Justifiation and Reliability

1. Belief retention over time is not automatically permitted. As new beliefs
arise, they may incriminate old ones, whose continuation is no longer appropri-
ate.
2. A system is right only if all the permissions it delivers, using any of its rules
or combinations thereof, are acceptable (though the permitted beliefs needn't all
be true). If a given member of system R delivers unacceptable permissions, per-
haps in combination with other members of R, then R is not right. Hence, if a
given rule is not capable of delivering only acceptable permissions when com-
bined with any other set of rules, then it does not belong to any right system.
3. Gilbert Harman, Thought(Princeton: Princeton University Press, 1973),p.
157.
4. Itis a delicate question just what 'abandonment' of belief consists in, psy-
chologically speaking. I shall broach this question in Chapter 10. Meanwhile, I
treat it as if it were unproblematic.
5. Proponents of relevance logic would deny that the argument from an in-
consistent set of premises to a random conclusion is valid. They would propose a

revision in the standard treatment of validity. (Cf. Alan Anderson and Nuel Bel-
nap, Jr., Entailment, vol. 1, Princeton: Princeton University Press, 1975.) But no
such revision is needed to save the intuition that a person is not justified in infer-
ring-forming a belief in-a random conclusion from inconsistent premises. All
that's needed is the recognition that facts of validity have no straightforward im-
plications for belief-forming license.
6. Strictly, the nonjustifiedness of Claude's belief in y only shows that there is
no right system containing the target rule plus the rules that licensed Claude's
belief in x. Does it show that there is no right system at all to which this rule be-
longs? Well, no constraints were placed on the system that licensed his belief in x.
Any arbitrary system could have been supposed. Hence, there is no system to
which the target rule might belong which would be right. The entire argument, of
course, rests on the presumed correctness of the framework principle, (Pl*) or
(P3.); but these seem quite unobjectionable.
7. There are nice metaphysical questions about the nature of processes. I am
not sure that calling processes 'causal chains' captures exactly what needs to be
said about processes. But since I lack satisfactory answers to the relevant ques-
tions, I will proceed without any further, detailed account of processes.
8. See Christopher Cherniak, "Feasible Inferences," Philosophy of Science,
48 (1981):248-268.
9. The point I have highlighted in this section has not gone wholly unno-
ticed. For example, Carl Hempel writes: "Formal logic tells us that if a given set
of statements is true then such and such other statements are true as well; but it
does not tell us what statements to believe or to act on. Indeed, the notion of ac-
cepting certain statements, like the notion of total evidence, is pragmatic in char-
acter and cannot be defined in terms of the concepts of formal deductive or
inductive logic" ("Inductive Inconsistencies," in Aspects of Scientific Explana-
tion, New York: Free Press, 1965, p. 66). Still, the significance of the point has not
been widely appreciated. Nor has the relevance of psychological processes been
generally acknowledged.
10. Rudolf Carnap, Logical Foundations of Probability, 2nd ed. (Chicago: Uni-
versity of Chicago Press, 1962).
11. I have misgivings about characterizing basic processes as unacquired pro- .
cesses. What about processes acquired through maturation, or the triggering of
innate propensities? Couldn't these still be considered basic processes? Perhaps
so. What I mainly mean to exclude from the class of processes are the acquired
skills, strategies, heuristics, and algorithms-especially task-specific ones-that
are encoded either as explicitly accepted instructions or as automatic procedures
(see Chapter 17). These are what I think of as 'methods'. A fully satisfactory for-
mulation of the process/method distinction, however, has eluded me. I shall rely
throughout on an informal grasp of this distinction.
12. See Daniel Kahneman, Paul Slovic, and Amos Tversky, eds., Judgment
under Uncertainty: Heuristics and Biases (Cambridge: Cambridge University
Press, 1982), and Richard Nisbett and Lee Ross, Human Inference: Strategiesand
Shorfcomings of Social Judgment (Englewood Cliffs, N.J.: Prentice-Hall, 1980).
This material will be explored in Part 11, especially Chapter 14.
392 Notes to Pages 97-1 00
13. On this point see Roderick Firth, "Epistemic Merit, Intrinsic and Instru-
mental," in P~oceedingsand Addresses of the Anerican Philosophical Association,
55 (1981):s-23.
14. Keith Lehrer's coherence theory of justification, in Knowledge (Oxford:
Oxford University Press, 1974), seems to be a specimen of weak coherentism.
With one qualification, to be explained below, Lehrer's theory says that a belief is
justified if the proposition believed is viewed by the cognizer as having a better
chance of truth than its competitors (see p. 212). But when a person believes any
proposition, he views it as having a better chance at truth than its competitors (at
least the competitors he thinks of). So even an idle dreamer's beliefs will satisfy
this condition. Lehrer's theory does contain the qualification that the belief
should be viewed from the cognizer's 'corrected' doxastic system, which is ob-
tained by deleting beliefs the cognizer would not retain as a "veracious inquirer,"
that is, as an "impartial and disinterested truth-seeker." But suppose a fantasizer
honestly believes--again, through fantasy-that fantasizing is the optimal
method for getting the truth. Then he is genuinely seeking truth (impartially and
disinterestedly); so his actual belief system is not subject to any correction. But
then the "veracious-inquirer" condition does no real work, and we are left with
the counterintuitive result obtained earlier.
Perhaps Lehrer would say that what counts as veracious belief formation must
not be judged by the inquirer's own lights, but by the facts of the matter. In that
case dreaming and fantasizing would not be methods of a veracious inquirer, and
this man's doxastic system would require drastic correction. But if this is the
proper interpretation of the view, it comes out more as closet reliabilism than as
subjective coherentism. Reference to the 'corrected doxastic system' really masks
a thoroughly reliabilist viewpoint.
Lehrer no longer subscribes to the view expressed in Knowledge. His more re-
cent position comes somewhat closer to reliabilism. This position is formulated in
"A Self Profile," in Radu Bogdan, ed., Keith Lehrer (Dordrecht: D. Reidel, 1979),
pp. 84-85. There he characterizes 'complete' justification as the conjunction of
'personal justifications-his old notion-plus 'verific justification'. Roughly, a be-
lief in p is 'verifically' justified just in case it coheres with the cognizer's true be-
liefs. But this notion of verific justification-and hence the notion of complete
justification-is excessively strong. A belief inferred from false premises is in dan-
ger of being unjustified solely because the premises are false. But, surely, the mere
falsity of previous beliefs should not endanger justifiedness. If believing those
premises was justified, and if the inferences were appropriate, then the resulting
beliefs should also be justified. Falsity of the premises endangers knowledge, but
not justifiedness. Hence, this strategy for introducing truth into the justification
picture does not work.
15. See Richard Jeffrey, The Logic of Decision, 2nd ed. (Chicago: University of
Chicago Press, 1983), chap. 11.
16. Lewis's term for this stronger notion of coherence is "congruence"; see An
Analysis of Knowledge and Valuation (La Salle, Ill.: Open Court, 1946), p. 338.
Chisholm's term for the same concept is "concurrence"; cf. Theory of Knowledge,
1st ed. (Englewood Cliffs, N.J.: Prentice-Hall, 1966),chap. 3; 2nd ed. (Englewood
Cliffs, N.J.: Prentice-Hall, 1977), chap. 4.
Notes to Pages 100-1 04 393
17. John Pollock also speaks of negative and positive coherence, in "A Pleth-
ora of Epistemological Theories," in George Pappas, ed., Jtcstijicationand Knowl-
edge (Dordrecht: D. Reidel, 1979). But his terminology marks a somewhat
different distinction.
18. See Carl Hempel, "Deductive-Nomologicalvs. Statistical Explanation," in
Herbert Feigl and Grover Maxwell, eds., Minnesota Studies in the Philosophy of
Science, vol. 3 (Minneapolis: University of Minnesota Press, 1962); Jaakko Hin-
tikka and J. Pietarinen, "Semantic Information and Inductive Logic," in Jaakko
Hintikka and Patrick Suppes, eds., Aspects of Inductive Logic (Amsterdam:
North-Holland, 1966);Isaac Levi, Gambling with Truth (New York: Knopf, 1967);
Risto Hilpinen, Rules of Acceptance and Inductive Logic (Amsterdam: North-
Holland, 1968); and Lehrer, Knowledge.
19. See Levi, Gambling with Truth, chap. 4 .
20. My framework principle, (P3*), employs a categorical notion of justifica-
tion, so I cannot easily incorporate this point into the current formulation of the
theory. But this feature could probably be revised to accommodate degrees of
justifiedness.
21. Apart from accounting for intuitions in all these cases, the truth ratio, or
reliabilist, theory of justifiedness holds out the prospect of coping satisfactorily
with the Humean riddle of induction. Although it does not fit naturally at this
juncture, let me sketch an approach to this problem, predicated on the correct-
ness of the reliabilist theory.
I propose to divide the traditional problem of induction into two problems.
The first is: Can we have justified beliefs based on inductive inference? The sec-
ond is the problem of second-order justifiedness: Can we have a justified belief
that we can have justified beliefs based on inductive inference? I think it quite
possible that Hume himself, and many followers of Hume, confused the first
problem with the second. At least they have held that a positive answer to the
first depends upon a positive answer to the second. I disagree with this conten-
tion. Nonetheless, the second problem is significant; and there is no reason why
one could not take the conjunction of these problems as the problem of induction.
If the reliability theory is correct, the first problem is not terribly serious. At
least it is 'logically' possible to have justified beliefs based on inductive inference,
whether or not it is humanly possible. Suppose we have a (basic) inductive process ,
with a sufficiently high truth ratio, at least a process whose output beliefs would
be true often enough if its input beliefs were true. Then this process could be
permitted by a right system of J-rules. Hence, beliefs generated by this process
could be justified, as long as the input beliefs to the process were justified. A criti-
cal point here is that our framework principle imposes no iterative requirement.
A person's belief is justified if it is permitted by a right rule system (I ignore the
no-undermining clause for simplicity); the person does not also have to be justi-
fied in believing that the belief is so permitted.
This takes care of the first problem of induction. But what about the second? If
we stick to the framework principle, there is no reason why a belief in the justifi-
cation-conferring power of induction could not be justified as a result of the self-
same inductive process! If the indicated inductive process is permitted by a right
rule system, then one might apply that same permitted process to beliefs in the ,
394 Notes to Page 104
process's past successes, and draw the conclusion that the process is successful
(reliable) in general. From this one could permissibly infer that the process is per-
mitted by right rules.
Of course, this inductive justification of induction is staunchly resisted by de-
fenders of Hume. The familiar charge is one of circularity. But how would this
charge be formulated in terms of my general theory? It would have to be couched
as a further restriction on justifiedness, to be incorporated into a new framework
principle. It might run: 'No belief about the permissibility of a process is justified
if the belief results from that selfsame process'. But does this restriction have any
plausibility? Surely not. It seems quite arbitrary. Would one say that a person
could not be justified in believing in the validity of modus ponens if he used
modus ponens to amve at this belief? (Here I waive the point that modus ponens
isn't really a process.) Acceptance of this stricture leads to the dilemma posed by
Lewis Carroll, in "What the Tortoise Said to Achilles," in Mind, vol. 4 (1895), pp.
278-280.But there is no good reason to accept it. If a process deserves to be per-
mitted, then its permission should extend to all subject matter, including its own
performance and its own permissibility.
Another attempt to address the problem of induction in a reliabilist vein is
made by James Van Cleve, in "Reliability, Justification, and Induction," in Peter
French, Theodore Uehling, Jr., and Howard Wettstein, eds., Midwest Studies in
Philosophy, vol. 9, Causation and Causal Theories (Minneapolis: University of
Minnesota Press, 1984). Van Cleve has an instructive discussion of circularity. He
says that an argument is viciously circular only if it is epistemically circular. He
then distinguishes premise circularity and rule circularity, as possible sources of
epistemic circularity. He indicates that the critical argument in the case of in-
duction does not suffer from premise circularity, and that rule circularity is not
really a source of epistemic circularity. I will not rehearse these arguments but
commend them to the interested reader.
22. This phrase is used by Hilary Kornblith, in "Justified Belief and Epistemi-
cally Responsible Action," The Philosophical Review, 92 (1983):33-48. However,
Kornblith uses it to express a different theory than the one intended in the text.
Kornblith's own theory is that a belief is justified just in case it is guided by a de-
sire to believe the truth. Such a theory is extremely weak, though. If a cognizer is
sufficiently benighted, such a desire may lead to very weird processes or 'evi-
dencecollecting' activities. It is questionable whether beliefs formed from such
processes or activities are justified. Furthermore, the theory may also be too
strong. Perceptual beliefs may presumably be justified; yet such beliefs are not
controlled or guided by a desire to believe the truth:
Let me return to the theory intended in the text: a truth-ratio maximization
theory, relativized to human cognitive endowments. There are various possible
refinements of this approach. Suppose, for example, that the best human percep-
tual processes can secure a truth ratio of .99, but the best human inductive pro-
cesses can only secure a truth ratio of .65. The simple maximizing criterion might
then disallow rule systems that sanction inductive processes, on the ground that
inductive processes would reduce the truth ratio below an otherwise attainable
maximum. This is (arguably) counterintuitive. However, the simple maximizing
criterion might be replaced by a more sophisticated criterion, which maximizes
within a category. For example, a criterion could license the best perceptual be-
lief-forming processes, the best inductive belief-forming processes, the best me-
morial belief-forming processes, and so on. This approach would encounter some
difficulties--how, exactly, should the categories be selected?-but it is still at-
tractive.
23. As formulated, (ARI) has a problem. It presupposes that each J-rule system
has a unique truth ratio. This is plausible if we consider only actual instantiations
of the permitted processes (by all cognizers, past, present, and future), assuming
these are finite. But presumably a truth ratio should be fixed by possible as well as
actual instantiations. For one thing, we want to allow a system to have a nonzero
truth-ratio propensity even though its authorized processes are never actually in-
stantiated. However, this makes the assumption of a unique truth ratio extremely
doubtful. There are many possible ways of instantiating a permitted set of pro-
cesses. Because of the permissive nature of the rules, one pattern of instantiation
might use a given subset of permitted processes (for example, inductive processes)
very often, while another pattern of instantiation might use this subset sparingly.
Let us call any pattern of permitted instantiations of a given system a compliance
profile of that system. The problem is that it is most unlikely that all compliance
profiles of a system have the same truth ratio. But if a rule system's truth ratio is
not inuuriant under diverse compliance profiles, the system does not have a
unique truth ratio.
One way to handle this problem is to require, for rightness, that all possible
compliance profiles of the target system meet the threshold truth ratio. (An ex-
ception could be made for the 'null' profile, which does not use any of the per-
mitted processes. This is a compliance profile in the sense that it does not uiolate
the rule system. But the truth ratio of this profile is 0.)This is a very stringent re-
quirement. A milder way of handling the problem is to require only that the me-
dian truth ratio of all the compliance profiles meet the threshold. But this seems
too weak. The worst truth ratio might be too low.
A third solution is to adopt the previous condition as necessary, but add a fur-
ther requirement that the amplitude of departures from the median value must
not be too great. To borrow a term from Brian Skyrms, the rule system should be
quite resilient. (Cf. Causal Necessity, New Haven: Yale University Press, 1980, p.
11.However, Skyrms's notion of resiliency is defined for probabilities.) Obviously,
many other measures of these sorts could be proposed, so the problem appears sol-
uble. But I will not try to pick a particular solution; I will simply assume that
(ARI) can be reformulated with a satisfactory solution.
Notice that similar problems would arise for resource-relative criteria, such as a
maximizing criterion. In fact, here the problems would be trickier because one
would have to compare various systems, each with multiple compliance profiles.
One possible solution is a 'maximax' solution: the system with the highest maxi-
mum compliance profile is best. Another possible solution is a 'maximin' condi-
tion: the system with the highest minimum compliance profile is best. Again, I
will not try to pick any particular solution from these candidates.
24. I assume that a system's truth ratio is relatively invariant across different
normal worlds.
25. Saul Kripke, Naming and Necessity (Cambridge, Mass.: Harvard Univer-
sity Press, 1980), and Hilary Putnam, "The Meaning of 'Meaning'," in Mind,
Languuge and Reality (Cambridge: Cambridge University Press, 1975).
26. Peter Unger, Philosophical Relativity (Minneapolis: University of Minne-

sota Press, 1984), pp. 84ff.Unger does not endorse normal-world chauvinism. His
own positive theory is quite different and ultimately supports a relativistic ap-
proach to conceptual analysis. I am not embracing this aspect of his discussion.
27. Hilary Putnam, "It Ain't Necessarily So," in Mathematics, Matter and
Method (Cambridge: Cambridge University Press, 1975).
28. Hilary Putnam, "Why Reason Can't Be Naturalized," in Realism and
Reason (Cambridge: Cambridge University Press, 1983).
29. Laurence BonJour, "Externalist Theories of Empirical Knowledge," in
Peter French, Theodore Uehling, Jr., and Howard Wettstein, eds., Midwest Stud-
ies in Philosophy, vol. 5, Studies in Epistemology (Minneapolis: University of
Minnesota Press, 1980).
30. See, for example, my "What Is Justified Belief?" in George Pappas, ed.,
Justification and Knowledge (Dordrecht: D. Reidel, 1979). Actually, BonJour fo-
cuses his criticism on Armstrong's version of reliabilism, in Belief, Truth and
Knowledge (Cambridge: Cambridge University Press, 1973),although Armstrong
does not advance a theory of iusti;fication.
31. See Keith Lehrer and Stewart Cohen, "Justification, Truth, and Coher-
ence," Synthese, 55 (1983):191-207; John Pollock, "Reliability and Justified Be-
lief," Canadian Journal of Philosophy, 14 (1984):103-114; Stewart Cohen,
"Justification and Truth," Philosophical Studies, 46 (1984):279-295.
32. At times I have been tempted to handle demon-world cases by saying that
beliefs in that world are not 'really' justified; they are merely 'apparently' justified
(see section 4.3). If this response were intuitively plausible, reliabilism could dis-
pense with the normal-worlds theory. However, while some people find this solu-
tion intuitively appealing, many do not.
33. Sarah Lichtenstein, Baruch Fischhoff, and Lawrence Phillips, "Calibration
of Probabilities: The State of the Art to 1980," in Daniel Kahneman, Paul Slovic,
and Amos Tversky, eds., Judgment under Uncertainty: Heuristics and Biuses
(Cambridge: Cambridge University Press, 1982), p. 307.
34. The main problem is whether good calibration is su$%ient for justified-
ness. The following example illustrates the problem. Suppose a weather forecaster
predicts a 20 percent chance of rain every day, despite the fact that meteorologi-
cal indicators differ from day to day. If the long-run proportion of rainy days in
the area is 20 percent, this forecaster's predictions are well calibrated, by defini-
tion. But it is questionable whether these predictions are justified, or otherwise
normatively sound.
One possible way of strengthening the calibration requirement is suggested by
Morris DeGroot and Stephen Fienberg, in "Assessing Probability Assessors," in
S. Gupta and J. Berger, eds., Statistical Decision Theory and Related Topics, III,
vol. 1(New York: Academic Press, 1982). Following DeGroot and Fienberg, let m
denote the relative frequency of days on which it rains, and let v(x) be the proba-
bility that a forecaster's prediction on a randomly chosen day will be x (or the
proportton of days on which his prediction is x). We can now characterize the
previous weather forecaster as having the following probability function:
v(m) = 1,
v(x) = 0 for x # m.
In other words, where the relative frequency of rain is -20,the probability of this .
forecaster making a 20 percent chance of rain prediction on a given day is 1. This
man is a least refined, though well-calibrated, forecaster. Contrast him with an-
other well-calibrated forecaster, characterized by the following probability func-
tion:
v(1) = m,
v(0) = 1-m,
v(x) = O f o r x Z 0 , 1.
The only probabilities of rain that this forecaster ever specifies are 0 and 1, and
since he is well calibrated, his predictions are always correct. This man is referred
to as a mst-refined forecaster. (In meteorology, the first forecaster is said to ex-
hibit zero sharpness and the second is said to exhibit perfect sharpness.)
Now it clearly seems too strong to require, for justificational rightness, that a
rule system yield mariml refinement as well as good calibration. But some
threshold of refinement might be set, as a further condition for rightness beyond
good calibration. This seems feasible since DeGroot and Fienberg develop a con-
dition for determining whether one well-calibrated forecaster is more refined
than another.
For another attempt to use calibration to explicate epistemic standards, see Bas
van Fraassen, "Calibration: A Frequency Justification for Personal Probability,"
in R. Cohen and L. Laudan, eds., Physics, Philosophy and Psychoanalysis (Dor-
drecht: D. Reidel, 1983).
35. Hilary Putnam, "Why Reason Can't Be Naturalized," in Realism and
Reason (Cambridge: Cambridge University Press, 1983), p. 231.
36. Roderick Firth, "Epistemic Merit, Intrinsic and Instrumental," p. 19.
37. Ibid., p. 17.
6. Problem Solving, Power, and Speed

1. My emphasis on relief from ignorance bears some similarity to Isaac Levi's
emphasis on relief from "agnosticism" (see his Gambling with Truth, New York:
Knopf, 1967). But the latter denotes relief from the absence of belief (or from the
weakness of belief) while the former denotes relief from the absence of true belief.
2. Larry Laudan, Progress and Its Problems (Berkeley: University of Califor-
nia Press, 1977).
3. John Dewey, Logic: The Theory of Inguiy (New York: Henry Holt, 1938).
4. Karl Popper, Objective Knowledge (Oxford: Oxford University Press,
1972);Thomas Kuhn, The Structure of Scientijk Revolutions (Chicago: University
of Chicago Press, 1962); Imre Lakatos, "Falsification and the Methodology of Sci-
entific Research Programmes," in Imre Lakatos and Alan Musgrave, eds., Criti-
cism and the Growth of Knowledge (Cambridge: Cambridge University Press,
1970); and Laudan, Progress and Its Problems.
5. Jaakko Hintikka, "On the Logic of the Interrogative Model of Scientific
Inquiry," Synthese, 47 (1981):69-83.
6. Sylvain Bromberger, "Why-Questions," in Robert Colodny, ed., Mind and
Cosmos (Pittsburgh: University of Pittsburgh Press, 1966); and Bas van Fraassen,
The Scientific Image (Oxford: Oxford University Press, 1980).
7. "Why-Questions," pp. 90-91.
8. Nuel Belnap, Jr., and Thomas Steel, Jr., The Logic of Questions and An-
swers (New Haven: Yale University Press, 1976).
9. For further treatment of the semantics of questions and answers, see
Jaakko Hintikka, "The Semantics of Questions and the Questions of Semantics,"
Acta Philosophical Fennica, 28 (1976), issue 4.
10. Discussions of problem solving in artificial intelligence portray it as get-
ting to an end state from some initial state. For surveys of this topic, see Patrick
Winston, ArtificMl Intelligence (Reading, Mass.: Addison-Wesley, 1977); Nils
Nilsson, Principles of Artijicial Intelligence (Palo Alto, Calif.: Tioga, 1980); and
Elaine Rich, ArtificMl Intelligence (New York: McGraw-Hill, 1983).
11. See Belnap and Steel, The Logic of Questions and Answers, pp. 5,109-121.
12. Progress and Its Problems, p. 24.
13. Cf. Lakatos, "Falsification and the Methodology of Scientific Research
Programmes"; Laudan, Progress and Its Problems; and Thomas Nickles, "Scien-
tific Discovery and the Future of Philosophy of Science," in Thomas Nickles, ed.,
Scientific Discovery, Logic, and Rationality (Dordrecht: D. Reidel, 1980).
14. Actually, pursuit may not be wholly different from the components men-
tioned above. A decision to pursue a specific research program is arguably predi-
cated on the cognizer's judgment of which program is most promising (in my
interpretation, most likely to lead to truth), which is a species of doxastic decision.
Having selected a research program, the scientist usually needs to generate new
hypotheses under the aegis of that program and then devise tests of these hypoth-
eses. So pursuit can itself be broken down into subcomponents, perhaps of the
same variety as previously mentioned.
15. Paraphrased from Michael Posner, Cognition: An Introduction (Glenview,
Ill.: Scott, Foresman, 1973), pp. 150-151.
16. Alternative modes of framing outcomes also have significant ramifications
for choice behavior, as shown in Amos Tversky and Daniel Kahneman, "The
Framing of Decisions and the Psychology of Choice," Science, 211
(1981):453-458, and Daniel Kahneman and Amos Tversky, "Choices, Values, and
Frames," American Psychologist, 39 (1984):341-350. However, representations
for behavioral choice do not concern me here.
17. G. Polya, How to Solve It, 2nd ed. (Princeton: Princeton University Press,
1971), pp. 10, 209.
18. Cognition, pp. 151-152. The example is drawn from C. N. Cofer, "The
Role of Language in Human Problem Solving," paper presented at a conference
on human problem solving, New York University, 1954.
19. Cf. A. de Groot, Thought and Choice in Chess (The Hague: Mouton, 1965),
and William Chase and Herbert Simon, "The Mind's Eye in Chess," in William
Chase, ed., Visual Infomution Processing (New York: Academic Press, 1973).
20. Walter Kintsch and T. A. van Dijk, "Toward a Model of Text Compre-
hension and Production," ,PsychologicalReoiew, 85 (1978):363-394.
21. Mary Gick and Keith Holyoak, "Analogical Problem Solving," Cognitive
Psychobgy, 12 (1980):306355.
22. For reviews of artificial intelligence approaches to strategies, see the
works cited in note 10.
23. The relevance of real-time constraints is discussed further in Chapter 13.
24. Actually, this is open to challenge. On some conceptions justificational sta-

tus depends on how well one collects evidence. See Hilary Kornblith, "Justified
Belief and Epistemically Responsible Action," Philosophical Review, 92
(1983):33-48. It may also depend on how well one generates alternative hypoth-
eses. I explore this suggestion in Chapter 11.
7. Truth and Realism

1. Putnam's principal writings in this area are: Meaning and the Moral Sci-
ences (London: Routledge and Kegan Paul, 1978); Reason, Truth and History
(Cambridge: Cambridge University Press, 1981); and several papers in Realism
and Reason (Cambridge: Cambridge University Press, 1983). Dummett's major
works in this area are: several papers in Truth and Other Enigmas (Cambridge,
Mass.: Harvard University Press, 1978); "What Is a Theory of Meaning? (II)," in
Gareth Evans and John McDowell, eds., Truth and Meaning (Oxford: Oxford Uni-
versity Press, 1976);Elements of Intuitionism (New York: Oxford University Press,
1977); and "Realism," Synthese, 52 (1982):55-112.
2. Dummett, "Realism," p. 55.
3. Ibid., p. 55 and elsewhere.
4. Dummett comments on the case of empty singular terms, but says this is a
kind of antirealism, the contrasting opposite of which is a realism of a Meinongian
kind. I find this implausible.
5. See Hilary Putnam, "On Truth," in Leigh Cauman et al., How Many
Questions? (Indianapolis: Hackett, 1983). Also cf. Scott Soames, "What Is a The-
ory of Truth?" The Journal of Philosophy, 81 (1984):411-429.
6. Reason, Truth and History, chap. 3.
7. "Realism and Reason," in Meaning and the Moral Sciences.
8. Meaning and the Moral Sciences, p. 125.
9. See Dummett, "Realism," p. 69.
10. C. S. Peirce, "How to Make Our Ideas Clear," in Justus Buchler, ed., Philo-
sophical Writings of Peirce (New York: Dover, 1955), p. 38. Note that this is,
strictly, a doxastic definition,not an epistemic one. It only invokes agreement in
opinion, whether or not this is justified or rational.
11. Reason, Truthand History, p. 55.
12. Meaning and the Moral Sciences, p. 125.
13. Brian Loar, "Truth Beyond All Verification," in Barry Taylor, ed., Essays
on Michael Durnmdt (The Hague: Martinus Nijhoff, forthcoming).
14. See William Wimsatt, "Robustness, Reliability and Multiple-Determina-
tion in Science," in M. Brewer and B. Collins, eds., Knowing and Validating in
the Social Sciences (San Francisco: Jossey-Bass, 1981). Among several works by
Campbell, he cites D. T. Campbell and D. Fiske, "Convergent and Discriminant
Validation by the Multi-Trait Multi-Method Matrix," Psychological Bulletin, 56
(1959):81-105, and T. D. Cook and D. T. Campbell, Quasi-Experimentation (Chi-
cago: Rand-McNally, 1979).The relevant works by R. Levins are "The Strategy of
Model Building in Population Biology," American Scientist, 54 (1966):421-431,
and Evolution in Changing Enuironments (Princeton: Princeton University Press,
1968).
15. See especially "What is a Theory of Meaning? (II)," p. 116.

16. Some critical discussions of Dummett's proposals include Loar, "Truth Be-
yond All Verification"; Colin McGinn, "Realist Semantics and Content-Ascrip-
tion," Synthese, 52 (1982):113-134; Michael Devitt, "Dummett's Anti-Realism,"
The Journal of Philosophy, 80 (1983):73-99; and Gregory Currie and Peter Eg-
genberger, "Knowledge and Meaning," Nous, 17 (1983):267-279.
17. Ludwig Wittgenstein, Tractatus Logico-Philosophicus, trans. C, K. Ogden
and F. P. Ramsey (London: Routledge and Kegan Paul, 1922).
18. Reason, Truth and History, pp. 52-53. Note, however, that in earlier work
Putnam had endorsed the notion of kinds being in the world. See "On Properties,"
in Nicholas Rescher et al., eds., Essays in Honor of Carl G. Hempel (Dordrecht: D.
Reidel, 1970).
19. Nelson Goodman, Ways of Worldmaking (Indianapolis: Hackett, 1978).
20. Ibid., p. 138.
21. Cf. Hilary Putnam, "Reflections on Goodman's Ways of Worldmaking,"
The ]oumal of Philosophy, 76 (1979):611.
22. See Putnam, "Realism and Reason," pp. 123-126, in Meaning and the
Moral Sciences, and Reason, Truth and History, chap. 2.
23. Alvin Plantinga, "How To Be an Anti-Realist," in Proceedings and Ad-
dresses of the American Philosophical Association, 56 (1982):47-70.
24. A similar characterization of scientific realism, minus the third thesis, is
given by Bas van Fraassen, in The Scientific Zrnuge (Oxford: Oxford University
Press, 1980):"Science aims to give us, in its theories, a literally true story of what
the world is like; and acceptance of a scientific theory involves the belief that it is
true" (p. 8).
25. More generally, Michael Dummett has observed that various kinds of re-
ductionism~can be compatible with realisms about those domains (see "Realism,"
p. 75).
26. Cf. Putnam, Lecture 11, in Meaning and the Moral Sciences, where he
draws on unpublished work by Boyd; Richard Boyd, "Scientific Realism and Nat-
uralistic Epistemology," in P. D. Asquith and R. N. Giere, eds., PSA 1980, vol. 2
(East Lansing, Mich.: Philosophy of Science Association, 1981): and W. H. New-
ton-Smith, The Rationality of Science (London: Routledge and Kegan Paul, 1981).
27. Larry Laudan raises doubts about the historical record in "A Confutation
of Convergent Realism," Philosophy of Science, 48 (1981):19-49.
28. Meaning and the Moral Sciences, pp. 24-2!5,37.
29. Parenthetically, we should be reminded that past applications of a method
do not exhaust the class of applications relevant to reliability. These would in-
clude present and future cases as well as counterfactual cases. But there is no rea-
son to suppose that historical uses of scientific methods would be
unrepresentative of long-run truth properties, as long as the very same methods
are in question.
30. This point has a connection with what has been called "verisimilitude," or
"truth-likeness." However, I do not mean to commit myself to any particular defi-
nition of verisimilitude, or to the prospects for a workable definition along the
lines initiated by Popper. This is a very complicated issue. For some proposed
definitions and critiques of versimilitude, see Karl Popper, Objective Knowledge
(Oxford: Oxford University Press, 1972);Ilkka Niiniluoto, "Truthlikeness in First-

Order Languages," in Jaakko Hintikka et al., eds., Essays on Mathematical and
Philosophical Logic (Dordrecht: D. Reidel, 1979), and "Scientific Progress,"
Synthese, 45 (1980):427-462; W. H. Newton-Smith, The Rationality of Science;
Pave1 Tichy, "On Popper's Definition of Verisimilitude," British Journal for the
Philosophy of Science, 25 (1974):155-160, and "Verisimilitude Revisited," Synth-
ese, 38 (1978):175-196; and Laudan, "A Confutation of Convergent Realism."
8. The Problem of Content

1. See W. V. Quine, The Roots of Reference (La Salle, Ill.: Open Court, 1973),
p. 35.
2. In "Mental Representation," in Ned Block, ed., Readings in Philosophy of
Psychology, vol. 2 (Cambridge, Mass.: Harvard University Press, 1981), Hartry
Field takes the belief relation to be a composite of two other relations: a relation
between a person and an internal sentence, and a relation between the sentence
and a proposition. For a recent defense of the view that takes propositions to be
the only objects of belief, see Robert Stalnaker, Inquiry (Cambridge, Mass.: MIT
Press, 1984).
3. Hilary Putnam, "Computational Psychology and Interpretation Theory,"
in Realism and Reason (Cambridge: Cambridge University Press, 1983).
4. Paul Churchland, "Eliminative Materialism and the Propositional Atti-
tudes," The Journal of Philosophy, 78 (1981):67-90. Also see Patricia Smith
Churchland, "A Perspective on Mind-Brain Research," The Journal of Philosophy,
77 (1980):185-207.
5. Imre Lakatos, "Falsification and the Methodology of Scientific Research
Programmes," in Imre Lakatos and Alan Musgrave, eds., Criticism and the
Growth of Knowledge (Cambridge: Cambridge University Press, 1970).
6. Cf. Ned Block and Jerry Fodor, "What Psychological States Are Not," in
Ned Block, ed., Readings in Philosophy of Psychology, vol. 1 (Cambridge, Mass.:
Harvard University Press, 1980); Ned Block, "Troubles with Functionalism," in
ibid.; Joseph Owens, "Functionalism and Propositional Attitudes," Nous, 17
(1983):529-549; Stephen Stich, From Folk Psychology to Cognitioe Science: The
Case Against Belief (Cambridge, Mass.: MIT Press, 1983); and Stephen Schiffer,
Remnants of Meaning, in preparation.
7. See Ernest Nagel, The Structure of Science (New York: Harcourt, Brace &
World, 1961), chap. 11.
8. Kenneth S c h h e r , "Approaches to Reduction," Philosophy of Science, 34
(1967):137-147.
9. See David Hull, Philosophy of Biological Science (Englewood Cliffs, N.J.:
Prentice-Hall, 1974), chap. 1. Hull himself, though, is cautious about any such
derivability.
10. Jerry Fodor, "Special Sciences, or The Disunity of Science as a Working
Hypothesis," in Block, ed., Readings in Philosophy of Psychology, vol. 1.
11. Stich, From Folk Psychology to Cognitioe Science, pp. 135-136.
12. Daniel Dennett, Content and Consciousness (New York: Humanities Press,
1969), p. 183.
13. See Inquiry, pp. 62-67.

14. Saul Kripke, "A Puzzle about Belief," in Avishai Margalit, ed., Meaning
and Use (Dordrecht: D. Reidel, 1979).
15. For further discussion of this point, see Brian Loar, "De Dido Ascription
and the Content of Names in Thought," forthcoming, and Ruth Barcan Marcus,
"Rationality and Believing the Impossible," The Journal of Philosophy, 80
(1983):321-338.
16. Tyler Burge, "Individualism and the Mental," in Peter French, Theodore
Uehling, Jr., and Howard Wettstein, eds., Midwest Studies in Philosophy, vol. 4,
Studies in Metaphysics (Minneapolis: University of Minnesota Press, 1979).
17. Terence Horgan and James Woodward, "Folk Psychology Is Here to
Stay," The Philosophical Reuiew, 94 (1985):197-226.
18. From Folk Psychology to Cognitive Science, p. 231.
19. I obviously have not exhausted all challenges to content. Two salient ex-
amples are W. V. Quine's indeterminacy of translation argument, in Word and
Object (Cambridge, Mass: MIT Press, 1960), chap. 2, and Saul Kripke's problems
about meaning or intending, in Wittgenstein on Rules and Private Language
(Cambridge, Mass.: Harvard University Press, 1982).These arguments would re-
'quire extensive examination, which cannot be undertaken here. Among the many
discussions of Quine's indeterminacy thesis, a recent one I find congenial is Peter
Unger's, in Philosophical Relativity (Minneapolis: University of Minnesota Press,
1984), pp. 15-20. This draws partly on Hilary Putnam's paper "The Refutation of
Conventionalism," in Mind, Language and Reality (Cambridge: Cambridge Uni-
versity Press, 1975).
20. See Hilary Putnam, "The Meaning of 'Meaning'," in Mind, Languuge and
Reality.
21. See Jerry Fodor, "Methodological Solipsism Considered as a Research
Strategy in Cognitive Psychology," in Representations (Cambridge, Mass.: MIT
Press, 1981).
22. Ibid., p. 227.
23. Ibid., p. 231.
24. "Propositional Attitudes," in Representations, p. 183.
25. Representations, p. 30.
26. This interpretation was suggested to me by Robert Harnish.
27. From Folk Psychology to Cognitive Science, chap. 8.
28. Cf. Tyler Burge, "Individualism and Psychology," presented at the 1984
Sloan Conference at the Massachusetts Institute of Technology.
29. This is suggested by Hartry Field, in "Thought Without Content," pre-
sented at the 1984 Sloan conference at the Massachusetts Institute of Technol-
ow.
30. See Zenon Pylyshyn, Computation and Cognition (Cambridge, Mass.: MIT
Press, 1984).
31. Robert Cummins, The Nature of Psychological Explanation (Cambridge,
Mass.: MIT Press, 1983).
32. Word and Object, p. 59.
33. Donald Davidson, "Psychology as Philosophy," in Essays on Actions and
Events (Oxford: Oxford University Press, 1980), p. 237.
34. Ibid., p. 237.

35. Donald Davidson, "On the Very Idea of a Conceptual Scheme," in Zn-
quiries into Truth and Interpretation (Oxford: Oxford University Press, 1984), p.
197.
36. Ibid., p. 196.
37. "Thought and Talk," in Inquiries, pp. 168-169.
38. Daniel Dennett, "Three Kinds of Intentional Psychology," in R. Healey,
ed., Reduction, Time and Reality (Cambridge: Cambridge University Press, 1981).
39. Stephen Schiffer, "Truth and the Theory of Content," in H. Parret and
J. Bouverese, eds., Meaning and Understanding (Berlin:Walter de Gruyter, 1981);
Field, "Mental Representation," sect. 5; Brian Loar, Mind and Meaning (Cam-
bridge: Cambridge University Press, 1981), chap. 8; Fred Dretske, Knowledge
a d the Flow of Znfmtion (Cambridge, Mass.: MIT Press, 1981); and Jerry
Fodor, "Psychosemantics, or: Where Do Truth Conditions Come From?'forth-
coming. At least one of these writerdchiffer-no longer subscribes to his ear-
lier position.
40. Christopher Cherniak, "Minimal Rationality," Mind, 90 (1981):161-183.
41. This and other criticisms of charity principles are presented by Paul Tha-
gard and Richard Nisbett, in "Rationality and Charity," Philosophy of Science, 50
(1983):250-267.
42. For other arguments to this effect, see Colin McGinn, "Charity, Interpre-
tation, and Belief," The Journal of Philosophy, 74 (1977):521-535.
9. Perception
1. Irvin Rock, The Logic of Perception (Cambridge, Mass.: MIT Press, 1983);
see chap. 1, especially pp. 17-19. David Marr is even better known for positing
multiple perceptual stages and outputs (see his Vision, San Francisco: W. H.
Freeman, 1982). But Marr's perceptual stages do not exemplify the kind of dis-
tinction I wish to emphasize here.
2. For example, see Marr, Vision, pp. 339-343.
3. D. H. Hubel and T. N. Weisel, "Receptive Fields, Binocular Interaction,
and Functional Architecture in the Cat's Visual Cortex," Journal of Physiology,
166 (1962):106-154; J. Y. Lettvin, H. R. Maturana, W. S. McCulloch, and W. H.
Pitts, "What the Frog's Eye Tells the Frog's Brain," Proceedings of the IRE, 47
(1959):1940-51.
4. G. C. Kinney, M. Marsetta, and D. J. Showman, Studies in Display Symbol
Legibility, pt. 21, "The Legibility of Alphanumeric Symbols for Digitalized Tele-
vision" (Bedford, Mass.: Mitre Corp., 1966), no. ESD-TR-66-117.
5. My discussion here and below draws on John R. Anderson, Cognitive Psy-
chology and Its Zmplicatbns (San Francisco: W. H. Freeman, 1980).
6. R. M. Warren, "Perceptual Restorations of Missing Speech Sounds," Sci-
ence, 167 (1970):392-393.
7. G. A. Miller and S. Isard, "Some Perceptual Consequences of Linguistic
Rules," Journal of Verbal Learning and Verbal Behavior, 2 (1963):217-228.
8. Jerry Fodor, The Modularity of Mind (Cambridge, Mass.: MIT Press,
1983).
9. Ibid., p. 69; emphases are mine.

10. Zenon Pylyshyn, Computation and Cognition (Cambridge, Mass.: MIT
Press, 1984).
11. See Rock, The Logic of Perception, pp. 12-13.
12. Rock also clearly acknowledges this (see Ibid., chap. 11).
13. Thomas Kuhn, The Structure of Scientific Revolutions (Chicago: Univer-
sity of Chicago Press, 1962). Other well-known proponents of the theory-laden-
ness of perception include N. R. Hanson, Patterns of Discmery (Cambridge:
Cambridge University Press, 1958),and Nelson Goodman, Ways of Worldmaking
(Indianapolis: Hackett, 1978).
14. This point is elaborated in Jerry Fodor, "Observation Reconsidered," Phi-
losophy of Science, 51 (1984):23-43,an article I encountered after completing the
bulk of this chapter.
15. G. Reicher, "Perceptual Recognition as a Function of Meaningfulness of
Stimulus Material," Journal of Experimental Psychology, 81 (1969):275-280.
16. D. D. Wheeler, "Processes in Word Recognition," Cognitiue Psychology, 1
(1970):59-85.
17. E. Tulving, G. Mandler, and R. Baumal, "Interaction of Two Sources of
Information in Tachistoscopic Word Recognition," Canadian Journal of Psychol-
ogy, 18 (1964):62-71.
18. The Logic of Perception, chap. 2.
19. See especially J. J. Gibson, The Perception of the Visual World (Boston:
Houghton Mifflin, 1950) and The Senses Considered as Perceptual Systems (Bos-
ton: Houghton Mifflin, 1966).
20. It does not matter to my present considerations whether the hypothesized
top-down processing is intramodular or extramodular.Fodor has argued that stud-
ies of context effects in speech perception-studies of Cloze values-are compati-
ble with his view that speech perception has access only to information within a
language perception module. What is being accessed, according to Fodor, is only
information about lexical items, something internal to the language recognition
module (see Modularity of Mind, pp. 77-80). This issue is not critical for my pre-
sent purposes, however. What concerns me here is the implications of top-down
processing for speed and reliability, whether the top-down processing is intramo-
dular or not.
21. Modularity of Mind, p. 70.
22. The Logic of Perception, p. 2.
23. The first two of these assumptions are built into John Anderson's model in
The Architecture of Cognition (Cambridge, Mass.: Harvard University Press,
1983). He expresses them in the terminology of "productions" and production
"application." I bypass the productions formulation here; but I will return to pro-
duction systems, and to the theme of partial matching, in Chapter 17.
24. A good presentation of F-ism along these lines is found in William Alston,
"Two T.ses of Foundationalism," The Journal of Philosophy, 73 (1976):165-185.
Also see Ernest Sosa, "The Raft and the Pyramid: Coherence versus Foundations
in the Theory of Knowledge," in Peter French, Theodore Uehling, Jr., and
Howard Wettstein, eds., Midwest Studies in Philosophy, vol. 5, Studies in Episte-
mology (Minneapolis: University of Minnesota Press, 1980).
25. On this topic see Roderick Firth, "Coherence, Certainty, and Epistemic
Priority," The Journal of Philosophy, 61 (1964):545-557, and Mark Pastin, "Mod-
est Foundationalism and Self-warrant," in George Pappas and Marshall Swain,
eds., Essays on Knowledge and Justification (Ithaca, N.Y .: Cornell University
Press, 1979).
26. See John Pollock, "A Plethora of Epistemological Theories," in George
Pappas, ed., Justification and Knowledge (Dordrecht: D. Reidel, 1979).
27. Cf. Roderick Chisholm's notion of "concurrence," in Theory of Knowl-
edge, 2nd ed. (Englewood Cliffs, N.J.: Prentice-Hall, 1977), p. 83. I referred to
other versions of positive coherence in Chapter 5.
28. See Pollock, "A Plethora of Epistemological Theories," p. 102.
29. Since there can be multiple right J-rule systems, on my theory, one right
system might be F-ist and another right system might be C-ist. This suggests the
commonly ignored possibility that both F-ism and C-ism might be right. But it is
also possible that F-ism could be viewed as the doctrine that all right J-rule sys-
tems have a certain F-ist property; and C-ism could say that all right J-rule sys-
tems have a certain C-ist property. On this approach the possibility of multiple
right systems would not provide for the cotenability of F-ism and C-ism.
30. See Fodor, Modularity of Mind, p. 107.
10. Memory
1. For example, see my A Theory of Human Action (Englewood Cliffs, N.J.:
Prentice-Hall, 1970; Princeton: Princeton University Press, 1976), chap. 4.
2. See Robert Crowder, "The Demise of Short-Term Memory," Acta Psycho-
bgica, 50 (1982):291-323.
3. John Anderson, The Architecture of Cognition (Cambridge, Mass.: Harvard
University Press, 1983), chaps. 1, 3.
4. I argued for this sortof distinction in an earlier paper, "Epistemology and
the Psychology of Belief," The Monist, 61 (1978):525-535.
5. See Daniel Dennett, "Brain Writing and Mind Reading," in Brainstom
(Montgomery, Vt.: Bradford Books, 1978).
6. Ibid., p. 45.
7. For example, see Robert Cummins, "Inexplicit Representations," forth-
coming in Myles Brand and Robert Harnish, eds., The Representation of Knowl-
edge and Belief (Tucson: University of Arizona Press).
8. This case is borrowed from Christopher Cherniak, "Rationality and the
Structure of Human Memory," Synthese, 57 (1983):163-186.
9. Rudolf Carnap, Logical Foundationsof Probability, 2nd ed. (Chicago: Uni-
versity of Chicago Press, 1962), and Carl Hempel, "Inductive Inconsistencies," in
Aspects of Scientific Explanation (New York: Free Press, 1965).
10. "Inductive Inconsistencies," p. 64; my emphases.
11. Cf. Cherniak, "Rationality and the Structure of Human Memory," pp.
164-165.
12. Gilbert Harman, Thought (Princeton: Princeton University Press, 1973),
pp. 158-159.
13. The complexity of the time course of culpability is discussed by Holly
Smith in "Culpable Ignorance," The Philosophical Reuiew, 92 (1983):543-571. It

is clear that normative judgments in the epistemic sphere should be sensitive to
such complexities of time.
14. J. D. Bransford, J. R. Barclay, and J. J. Franks, "Sentence Memory: A Con-
structive versus Interpretive Approach," Cognitiue Psychology, 3 (1972):193-209.
15. R. A. Sulin and D. J. Dooling, "Intrusion of a Thematic Idea in Retention
of Prose," Journal of Eqwimental Psychology, 103 (1974):255-262.
16. D. J. Dooling and R. E. Christiaansen, "Episodic and Semantic Aspects of
Memory for Prose," Journal of Experimental Psychology: Human Learning and
Memory, 3 (1977):428-436.
17. T. S. Hyde and J. J. Jenkins, "Recall for Words as a Function of Semantic,
Graphic, and Syntactic Orienting Tasks," Journal of Verbal Learning and Verbal
Behauior, 12 (1973):471-480.
18. Elizabeth Loftus, Eyewitness Testimony (Cambridge, Mass.: Harvard Uni-
versity Press, 1979).
19. R. N. Shepard, "Recognition Memory for Words, Sentences and Pictures,"
Journal of Verbal Learning and Verbal Behaviot, 6 (1967):156-163.
20. C. Bird, "The Iduence of the Press upon the Accuracy of Report," Jour-
nal of Abnormal and Social Psychology, 22 (1927):123-129.
21. E. F. Loftus, D. G. Miller, and H. J. Bums, "Semantic Integration of
Verbal Information into a Visual Memory," Journal of Experimental Psychology:
Human Learning and Memory, 4 (1978):19-31.
22. F. I. M. Craik and R. S. Lockhart, "Levels of Processing: A Framework for
Memory Research," Journal of Verbal Learning and Verbal Behavior, 11
(1972):671-684.
23. S. Bobrow and G. H. Bower, "Comprehension and Recall of Sentences,"
Journal of Experimental Psychology, 80 (1969):455-461.
24. John Anderson, Cognitiue Psychology and Its Implications (San Francisco:
W. H. Freeman, 1980), pp. 193-194.
25. L. M. Reder, "The Role of Elaborations in Memory for Prose," Cognitive
Psychology, 11 (1979):221-234, and R. J. Spiro, "Constructing a Theory of Recon-
structive Memory: The State of the Schema Approach," in R. C. Anderson, R. J.
Spiro, and W. E. Montague, eds., Schooling and the Acquisition of Knowledge
(Hillsdale, N.J.: Erlbaum, 1977).
26. Richard Nisbett and Lee Ross, Human Inference: Strategies and Shortcom-
ings of Sociul Judgment (Englewood Cliffs, N.J.: Prentice-Hall, 1980).
27. L. Ross, M. R. Lepper, and M. Hubbard, "Perseverance in Self Perception
and Social Perception: Biased Attributional Processes in the Debriefing Para-
digm," Journal of Personality and Social Psychology, 32 (1975):880-892. This and
other articles by Ross and colleagues are summarized by Nisbett and Ross, in
Human Inference, pp. 1768.
28. However, Nisbett and Ross sometimes overstate the results of these ex-
periments. For example, they write: "The evidence suggests that [Francis] Bacon
was correct in his assertion that 'the human understanding when it has once
adopted an opinion draws all things to support and agree with it' . . . Conflicting
evidence is treated as if it were supportive of beliefs, impressions formed on the
basis of early evidence survive exposure to inconsistent evidence presented later,
and beliefs survive the total discrediting of their evidence base" (p. 179). This
summary statement does not do justice to the facts of the experiment I have just
reviewed. After all, most of the initial impressions in this experiment were suc-
cessfully eliminated by process debriefings, and some of the effects of the initial
impressions were eliminated by outcome debriefings. Admittedly, there are many
other studies in their chapter on which Nisbett and Ross's summary statement
draws. But, all in all, the evidence they marshal does not support the very strong
statement quoted from Bacon. And their next sentence suggests that conflicting
evidence is always treated as if it were supportive of beliefs, that beliefs always
survive the discrediting of their evidence, and so on; whereas, in fact, these things
only sometimes happen.
29. Ibid., p. 181.
30. P. C. Wason and P. N. Johnson-Laird, Psychology of Reasoning: Structure
and Content (London: Batsford, 1965); M. Snyder and W. B. Swann, "Behavioral
Confirmation in Social Interaction: From Social Perception to Social Reality,"
Journal of Experimental SocMl Psychology, 14 (1978):148-162; and Snyder and N.
Cantor, "Testing Theories about Other People: Remembering All the History
that Fits," unpublished manuscript, University of Minnesota, 1979.
31. L. Ross, M. R. Lepper, F. Strack, and J. L. Steinmetz, "Social Explanation
and Social Expectation: The Effects of Real and Hypothetical Explanations upon
Subjective Likelihood," Journal of Personality and Social Psychology, 35
(1977):817-829.
32. R. Hastie and P. A. Kumar, "Person Memory: Personality Traits as Organ-
izing Principles in Memory for Behavior," Journal of Personality and Social Psy-
chology, 37 (1979):25-38.
33. Human Inference, p. 167.
34. The literature on artificial intelligence contains a good bit of discussion of
belief revision and dependency-directed backtracking. See Jon Doyle, "A Model
for Deliberation, Action, and Introspection," (Cambridge, Mass.: MIT Artificial
Intelligence Laboratory, 1980), no. TR-581; Doyle, "A Truth Maintenance Sys-
tem," Artificial Intelligence, 12 (1979):231-272; and R. M. Stallman and. G. J.
Sussman, "Forward Reasoning and Dependency-Directed Backtracking in a Sys-
tem for Computer-Aided Circuit Analysis," Artificial Intelligence, 9
(1977):135-196. I cannot pursue this technical literature in detail. But the com-
plications involved in a system of backtracking show that it may be excessive to
require ubiquitous backtracking. At a minimum this may prove excessive from a
resource-relative perspective.
35. Similar ideas are expressed by Gilbert Harman, in "Positive versus Nega-
tive Undermining in Belief Revision," Nous, 18 (1984):39-49. My own thoughts
on this were arrived at independently.
36. S. A. Madigan, "Intraserial Repetition and Coding Processes in Free Re-
call," Journal of Verbal Learning and Verbal Behavior, 8 (1969):828-835.
37. See Anderson, Cognitive Psychology and Its Implications, chap. 7.
38. W. Penfield, "The Interpretive Cortex," Science, 129 (1959):1719-25.
39. T. 0.Nelson, "Savings and Forgetting from Long-Term Memory," Journal
of Verbal Learning and Verbal Behavior, 10 (1971):568-576.
40. D. E. Rumelhart, P. Lindsay, and D. A. Norman, "A Process Model for
Long-Term Memory," in E. Tulving and W. Donaldson, eds., Organization of

Memory (New York: Academic Press, 1972), pp. 231-232.
41. Ibid., p. 232.
11. Constraints on Representation

1. Jerry Fodor, Modularity of Mind (Cambridge, Mass.: MIT Press, 1983), p.
120.
2. Noam Chomsky, Reflections on Language (New York: Pantheon, 1975),
pp. 10-11.
3. Cf. A. M. Collins and M. R. Quillian, "Retrieval Time from Semantic
Memory," Journal of Verbal Learning and Verbal Behauior, 8 (1969):MO-247, and
A. M. Collins and E. F. Loftus, "A Spreading-Activation Theory of Semantic Pro-
cessing," Psychological Review, 82 (1975):407428.
4. For studies of the formal properties of natural-language taxonomies, see
Fred Sommers, "The Ordinary Language Tree," Mind, 68 (1959):160-185 and
"Types and Ontology," Philosophical Review, 72 (1963):327-363. See also Frank
Keil, Semantic and Conceptual Development (Cambridge, Mass.: Harvard Univer-
sity Press, 1979).
5. D. Navon, "Forest before Trees: The Precedence of Global Features in
Visual Perception," Cognitive Psychology, 9 (1977):353-383; and P. H. Winston,
"Learning Structural Descriptions from Examples," in P. H. Winston, ed., The
Psychology of Computer Vision (New York: McGraw-Hill, 1975).
6. R. A. Bjork, "All-or-None Subprocesses in the Learning of Complex Se-
quences," Journal of Mathematical Psychology, 5 (1968):182-195; H. A. Simon
and K. Kotovsky, "Human Acquisition of Concepts for Sequential Patterns," Psy-
chological Review, 70 (1963):534-546; and J. G. Greeno and H. A. Simon, "Pro-
cesses for Sequence Production," Psychological Reuiew, 81 (1974):187-196.
7. W. A. Kintsch and T. A. van Dijk, "Toward a Model of Text Comprehen-
sion," Psychological Reuiew, 85 (1978):363-394.
8. G. A. Miller, E. H. Galanter, and K. H. Pribram, Plans and the Structure of
Behuoior (New York: Holt, Rinehart and Winston, 1960); Alvin Goldman, A The-
ory of Human Action (Englewood Cliffs, N.J.: Prentice-Hall, 1970; Princeton:
Princeton University Press, 1976).
9. G. W. Ernst and A. Newell, GPS: A Case Study in Generality and Problem
Solving (New York: Academic Press, 1969); A. Newel1 and H. A. Simon, Human
Problem Solving (Englewood Cliffs, N.J.: Prentice-Hall, 1972); H. A. Simon,
"Complexity and the Representation of Patterned Sequences of Symbols," Psy-
chological Review, 79 (1975):369-382.
10. A. Stevens and P. Coupe, "Distortions in Judged Spatial Relations," Cog-
nitive Psychology, 10 (1978):422-437.
11. G. H. Bower and A. L. Glass, "Structural Units and the Redintegrative
Power of Picture Fragments," Journal of Experimental Psychology: Human
Leczrning and Memory,2 (1976):456-466.
12. Stephen Palmer, "Hierarchical Structure in Perceptual Representation,"
Cognitive Psychology, 9 (1977):441-474.
13. Frank Restle, "Theory of Serial Pattern Learning: Structural Trees," Psy-
chological Review, 77 (1970):481-495.
14. F. Restle and E. Brown, "Organization of Serial Pattern Learning," in

G. Bower, ed., The Psychology of Learning and Motivation: Advances in Research
and Theory, vol. 4 (New York: Academic Press, 1970).
15. F. Lerdahl and R. Jackendoff, A Generatiue Theory of Music (Cambridge,
Mass.: MIT Press, 1983).Work on this theme has also been done by Carol Krum-
hand and colleagues. See C. L. Krumhansl and E. J. Kessler, "Tracing the Dy-
namic Changes in Perceived Tonal Organization in a Spatial Representation of
Musical Keys," Psychological Reuiew, 89 (1982):334-368, and Krumhansl and
R. N. Shepard, "Quantification of the Hierarchy of Tonal Functions within a
Diatonic Context," Journal of Experimental Psychology: Human Perception and
Performurace, 5 (1979):579-594.
16. Diana Deutsch and John Feroe, "The Internal Representation of Pitch Se-
quences in Tonal Music," Psychological Review, 88 (1981):503-522.
17. Heinrich Schenker, Free Composition, ed. and trans. Ernst Oster (New
York: Longman, 1979).
18. C. P. E. Bach, Essay on the True Art of Playing Keyboard Instruments, ed.
and trans. W. J. Mitchell (New York: Norton, 1949).
20. See Donald Hoffman, "The Interpretation of Visual Illusions," Scientific
American, 249 (1983):154-162, and Hoffman and W. A. Richards, "Parts of Rec-
ognition," Cognition, 18 (1984):65-96.
21. Shimon Ullman, The Interpretation of Visual Motion (Cambridge, Mass.:
MIT Press, 1979).
22. Ibid., chap. 4.
23. See Gunnar Johansson, "Visual Motion Perception," Scientific American,
232 (1975):76-88.
24. Reported in Hoffman, "The Interpretation of Visual Illusions," p. 157.
25. Noam Chomsky, Rules and Representations(New York: Columbia Univer-
sity Press, 1980), p. 252.
26. Marshall Edelson, "Language and Dreams: The Interpretation of Dreams
Revisited," Psychoanalytic Study of the ChiM, 27 (1972):203-282.
27. The Architecture of Cognition, chap. 2.
28. Amos Tversky, "Features of Similarity," Psychological Reuiew, 84
(1977):327-352.
29. George Lakoff and Mark Johnson, Metaphors We Live By (Chicago: Uni-
versity of Chicago Press, 1980).
30. G. Polya, How to Solue It (Princeton: Princeton University Press, 1957).
31. M. E. Atwood and P. G. Polson, "A Process Model for Water Jug Prob-
lems," Cognitive Psychology, 8 (1976):191-216.
32. For other attempts to study the role of analogy in theory construction, see
D. Gentner, "The Structure of Analogical Models in Science" (Cambridge, Mass.:
Bolt, Beranek and Newman, 1980),report no. 4451; L. Darden, "Artificial Intelli-
gence and Philosophy of Science: Reasoning by Analogy in Theory Construction,"
in Peter Asquith and Thomas Nickles, eds., PSA 1982, vol. 2 (East Lansing, Mich.:
Philosophy of Science Association, 1983); and Mary Hesse, Models and Analogies
in Science (Notre Dame, Ind.: University of Notre Dame Press, 1966).
33. Something like this may be behind the evidence about similarity presented
in A. Tversky and I. Gati, "Similarity, Separability, and the Triangle Inequality,"

Psychological Review, 89 (1982):123- 154.
34. "Parts of Recognition," p. 67.
35. Here too I have drawn from "Parts of Recognition," pp. 75-76.
36. K. A. Ericsson, W. G. Chase, and S. Fallon, "Acquisition of a Memory
Skill," Science, 208 (1980):1181-82.
37. "The Internal Representation of Pitch Sequences in Tonal Music," pp.
518-519.
38. Cf. D. N. Osherson, M. Stob, and S. Weinstein, "Learning Theory and Nat-
ural Languages," Cognition, 17 (1984):l-28; E. M. Gold, "Language Identifica-
tion in the Limit," Informution and Control, 10 (1967):447-474; K. Wexler and
P. Culicover, Fonnal Principles of Lcrnguage Acquisition (Cambridge,Mass.: MIT
Press, 1980); R. J. Matthews, "The Plausibility of Rationalism," The Journal of
Philosophy, 8 1 (1984):492-515.
39. W. V. Quine, "Natural Kinds," in Ontological Relativity and Other Essays
(New York: Columbia University Press, 1969), and The Roots of Reference (La
Salle, Ill.: Open Court, 1973), p. 23.
40. See Robert Merton, "Singletons and Multiples in Science," in The Sociol-
ogy of Science (Chicago: University of Chicago Press, 1973).
12. Internal Codes

1. John Locke, An Essay concerning Human Understanding (1690), bk. 2,
chap. 10, sect. 7.
2. David Hume, A Treatise of Human Nature (1739), bk. 1, pt. 1, sect. 1.
3. James Mill, Analysis of the Phenomena of the Human Mind (1829) p. 52.
4. Talk of a 'propositional code' implies that the mind has a form of represen-
tation that uses propositions. It suggests that propositions or propositionlike enti-
ties are vehicles of thought. This is quite different from the (milder) view that
what is represented in thought are propositions. This latter view is silent about
how the mind represents. It can be construed as merely holding that propositions
are an appropriate extrinsic index of thought, not that prop&itions are mental
instruments. (Cf. Robert Stalnaker, Inquiry, Cambridge, Mass.: MIT Press, 1984,
chap. 1.)
Unfortunately, it is far from clear just what vehicles, instruments, or data
structures are being postulated by a theorist who talks of a propositional code.
Must the vehicles be like sentences? In what respects like sentences? Despite this
lack of clarity, I shall continue to employ the terminology of propositional codes
with regularity. It must be insisted, however, that the charge sometimes leveled
against image theorists, that they are insuffciently explicit about the nature of the
code they postulate, can also be turned against (some) proposition theorists. They
too are insufficiently specific about the kind of representational vehicles to which
they are committed.
5. D.M. Armstrong, Perception and the Physical WorM (London: Routledge
and Kegan Paul, 1961), and George Pitcher, A Theory of Perception (Princeton:
Princeton University Press, 1971).
6. J. M. Shorter, "Imagination," Mind, 61 (1952):528-542, and Daniel Den-
nett, Content and Consciousness (New York: Humanities Press, 1969). Similar
views on imagery may be associated with Wittgenstein's Philosophical Inuestiga-

tions (New York: Macmillan, 1953).
7. John Anderson and Gordon Bower, Human Associative Memory (New
York: V. H. Winston, 1973); Herbert Simon, "What is Visual Imagery? An Infor-
mation Processing Interpretation," in L. W. Gregg, ed., Cognition in Learning
and Memory (New York: Wiley, 1972); H. H. Clark and W. G. Chase, "On the
Process of Comparing Sentences against Pictures," Cognitiue Psychology, 3
(1972):472-517; and M. Minsky and S. Papert, "Artificial Intelligence Progress
Report" (Cambridge, Mass.: MIT Artificial Intelligence Laboratory) memo no.
252.
8. See Zenon Pylyshyn, "What the Mind's Eye Tells the Mind's Brain: A Cri-
tique of Mental Imagery," Psychological Bulletin, 80 (1973):1-24.
9. J. S. Bruner, R. 0. Oliver, and P. M. Greenfield, Studies in Cognitiue
Growth (New York: Wiley, 1966).
10. R. N. Shepard and S. Chipman, "Second Order Isomorphisms of Internal
Representations: Shapes of States," Cognitiue Psychology, 1 (1970):1-17; Shepard
and J. Metzler, "Mental Rotation of Three-Dimensional Objects," Science, 171
(1971):701-703; Shepard and C. Feng, "A Chronometric Study of Mental Paper
Folding," Cognitiue Psychology, 3 (1972):228-243; Shepard, "Form, Formation,
and Transformation of Internal Representations," in R. L. Solso, ed., Infonnation
Processing and Cognition: The Loyola Symposium (Hillsdale, N.J.: Erlbaum,
1975).
11. "What the Mind's Eye Tells the Mind's Brain"; see also Zenon Pylyshyn,
"Imagery and Artificial Intelligence," in C. W. Savage, ed., Perception and Cog-
nition: Issues in the Foundations of Psychology (Minneapolis: University of Min-
nesota Press, 1978);"The Rate of 'Mental Rotation' of Images: A Test of a Holistic
Analogue Hypothesis," Memory and Cognition, 7 (1979):19-28; "The Imagery
Debate: Analogue Media versus Tacit Knowledge," Psychological Reuiew, 88
(1981):16-45; Computation and Cognition (Cambridge, Mass.: MIT Press, 1984),
chaps. 7-8.
12. Stephen Kosslyn, "Scanning Visual Images: Some Structural Implica-
tions," Perception and Psychophysics, 14 (1973):90-94; "Information Representa-
tion in Visual Images," Cognitive Psychology, 7 (1975):341-370; "Measuring the
Visual Angle of the Mind's Eye," Cognitiue Psychology, 10 (1978):356-389; Image
and Mind (Cambridge, Mass.: Harvard University Press, 1981); Kosslyn and J. R.
Pomerantz, "Imagery, Propositions, and the Form of Internal Representations,"
Cognitive Psychology, 9 (197752-76; Kosslyn, S. Pinker, G. E. Smith, and S. P.
Shwartz, "On the Demystification of Mental Imagery," Behavioral and Brain Sci-
ences, 2 (1979):535-581; reprinted in Ned Block, ed., Imagery (Cambridge, Mass.:
MIT Press, 1981).
13. Both the pictorialist formulation and the perception-similitude formula-
tion are presented by Ned Block, in his introduction to Block, ed., Imagery; also
see his "Mental Pictures and Cognitive Science," Philosophical Reuiew, 92
(1983):499-541.
14. See Kosslyn, Pinker, Smith, and Shwartz, "On the Demystification of
Mental Imagery," in Block, ed., Imagery, p. 132.
15. Image and Mind, p. 33.
16. Ibid., pp. 30, 31.
17. Ibid., p. 13.

18. Ibid., p. 18.
19. Ibid., p. 18; my emphasis.
20. Ibid.; my emphasis.
21. Stephen Kosslyn, "The Medium and the Message in Mental Imagery," in
Block, ed., Imagery, p. 207; my emphases.
22. R. N. Shepard, "The Mental Image," American Psychologist, 33
(1978):135.
23. R. N. Shepard and Lynn Cooper, Mental Images and Their T r a n s f m -
tions (Cambridge, Mass.: MIT Press, 1982), p. 14; emphasis added.
24. Zenon Pylyshyn, "The Imagery Debate: Analog Media versus Tacit
Knowledge," in Block, ed., Imagery, pp. 199-201.
25. Much of this evidence is culled from Block's introduction to Block, ed.,
Imagery.
26. C. W. Perky, "An Ekperimental Study of Imagination," American Journal
of Psychology, 21 (1910):422452. Also see S. J. Segal and V. Fusella, "Influence of
Imaged Pictures and Sounds on Detection of Visual and Auditory Signals," JOUF
nal of Experimental Psychology, 83 (1970):458-464.
27. For summaries, see Shepard, "The Mental Image," and R. A. Finke,
"Levels of Equivalence in Imagery and Perception," Psychological Review, 87
(1980):113- 139.
28. R. A. Finke and M. J. Schmidt, "Orientation-Specific Color Aftereffects
following Imagination," Joumal of Experimental Psychology: Human Perception
and P e r f m n c e , 3 (1977):599-606.
29. N. Pennington and S. M. Kosslyn, "Measuring the Acuity of the Mind's
Eye," (in preparation); reported by Block in his introduction to Block, ed., Imag-
ery.
30. Stephen Palmer, "Fundamental Aspects of Cognitive Representation," in
E. Rosch and B. Lloyd, eds., Cognition and Categorization (Hillsdale, N.J.: Erl-
baum, 1978).
31. I paraphrase my presentation from Kosslyn, "The Medium and the Mes-
sage in Mental Imagery."
32. Image and M i d , p. 277. (The original source misspells 'Simonides'; I have
corrected it here.) Kosslyn credits A. Paivio, Imagery and Verbal Processes (New
York: Holt, Rinehart and Winston, 1971).
33. Shepard and Cooper, Mental Tasks and Their Transfmtions, p. 36.
34. See Shepard and Cooper's introduction to Mental Zmages and Their
Transfinmutions, pp. 6-7. For further treatment of this theme, see Judith
Wechsler, ed., On Aesthetics in Science (Cambridge, Mass.: MIT Press, 1978).
35. See "The Imagery Debate: Analog Media versus Tacit Knowledge," and
Computation and Cognition.
36. Cf. Computation and Cognition, chap. 5.
37. Ralph Haber, "How We Remember What We See," in Richard Atkinson,
ed., Coirtemporay Psychology: Readingsfiom Scientijk American (San Francisco:
W. H. Freeman, 1971).
38. Personal communication from Mary Potter to Jerry Fodor, reported in
Fodor, Modularity of Mind (Cambridge, Mass.: MIT Press, 1983), pp. 62-63,
39. Richard Nisbett and Lee Ross, Human Inference: Strategies and Shortcorn-
ings of Social Judgment (Englewood Cliffs, N.J.: Prentice-Hall, 1980), chap. 3.
40. See Janellen Huttenlocher, "Constructing Spatial Images: A Strategy in
Reasoning," in P. N. Johnson-Laird and P. C. Wason, eds., Thinking (Cambridge:
Cambridge University Press, 1977).
41. Ibid., pp. 89-90.
42. H. H. Clark, "Linguistic Processes in Deductive Reasoning," in Johnson-
Laird and Wason, eds., Thinking.
43. David Hilbert, "On the Infinite," in Paul Benacerraf and Hilary Putnam,
eds., Philosophy of Mathematics (Englewood Cliffs, N.J.: Prentice-Hall, 1964).
44. Michael Resnik, "Mathematics as a Science of Patterns: Epistemology,"
Nous, 16 (1982):95-105. A similar view appears to have been maintained by the
Bourbaki mathematicians.
45. See Philip Kitcher, The Nature of Mathematical Knowledge (New York:
Oxford University Press, 1984), chap. 6.
46. Penelope Maddy, "Perception and Mathematical Intuition," The Philo-
sophical Reuiew, 89 (1980):163-196.
47. G. H. Hardy, A Mathmautician's Apology (Cambridge: Cambridge Univer-
sity Press, 1941), p. 24.
48. Human Inference, pp. 44-45.
49. W. C. Thompson, R. M. Reyes, and G. H. Bower, "Delayed Effects of
Availability on Judgment," unpublished manuscript, Stanford University, re-
ported in Nisbett and Ross, Human Inference, p. 52.
50. R. Hamill, T. D. Wilson, and R. E. Nisbett, "Ignoring Sample Bias: Infer-
ences about Collectivities from Atypical Cases," unpublished manuscript, Uni-
versity of Michigan, reported in Nisbett and Ross, Human Inference, p. 57.
51. R. E. Nisbett, E. Borgida, R. Crandall, and H. Reed, "Popular Induction:
Information is Not Always Informative," in J. S. Carroll and J. W. Payne, eds.,
Cognition and Social Behuuior (Hillsdale, N.J.: Erlbaum, 1976).
52. Human Inference, pp. 51-52.
53. Shelley Taylor and Suzanne Thompson, "Stalking the Elusive 'Vividness*
Effect," Psychological Review, 89 (1982):155-181.
13. Deductive Reasoning

1. Rudolf Carnap, "The Aim of Inductive Logic," in'^. Nagel, P. Suppes, and
A. Tarski, eds., Logic, Methodology and Philosophy of Science (Stanford: Stanford
University Press, 1962), p. 307.
2. Christopher Cherniak, "Computational Complexity and the Universal Ac-
ceptance of Logic," The Journal of Philosophy, 81 (1984):739-758.
3. I draw here on Cherniak's article, "Computational Complexity and the
Universal Acceptance of Logic." Among other papers on this topic that he cites
are two easily accessible articles: H. Lewis and C. Papadimitriou, "The Efficiency
of Algorithms," Scientific American, 238 (1978):96-109, and L. Stockmeyer and
A. Chandra, "Intrinsically Difficult Problems," Scientific Arn&n, 240
(1979):140-159.
4. For an explanation of the notion of an NP-complete problem, see Cher-
niak, "Computational Complexity and the Universal Acceptance of Logic," p.

744.
5. Lewis Carroll, "What Achilles Said to the Tortoise," Mind, 4
(1895):278-280.
6. For some treatments of this topic, however, see H. H. Clark, "Semantics
and Comprehension," in T. A. Sebeok, ed., Current Trends in Linguistics, vol. 12
(The Hague: Mouton, 1974),on the topic of negation; P. N. Johnson-Laird and J.
Tagart, "How Implication Is Understood," American Journal of Psychology, 2
(1969):367-373, on the topic of conditionals; and S. Fillenbaum, "Mind Your p's
and q's: The Role of Content and Context in Some Uses of AND, OR, and IF," in
Gordon Bower, ed., The Psychology of Learning and Motiuation, vol. 11 (New
York: Academic Press, 1977).
7. P. N. Johnson-Laird, "Models of Deduction," in R. J. Falmagne, ed., Rea-
soning: Representation and Process in Children and Adults (Hillsdale, N.J.: Erl-
baum, 1975); D. N. Osherson, "Logic and Models of Logical Thinking," in ibid,;
M. D. S. Braine, "On the Relation between the Natural Logic of Reasoning and
Standard Logic," Psychological Review, 85 (1978):l-21; and L. Rips, "Cognitive
Processes in Reasoning,'' Psychological Reuiew, 90 (1983):38-71.
8. A. Newel1 and H. A. Simon, Human Problem Solving (Englewood Cliffs,
N.J.: Prentice-Hall, 1972).
9. D. Hewitt, "Description and Theoretical Analysis (Using Schemata) of
PLANNER" (Cambridge, Mass.: MIT Artificial Intelligence Laboratory, 1972),
no. AI-TR-258; D. V. McDermott and G. J. Sussman, "The CONNIVER Refer-
ence Manual" (Cambridge, Mass.: MIT Artificial Intelligence Laboratory), memo
no. 259.
10. W. V. Quine, Methods of Logic (New York: Holt, Rinehart and Winston,
1950, 1959).
11. Philip Johnson-Laird, Mental Models (Cambridge, Mass.: Harvard Univer-
sity Press, 1983).
12. Ibid., pp. 53-54.
13. Here my exposition of Johnson-Laird partly follows Lance Rips's "Mental
Muddles," in Myles Brand and Robert Harnish, eds., The Representation of
Knowledge and Belief (forthcoming).
14. Mental Models, p. 104.
15. Ibid.; also reported in P. N. Johnson-Laird and M. J. Steedman, "The Psy-
chology of Syllogisms," Cognitive Psychology, 10 (1978):64-99.
16. L. J. Rips and S. L. Marcus, "Supposition and the Analysis of Conditional
Sentences," in M. A. Just and P. A. Carpenter, eds., Cognitive Processes in Com-
prehension (Hillsdale, N.J.: Erlbaum, 197'7).
17. J. E. Taplin and H. Staudenmayer, "Interpretation of Abstract Conditional
Sentences in Deductive Reasoning," Journal of Verbal Learning and Verbal Be-
ha&, 12 (1973):530-542.
18. R. S. Woodworth and S. B. Sells, "An Atmosphere Effect in Formal Syllo-
gistic Reasoning," Journal of Experimntal Psychology, 18 (1935):451460.
19. See P. C. Wason and P. N. Johnson-Laird, Psychology of Reasoning:
Structure and Content (London: Batsford, 1965; Cambridge, Mass.: Harvard Uni-
versity Press, 1972), chaps. 13 and 14.
20. The opus classicus here is W . V. Quine, "Two Dogmas of Empiricism," in

From a Logical Point of View (Cambridge, Mass.:Harvard University Press, 1953).
21. David Hume, A Treatise of Human Nature (1739),bk. 1, pt. 4, sect. 1.
22. Cf. Renh Descartes, ReguZue, especially rules 7 and 11.
23. On similar issues, see Philip Kitcher, The Nature of Mathematical Knowl-
edge (New York: Oxford University Press, 1984), chap. 2.
24. This possibility was raised by Holly Smith.
25. See Mental Models, p. 109.
14. Probability Judgments

1. These quotations are taken from Amos Tversky and Daniel Kahneman,
"Extensional versus Intuitive Reasoning: The Conjunction Fallacy in Probability
Judgment," Psychological Review, 90 (1983):293-3 15.
2. Daniel Kahneman and Amos Tversky, "Subjective Probability: A Judg-
ment of Representativeness," in D. Kahneman, P. Slovic, and A. Tversky, e&.,
Judgment under Uncertainty: Heuristics and Biases (Cambridge: Cambridge Uni-
versity Press, 1982), p. 46. Several other papers by Tversky and Kahneman, also
reprinted as chapters of this volume, strike similar themes, including "Judgment
under Uncertainty: Heuristics and Biases," "Belief in the Law of Small Numbers,"
and "On the Psychology of Prediction".
3. Richard Nisbett and Lee Ross, Human Inference: Strategies and Shortcom-
ings of Social Judgmat (Englewood Cliffs,N.J.: Prentice-Hall, 1980), p. 3.
4. Ibid., p. 6.
5. Ibid.
6. R. E. Nisbett and E. Borgida, "Attribution and the Psychology of Predic-
tion," Journal of Personality and Social Psychology, 32 (1975):935.
7. P. Slovic, B. Fischhoff, and S. Lichtenstein, "Cognitive Processes and Soci-
etal Risk Taking," in J. S. Carroll and J. W. Payne, eds., Cognition and Social Be-
hauior (Hillsdale, N.J.: Erlbaum, 1976), p. 169.
8. L. Jonathan Cohen, "Can Human Irrationality Be Experimentally Demon-
strated?" The Behavioral and Brain Sciences, 4 (1981):317-331. Commentaries on
this paper, including ones by Fischhoff, Kahneman, Nisbett, and Tversky, appear
in the same journal issue, pp. 331-359.
9. These are summarized in "Judgment under Uncertainty: Heuristics and
Biases," and in the other articles cited in note 2, which are reprinted in Kahne-
man, Slovic, and Tversky, eds., Judgment under Uncertainty.
10. Amos Tversky and Daniel Kahneman, "Judgments of and by Representa-
tiveness," in Kahneman, Slovic, and Tversky, eds., Judgment under Uncertainty.
11. Ibid., pp. 88-89.
12. "Extensional versus Intuitive Reasoning," p. 304.
13. However, the phrase might.be applied to states that a cognizer could be
in, not just to states he is in. That is, the phrase may refer to a state it would be
rational, or justified, for him to be in, even if he is not actually in it.
14. A survey of these issues is found in Ellery Eells, Rational Decision and
Causality (Cambridge: Cambridge University Press, 1982).
15. "Extensional versus Intuitive Reasoning," p. 311.
16. R. E. Nisbett, D. H. Krantz, C. Jepson, and Z. Kunda, "The Use of Statisti-

cal Heuristics in Everyday Inductive Reasoning," Psychological Review, 90
(1983):339-363.
17. B. Fischhoff, P. Slovic, and S. Lichtenstein, "Subjective Sensitivity Analy-
sis," Organizational Behavior and Hunaun P ~ m n c e23 , (1979):339-359.
18. See p. 310.
19. Daniel Kahneman and Amos Tversky, "On the Study of Statistical Intui-
tions," in Kahneman, Slovic, and Tversky, eds., Judgment under Uncertainty, p.
494.
24). L. J. Rips, E. J. Shoben, and E. E. Smith, "Semantic Distance and the
Verification of Semantic Relations," Journal of Verbal Learning and Verbal Be-
havior, 12 (1973):l-20; and E. Rosch, "On the Internal Structure of Perceptual
and Semantic Categories," in T. E. Moore, ed., Cognitive Development and the
Acquisition of Lrrnguage (New York: Academic Press, 1973).
21. See Edward Smith and Douglas Medin, Concepts and Categories (Cam-
bridge, Mass.: Harvard University Press, 1981).
15. Acceptance and Uncertainty

1. Varying terminologies make it unclear to whom, exactly, this position can
be attributed. But it seems to be attributable at least to Isaac Levi, Gambling with
Truth (New York: Knopf, 1967), and to Keith Lehrer, Knowledge (Oxford: Oxford
2. In The Enterprise of Knowledge (Cambridge, Mass.: MIT Press, 1980),
Isaac Levi proposes a theory intended to rationalize the revision of subjective
probabilities equal to 1.For criticisms of this theory, however, see reviews of the
book by Patrick Maher, Philosophy of Science, 51 (1984):690-692, and by Mark
Kaplan, The Philosophical Reoiew, 92 (1983):310-316.
3. Gilbert Harman, Thought (Princeton: Princeton University Press, 1973),p.
118; also Mark Kaplan, "A Bayesian Theory of Rational Acceptance," TheJournal
of Philosophy, 78 (1981):305-330.
4. For a survey, see Ellery Eells, Rational Decision and Causality (Cam-
bridge: Cambridge University Press, 1982), chaps. 1, 2.
5. Amos Tversky, "Intransitivity of Preferences," Psychological Reuiew, 76
(1969):31-48.
6. K. R. MacCrimmon, "Descriptive and Normative Implications of the De-
cision-Theory Postulates," in K. Borch and J. Mossin, eds., Risk and Uncertainty
(New York: St. Martin's, 1968).
7. Cf. Amos Tversky, "A Critique of Expected Utility Theory," Erkenntnis, 9
(1975):163-173, and Daniel Kahneman and Amos Tversky, "Prospect Theory: An
Analysis of Decision under Risk," Econometricq 47 (1979):263-291.
8. I owe this point to John Pollock.
9. See David Rumelhart and Donald Norman's introduction to Geofiey
Hinton an$. James Anderson, eds., Parallel Models of Associative Memory (Hills-
dale, N.J.: Erlbaum, 1981).
10. See J. A. Feldman and D. H. Ballard, "Connectionist Models and Their
Properties," Cognitive Science, 6 (1982):W.
11. Ibid., and Lokendra Shastri and J. A. Feldman, "Semantic Networks and
Neural Nets" (Rochester: University of Rochester, Computer Science Depart-

ment, 1984), technical report no. 131. Also see Feldrnan, "A Connectionist Model
of Visual Memory," in Hinton and Anderson, eds., Parallel Models of Associative
Memory.
12. Actually the fonnulcr seems to imply that a unit sets itself to zero if it
knows of a higher or equal input. But I assume that Feldman and Ballard mean
just a higher input, which is what they say (p. 228).
13. See C. S. Peirce, "The Fixation of Belief' and "How To Make Our Ideas
Clear," in Justus Buchler, ed., The Philosophical Writings of Peirce (New York:
Dover, 1955).
14. There is a nonstandard model of probability, proposed by A. P. Dempster
and G. A. Shafer, in which this requirement is relaxed. See Dempster, "A General-
ization of Bayesian Inference," Jounzal of the Royal Statistical Society, ser. B, 30
(1968):U15-247, and Shafer, A Muthematical Theory of Eddeme (Princeton:
Princeton University Press, 1976). However, I focus on the standard model.
15. Notice that a theory of propositional uncertainty could include not only
propositional contents with explicit probability values, but also modal proposi-
tions like 'It might rain today'.
16. Cf. Lehrer, Knowledge, chap. 8.
17. The problem of little or no information is mentioned in Shastri and Feld-
man, "Semantic Networks and Neural Nets," pp. 2628.
18. Use of scenarios in bolstering the strength of hypotheses has been criti-
cized by Tversky and Kahneman, on the grounds that they can lead to conjunc-
tion errors. See their article "Extensional versus Intuitive Reasoning,"
Psychological Reuiew, 90 (1983):293315. For example, groups of subjects (in
1982) estimated the probability of "a massive flood somewhere in North America
in 1983, in which more than 1,000 people drown" and the probability of "an
earthquake in California sometime in 1983, causing a flood in which more than
1,000 people drown." The estimates of the conjunction-earthquake and flood-
were significantly higher than estimates of the flood. Presumably, a reminder that
a devastating flood could be caused by an earthquake made the conjunction--a
'coherent' s6quenc-seem more probable. similarly, Tversky and- Kahneman
note, a plausible story of how a victim might have been killed by someone other
than the defendant may convince a jury of the existence of reasonable doubt. An
attorney can fill in guesses regarding unknown facts, such as motive or mode of
operation, thereby strengthening a case through improving its coherence, and a
political analyst can improve scenariosby adding plausible causes and representa-
tive consequences. All such uses of coherence-promoting scenarios are regarded
by Tversky and Kahneman with suspicion.
These examples may illustrate how certain uses of coherence can lead cognizers
astray. But they hardly foreclose the prospect that some versions of the coherence
processes in question could help a setof processes achieve a reasonably high truth
ratio, high enough to qualify for admission into a right system of J-rules. Indeed, it
is not even clear that these examples are (all) misuses of coherence. Maybe sub-
jects' estimates of the chances of flood alone were not very accurate. Being re-
minded that a flood can be caused by an earthquake could improve the accuracy
of one's estimate of flood chances-and of earthquake-plus-flood chances.
19. Speaking of 'nature's way', it is worth noting the existence of WTA mecha-
nisms elsewhere in nature. Consider, for example, the WTA mechanisms of mat-
ing found in certain species. Elk and bighorn sheep have battles among their
males to determine dominance within the herd. The females of the herd all mate
exclusively with the single victorious male. This mechanism has some obvious ge-
netic advantages. It allows the genes of only the strongest or most battleworthy
male to be transmitted to subsequent generations. It also has some apparent dis-
advantages: it does not make use of the entire gene pool of the male population.
The viability of such species, however, suggests that this is a tenable mating
mechanism. Analogous advantages of WTA mechanisms in cognition may also be
identified.
16. Belief Updating

1. One variant of conditionalization is that presented by Richard Jeffrey in
The Logic of Decision, 2nd ed. (Chicago: University of Chicago Press, 1983),chap.
11.
2. Hillel Einhorn and Robin Hogarth, "A Contrast Model for Updating Be-
liefs," University of Chicago, Center for Decision Research (1984). I am grateful
to Professors Einhorn and Hogarth for allowing me to present material from this
unpublished paper.
3. J. D. Shanteau, "An Additive Mode1 for Sequential Decision Making,"
Journal of Experimental Psychology, 85 (1970):181-191.
4. N. H. Anderson, Foundations of Information Integration Theory (New
York: Academic Press, 1981).
5. For example, see Amartya Sen, Collective Choice and Social Welfare (San
Francisco: Holden-Day, 1970), chap. 10, and C. Plott, "Axiomatic Social Choice
Theory: An Overview and Interpretation," American Journal of Political Science,
20 (1976):511-596.
6. Hans Reichenbach, Theory of Probability (Berkeley: University of Califor-
nia Press, 1949), pp. 470-482.
7. See Karl Popper, The Logic of Scientific Discovery (London: Hutchinson,
1959), p. 267.
8. Imre Lakatos, "Falsification and the Methodology of Scientific Research
Programmes," in Imre Lakatos and Alan Musgrave, eds., Criticism and the
Growth of Knowledge (Cambridge: Cambridge University Press, 1970).
9. Larry Laudan, Progress and Its Problems (Berkeley: University of Califor-
nia Press, 1977), pp. 196ff.
10. Robin Hogarth, "Beyond Discrete Biases: Functional and Dysfunctional
Aspects of Judgmental Heuristics," Psychological Bulletin, 90 (1981):197-217.
17. Production Systems and Second-Order Processes

1. E. L. Post, "Formal Reductions of the General Combinatorial Decision
Problem," American Journal of Mathematics, 65 (1943):197-268.
2. A. .Newel1 and H. A. Simon, Human Problem Soloing (Englewood Cliffs,
N.J.: Prentice-Hall, 1972); also see Newell, "Production Systems: Models of Con-
trol Structures," in W. G. Chase, ed., Visual Information Processing (New York:
Academic Press, 1973).
Notes to Pages 360-3 72 41 9

4. John Anderson, Language, Memory, and Thought (Hillsdale, N.J.: Erl-
baum, 1976).
5. Donald Norman, "Categorization of Action Slips," Psychological Review,
88 (1981):1-15.
6. See G. A. Miller, E. Galanter, and K. H. Pribram, Plans and the Structure
of Behavior (New York: Holt, Rinehart and Winston, 1960), and Newel1 and
Simon, Human Problem Soluing.
7. B. Hayes-Roth and F. Hayes-Roth, "A Cognitive Model of Planning,"
Cognitiue Science, 3 (1979):275-310.
8. The following two productions and macroproduction are quoted from Ar-
chitecture of Cognition, pp. 235-236.
9. Anderson also discusses what he calls "action discrimination," as opposed
to "condition discrimination," but I will ignore that.
10. Cf. P. H. Winston, "Learning Structural Descriptions from Examples"
(Cambridge, Mass.: MIT Artificial Intelligence Laboratory, 1970),no. AI-TR-231,
and "Learning to Identify Toy Block Structures," in R. L. Solso, ed., Contempo-
r a y Issues in Cognitiue Psychology: The Loyola Symposium (Washington, D.C.:
Winston, 1973).
Illustration Credits
Figure 9.1: redrawn from 0.G. Selfridge, "Pattern Recognition and Modem
Computers," in Proceedings of the Western Joint Computer Conference (New
York: Institute of Electrical and Electronics Engineers, 1955), p. 92. Copyright
o by the American Federation of Information Processing Societies.
Figure 9.2: redrawn from Irvin Rock, % Logic of Perception (Cambridge,
Mass.: MIT Press, 1983), p. 309.
Figure 9.3: redrawn from Stephen Palmer, "Symmetry, Transformation, and
the Structure of Perceptual Systems," in J. Beck, ed., Organization and
Representation in Perception (Hillsdale, N.J.: Erlbaum, 1982),p. 122.
Figure 10.1: redrawn from L.Ross, M. Lepper, and M. Hubbard,
"Perseverance in Self Perception and Social Perception: Biased Attributional
Processes in the Debriefing Paradigm," Journal of Personality and Social
Psychology, 32 (1975):887. Copyright Q 1975 by the American Psychological
Association. Adapted by permission of the publisher and author.
Figure 10.2: redrawn from Stephen Madigan, "Intraserial Repetition and
Coding Processes in Free Recall," Journal of Verbal Learning and Verbal
B e h a w , 8 (1969):829.
Figure 11.1: redrawn from John Anderson, Cognitive Psychology and Its
Implications (San Francisco: W . H. Freeman, copyright o 1980),p. 53.
Figure 11.2: redrawn from G. H. Bower and A. L. Glass, "Structural Units
and the Redintegrative Power of Picture Fragments," Journal of Experimental
Psychology: Human Learning and Memory,2 (1976):459,460. Copyright o 1976 .
by the American Psychological Association. Adapted by permission of the
publisher and author.
Figure 11.3: redrawn from Stephen Palmer, "HierarchicaI Structure in
Perceptual Representation," Cognitive Psychology, 9 (1977):452.
Figures 11.4 and 11.5: redrawn from Diana Deutsch and John Feroe, "The
Internal Representation of Pitch Sequences in Tonal Music," Psychological
Reoiew, 88 (1981):504, 505. Copyright @ 1981 by the American Psychological
Association. Adapted by permission of the publisher and author.
Figures 11.6 and 11.7: redrawn from Donald Hoffman, "The Interpretation
of Visual Motion," Scientific American, 249 (1983):157, 159. Copyright o 1983
by Scientific American, Inc. All rights reserved.
Figure 12.1: redrawn from Roger Shepard and Lynn Cooper, Mental Images
and Their Transformations (Cambridge, Mass.: MIT Press, 1982), p. 21.
422 Illustration Credits
Figure 13.1: redrawn from P. C. Wason and P. N. Johnson-Laird, Psychology

of Reasoning: Structure and Content (Cambridge, Mass.: Harvard University
Press, 1972), p. 173.
Figures 15.1-15.5: redrawn from J. A. Feldman and D. H. Ballard,
"Connectionist Models and Their Properties," Cognitive Science, 6 (1982):213,
219, 221, 229.
Figures 16.1-16.6: redrawn from Hillel Einhorn and Robin Hogarth's
unpublished paper "A Contrast Model for Updating Beliefs," with the authors'
permission.
Figures 17.1-17.3: redrawn from John Anderson, The Architecture of
Cognition (Cambridge, Mass.: Harvard University Press, 1983), pp. 10, 19, 132.
Author Index
Alston, William, 388nn30,5, 390n30, Campbell, Donald, 2, 149

404& Cantor, Georg, 317
Anderson, Alan, 391n5 Cantor, M., 217
Anderson, John, 201,205, 212,235, Carnap, Rudolf, 6, 87, 89, 204, 279,
360-366,368-373,375,376,403n5, 389nn15,17
404n23.407n37 Carneades, 28
Anderson, N. H., 352 Carroll, Lewis, 285, 394n21
Arcesilaus, 28 Chandra, A,, 413n3
Armstrong, D. M., 43,254,3961130 Chase, W. G., 244,254, 398n19
Atwood, M. E., 240 Chastain, Charles, 242
Austin, John, 65 Cherniak, Christopher, 175, 282, 391n8,
405nn8,11, 413nn3,4
Bach, C. P. E., 235 Chipman, S., 258, 411n10
Bacon, Francis, 4 Chisholm, Roderick, l00,389n21,405n27
Ballard, Dana, 329-334 Chomsky, Noam, 228,235,238-239
Barclay, J. R., 209 Christiaansen, R. E., 209
Baumal, R., 190, 191 Church, Alonzo, 281,300
Bayes, Thomas, 344,345 Churchland, Patricia, 401n4
Bayle, Pierre, 31 Churchland. Paul, 164-165
Belnap, Nuel, 127,39ln5,398nll Clark, H. H., 254, 271, 414n6
Berkeley, George, 34, 35 Cofer, C. N., 398n18
Bird, C., 210 Cohen, L. Jonathan, 305,307
Bjork, R. A., 408116 Cohen, Stewart, 3961131
Block, Ned, 401n6,411n13,412n25 Coleman, Jules, 389n9
Bobrow, S., 212 Collins, A. M., 408n3
Bohr, Niels, 241 Cook, T. D., 399n14
BonJour, Laurence, 109-110, 111- Cooper, Lynn, 258,412nn33,34
112 Coupe, P., 230
Borgida, E., 274,305 Craik, F. I. M., 212
Bower, Gordon, 212,230-232,273 Crandall, R., 274
Boyd, Richard, 157 Crick, Francis, 265
Braine, M. D. S., 287 Crowder, Robert, 4051-12
Bransford, J. D., 2OQ Culicover, Peter, 4101138
Bromberger, Sylvain, 125, 126 Cummins, Robert, 172, 405n7
Brown, E., 233 Currie, Gregory, 400n16
Bruner, Jerome, 255
Burge, Tyler, 168, 4021-128 Darden, Lindley, 4091132
Bums, H. J., 406n21 Darwin, Charles, 241
424 Author Index
Davidson, Donald, 173-177 Gettier, Edmund, 46, 47, 54,3861121,

DeGroot, A., 3981119 3881126
DeGroot, Morris, 3961134 Gibson, James, 191
Dempster, A. P., 417n14 Gick, Mary, 133
Dennett, Daniel, 167, 174, 201-202, 254, Glass, A. L., 230-232
405n5 Gold, E. M., 4101138
Descartes, Renb, 7, 31,33, 79, 176, Goldman, Alvin, 46, 47, 49, 384113,
195-196,301,3&?n5,38On17 385n14,387n13,396n30,405nn1,4,
Deutsch, Diana, 234-235, 238, 245 408n8
Devitt, Michael, 400n16 Goodman, Nelson, 66, 152, 153-154,
Dewey, John, 8, 125 385n 10
Dooling, D. J., 209 Greenfield, P. M., 255
Doyle, Jon, 4071134 Greeno, James, 408n6
Dretske, Fred, 46, 56, 174, 383n4
Dummett, Michael, 142-143, 145, 147, Haber, Ralph, 267
150,399nn4,9,400n25 Hamill, R., 274
Hardy, G. H., 272
Edelson, Marshall, 239 Harman, Gilbert, 83,89,206-207,325,
Eells, Ellery, 415n14, 416n4 383n4,38Sn15,390n28,407n35
Eggenberger, Peter, 4001116 Harnish, Robert, 4021126
Einhorn, Hillel, 344452,354 Hart, H. L. A., 65
Einstein, Albert, 77, 265 Hastie, R., 218
Ericsson, K. A., 244 Hayes-Roth, B., 368
Emst, G. W., 408n9 Hayes-Roth, F., 368
Escher, M. C., 315 Hempel, Carl, 101,204,391n9
Heraclitus, 31
Fallon, S., 244 Hesse, Mary, 409n32
Feldman, Jerome, 329334,336,417n17 Hewitt, D., 288
Feldrnan, Richard, 3871121 Hilbert, David, 271
Feng, C., 411n10 Hilpinen, Risto, 101
Feroe, John, 234-235,238,245 Hintikka, Jaakko, 101, 125,398n9
Field, Hartry, 174,388n1,401n2,402n29 H o h a n , Donald, 235-236, 237, 243, 246
Fienberg, Stephen, 396n34 Hogarth, Robin, 344352,354,358
Fillenbaum, S., 414n6 Hollis, Martin, 389n19
Finke, R. A., 259 Holyoak, Keith, 133
Firth, Roderick, 117-1 18,390n31, 392n13, Horgan, Terence, 169
405n25 Hubbard, H., 214
Fischhoff, Baruch, 114,305,319 Hubel, D. H., 403113
Fiske, D., 399x114 Hull, David, 401n9
Fleming, Alexander, U)3, 204 Hume, David, 6,32, 3536, 253,301,378,
Flinchbaugh, Bruce, 237 393n21
Fodor, Jerry, 186, 170-171, 174, 187-188, Huttenlocher, Janellen, 270
192,197,228,245,401n6,404~14,20 Hyde, T. S., 210,212
Foot, Philippa, 389n23
Franks, J. J., 209 Isard, S., 186
Frege, Gottlob, 6, 143, 314
Freud, Sigmund, 239 Jackendoff, Ray, 233
Fusella, V.,412n26 Jeffrey, Richard, 345,3921115. 418nl
Jenkins, J. J., 210, 212
Galanter, E. H., 408n8, 419n6 Jepson, C., 319
Gati, I., 4101133 Johansson, Gunnar, 237
Gentner, D., 4091132 Johnson, Mark, 239
Author Index 425
Johnson-Laird, Philip, 217, 287,290-293. Marcus, S. L., 296

298-!299,304,414n6 Marr, David, 403nn1,2
Marsetta, M., 403n4
Kahneman, Daniel, 305311,317422, Matthews, R. J., 410n38
327,39ln12,398n16,417n18 Maturana, H. R., 4031-13
Kant, Immanuel, 6, 7,31, 34,35, 152, 153, McCulloch, W. S., 403113
227,228,268 McDermott, D. V., 288
Kaplan, Mark, 416nn2,3 McGinn, Colin, 47-48, 56, 386x16, 4001116,
Keil, Frank, 408114 403n42
Kessler, E. J., 409n15 Medin, Douglas, 385n19.4161121
Kinney, G. C., 403n4 Merton, Robert, 248
Kintsch, Walter, 133, 408x17 Metzler, J., 411n10
Kitcher, Philip, 271-272, 415n23 Mill, James, 253
Klein, Peter, 385x13 Mill, John Stuart, 34,35, 295
Kornblith, Hilary, 394n22, 3991124 Miller, D. G., 4061121
Kosslyn, Stephen, 255, 257-2434 Miller, George, 186,408118,419n6
Kotovsky, K., 408n6 Minsky, M., 2S4
Krantz, David, 319 Montague. Phillip, 3891123
Kripke, Saul, 107, 168,4021119 Moore, G. E., 3891123
Krumhansl, Carol, 409x115
Kuhn, Thomas, 8,66, 125, 188 Nagel, Ernest, 165-166
Kumar, P. A., 218 Navon, D., 408n5
Kunda, Z., 319 Nelson, T. O., 223
Neurath, Otto, 3
Lakatos, Imre, 1%. 164,358,398n13 Newell, Allen, 288,359, 408x19, 4191-16
Lakoff, George, 239 Newton-Smith, W. H., 157, 4011130
Laudan, Larry, 125,129-130,358,3981113, Nickles, Thomas, 3981113
400n27.401n30 Niiniluoto, Ilkka, 401n30
Lehrer, Keith, 79,101,388n27,392n14, Nilsson, Nils, 398n10
396n31,416nl, 417nl6 Nisbett, Richard, 71,214-219.268-269.
Lepper, M. R, 214,217 273-277,305,319,391n12,403n41
Lerdahl, F., 233 Norman, Donald, 223424,365, 416n9
Lettvin, J. Y., 403n3 Nozick, Robert, 45-46, 48, 56
Levi, Isaac, l01,393n19,397nl,416nn1,2
Levins, R., 149 Oakley, I. T., 388nl
Lewis, H., 413n3 Oliver, R. O., 255
Lewis, C. I., 100 Osherson, D. N., 287,4101138
Lichtenstein, Sarah, 114,305,319-320 * Owens, Joseph, 401116
Lindsay, Peter, 223-224
Loar, Brian, 148,174,402n15 Paivio, A., 412n32
Locke, John, 6,253,378 Palmer, Stephen, 232-233,246,260
Lockhart, R. L., 212 Papadimitriou, C., 413n3
Loftus, Elizabeth, 210-211,408n3 Papert, S., 254
Lukes, Steven, 38&119 Pappas, George, 384111
Pastin, Mark, 390x131,4051125
MacCrirnmon, K. R, 327 Paxson, Thomas, 3881127
Maddy, Penelope, 272 Peirce, C. S., 8, 145,334,336, 357
Madigan, Stephen, 221 Penfield, W., 223
Maher, Patrick, 416n2 Pennington, N., 259
Malthus, Thomas. 241 Perky, C. W., 259
Mandler, George, 190, 191 Phillips, Lawrence, 114
Marcus, Ruth, 402n15 Pietarinen, J., 101
426 Author Index
Pinker, S., 411nn12,14 Savage, Leonard, 327

Pitcher, George, 254 SchafFner, Kenneth, 165
Pitts, W. H., 403n3 Schenker, Heinrich, 235
Plantinga, Alvin, 155 Schiffer, Stephen, l74,388n1,401n6
Plato, 8 , 3 1 Schmidt, J., 259
Plott, C., 418n5 Schrnitt, Frederick, 386n4, 3871120
Pollock, John, 3891116,39On24,393n17, Segal, S. J., 412n2.6
396x13 1,405nn26,28,4 16n8 Sellars, Wilfrid, 388x14
Polson, P. G., 240 Sells, S. B., 297
Polya, G., 132, 240 Sen, Amartya, 418n5
Pornerantz, J. R., 411n12 Sextus Empiricus, 28-29, 31, 32
Popkin, Richard, 384n1, 385114 Shafer, Glenn, 417n14
Popper, Karl, 22, 125,357,377,4001130 Shanteau, J. D., 350,351
Posner, Michael, 132 Shastri, Lokendra, 329,417n17
Post, E. L., 359 Shepard, Roger, 210,255,258,259,265,
Potter, Mary, 267 4091115
Pribram, Karl, 408n8,419n6 Shoben, E. J., 321
Ptolemy, 130 Shoemaker, Sydney, 38
Putnam, Hilary, 35,107-108,109, 110, Shope, Robert, 387n11
116-117, 142, 144, 146-147, 148, Shorter, J. M., 254
151-152, 153-154, 155-156, 157, 158, Showman, D. J., 4031-14
164, 170,402n19 Shwartz, S. P., 411nn12,14
Pylyshyn, Zenon, 187, 255,258-259, Simon, Herbert, 254,288,359,398n19,
265-266,268,402n30,411n8 408nn6,9, 419n6
Pyrrho, 28 Simonides, 264
Sinclair, Upton, 273
Quillian, M. R., 408x13 Skyrms, Brian, 46,395n23
Quine, W. V., 2, 7,38, 57, 163, 173, 197, Slovic, P., 305, 319, 391n12
206,247,288,299,300,314,378,383n5, Smith, Edward, 321,385n19,416n21
402n 19 Smith, G. E., 411nn12.14
Smith, Holly, 384x14, 405n13, 415n24
Ramsey, Frank, 44 Snyder, M., 217
Rawls, John, 64-65, 66, 67 Soames, Scott, 3991-15
Reder, L. M., 213 Sommers, Fred, 408n4
Reed, H., 274 Sosa, Ernest, 388114, 404n24
Reichenbach, Hans, 6, 357 Spinoza, Benedict, 79
Reicher, G., 190 Spiro, R. J., 213
Rescher, Nicholas, 384n1 Stace, Walter, 390n23
Resnik, Michael, 271 Stallman, R. M., 407x134
Restle, Frank, 233 Stalnaker, Robert, 168,40ln2,410n4
Reyes, R. M., 273 Staudenmayer, H., 297
Rich, Elaine, 398n10 Steedman, M. J., 414n15
Richards, W. A,, 243,409nU) Steel, Thomas, 127,398n11
Rips, Lance, 242,287-289,292-299,303, Steinmetz, J. L., 217
321 Stevens, A., 230
Rock, Irvin, 185, 190, 192,404nnl1,12 Stich, Stephen, 71, 167, 169, 171, 401n6
Rorty, Richard, 8, 18, 66, 75, 79 Stob, M., 410n38
Rosch, Elemor, 321 Stockmeyer, L., 413n3
Ross, Lee, 214-219,268-269,273-277, Stowe, Harriet, 273
305,344,391n12 Strack, F., 217
Rumelhart, David, 223-224,416n9 Strawson, Peter, 143
Russell, Bertrand, 32,93,314 Stroud, Barry, 38
Author Index 427
Sulin, R. A., 209 Van Cleve, James, 394n21

Swman, G. J., 288,407n34 Van Dijk, T. A., 133, 408n7
Swain, Marshall, 386n2 Van Fraassen, Bas, 125, 397x134, 4001124
Swann, W. B., 217
Tagart, J., 414n6 Warren, R. M., 186

Taplin, J. E., 297 Wason, P. C., 217, 297
Tarski, Alfred, 144 Watson, James, 265
Taylor, Shelley, 275-276 Wechsler, Judith, 412n34
Tesla, Nicola, 285 Weierstrass, Karl, 269
Thagard, Paul, 4031141 Weinstein, Scott, 410n38
Thompson, Suzanne, 275-276 Weisel. T. N., 403n3
Thompson, W. C., 273 Wheeler, D. D., 190
Tichy, Pavel, 401n30 Wilson, T. D., 274
Tulving, Endel, 190, 191 Wimsatt, William, 149
Tversky, Amos, 239,305311,3174322, Winston, Patrick, 372, 398n10, 408n5
327,39ln12,398nl6,410n33,417n18 Wittgenstein, Ludwig, 8, 64, 66, 383111,
400n 17
Ullman, Shimon, 237, 246 Woodward, James, 169
Unger, Peter, 107-108, 402n19 Woodworth, R. S., 297
Subject Index
Abandonment of belief, 223 Basic processes. See Processes

Absolute resource-independent criterion Bayesianism, 99, 101, 326, 328,343,
(ARI), 106 344445,355
Absolutism, 3, 69 Behaviorism, logical, 35
Abstract entities. 31 Belief, 3, 4, 13-16, 101, 162-177; categori-
Abstraction, 133 cal, 14,324326,333437; degrees of,
Acceptance, 101,324326,333337. See 14, 89-92, 113-115,324328; objects of,
also Belief 15, 163: reduction of, 16; replacement
Access, epistemic, 36, 148, 154 of, 16, 164-170; refinement of concept
Accretion model, 347,348 of, 16, 182, 199-200; and J-rules, 77-78;
Activated belief, 201-U)4. See also Activa- formation and logic, 82; -dependent vs.
tion; Belief; Memory; Retrieval -independent J-rules, 83; reality of par-
Activation: of long-term memory, 86, tial beliefs questioned, 114, 326-3528;
204407,384; spreading, S 2 0 6 . See and cognitive science, 167, 169; percep-
also Memory; Retrieval; Source nodes tual, 185, 197-198, 200; dispositional vs.
Act Utilitarianism. See Utilitarianism occurrent, 200-204; activated vs. unacti-
Analogy, 133,239-242,270. See also vated, U)1-204; perseverance, 214-221;
Matching; Similarity abandonment, 223; revision and updat-
Analysis, conceptual, 2,38-39; and skepti- ing, 224-226,344358. See also Accep
cism, 35-38,57 tance; Content; Doxastic attitudes; Folk
Analytic, 38, 57, 299 psychology; J-rules; Subjective probabil-
Anchoring and adjustment model, 345-357 ity; Uncertainty
Antinomies, 31 Biological fitness, 98
Antirealism. See Realism Bivalence, 143
A posteriotl, 299.301302 Bootstrapping, epistemic, 120
Apparent justification, 73-74 Bottom-up processing, 184-188
A ptiori. 34,299,301402 Brain. See Neural networks
Artificial intelligence, 134-135, 141, 283,
288,329 Calibration, 114-115, 116, 396n34; and
Art, and cognition, 139,238 the anchoring-and-adjustmentprocess,
Assertion, 13, 15-16, 17, 21 353-358; calibrational convergence,
Attention, 194, 351-352, 354 357-358
Autonomy, of epistemology, 1-2, 8-9 Cartesian demon. See Demon, Cartesian;
Skepticism
Backtracking, 220,4071-134 Categorization, 321-323
Basic beliefs, 79, 195, 198. See also Foun- Causal theory of knowledge, 42-44; and
dationalism skepticism, 36
Basic logical abilities, 284-286, 294 Causation, 35-36; and right J-rules, 85-87
430 Subject Index
Central system, 187, 197,200 Constraints, representational, 227-251

Certainty, 29,34, 68,300,301 Construction, of representations, 227-228
Church's Theorem, 281,300 Constructivism: mathematical, 34,143,
Circularity, 3; and skepticism, 32; in ac- 145; metaphysical, 153-154; perceptual,
count of justification, 61; and truth- 190-191
linkededness, 116-118; in epistemic ap- Content, 15; problems of, 162-177; and
proach to truth, 147 psychology, 170-173; and rationality,
Clairvoyance, 109-1 10, 111-112 173-175; and reliability, 173, 174,
Class inclusion, 229 175-177
Closure principle, 56 Contingent, 299,300,302
Codes, 164,252-277; sense-linked vs. Contradictories, 88
sense-neutral, 252254, 273-277; propo- Contrast effect, 345,346,349350,352,
sitional, 254: and code systems, 354
260-268; and memory, 266-267, Correspondence theory of truth, 151-
274-277. See also Imagery; Visual repre- 156
sentations Creativity, 248-249. See a b Originality
Cognitive impenetrability, 187,266 Credal attitudes. See Belief; Doxastic atti-
Cognitive processes. See Processes; Psy- tudes; Subjective probability
chology Credal residue, 224-226
Cognitive science, 1, 7, 96, 181-184,324, Credence, degrees of. See Belief; Subjec-
378-380; and belief, 16-17, 162-177; tive probability
and problem-solving activities, 132-135, Criterion: in verificationist sense,24; and
227,249. See also Psychology circular reasoning, 32; of rightness for J-
Cognitive-state transition rules. See J-rules rules, 63-89; of rightness for moral
Coherence: negative and positive, 99-100, rules, 64; theological, 64; of rightness
196; and coalitions, 339-340. See also for principles of justice, 64-65; of valid-
Coherentism ity of a law, 65; and reflective equilib
Coherentism: as rightness criterion for J- riurn, 66; sample criteria of J-rule
rules, 66,68,72,98,99-100,118-119; as rightness, 66-69; not derivable from
theory of right J-rule contents, 79-80; logic, 81-89; for second-order as well as
vs. foundationalism, 80, 194-198; and first-order processes, 94, 113, 114-115,
reliabilism, 80; as theory of truth, 144; 373-377; for methods, 94-95; conse-
autonomous or nonautonomous, 196; ho- quentialist, 95-103; reliabilist, 103-109;
listic or nonholistic, 197-198. See also comparative vs. absolute, 104; maximiz-
Coherence; Foundationalism; Holism ing vs. satisficing, 104-105; resource-in-
Communication, 5, 137, 141,247 dependent, 104-106; resource-relative,
Competition, 333335,337,338,339,340. 104-106. 115-116,251, 343; application
See also Neural networks; Uncertainty of, 118-119. See aZso J-rules; Justifica-
Completeness, of a rule system, 66 tion; Rightness
Complexity, 1!23, 141, 302; theory, Culture, 1, 140-141
282283,300 Cultural relativism. See Relativism
Computation, 170,282,328,329330,339
Computer, 172,254,360 Data structures, 261-263
Concept, 321-323; formation, 21, 131; in Decision, doxastic, 131, 136
mathematics, 271-272 Declarative vs. procedural knowledge. See
Confirmation, 89-91, 147; bias, 217-218; Knowledge
attitudes toward (and disconfirmation), Deductive reasoning. See Reasoning
347,348,349,355. See a h Evidence Demon, Cartesian, 33,37, 110, 113, 176.
Conjunction fallacy, 308310,319,320 See aZso Skepticism
Connectionism. See New connectionism Deontic, 25
Consequentialism, 3, 95, 97-103; objective Deontology, 3, 97
vs. subjective, 102-103 Depth of processing, 207,212
Subject Index 431
Derivability, of right J-rules from logic, Error, 26; and skepticism, 30-32. See also
81-89 Skepticism
Descriptional representations. See Repre- Evaluation, 20, 137-138, 247; as episte-
sentations mology's mission, 2 3 ; objects of, 3-6; of
Detachment, rules of, 73, 115 psychological processes, 4-5; factual
Dichotomy in epistemology, 299-304 basis of, 22-25, 63, 69, 78; regulative vs.
Dimensions of evaluation. See Epistemic nonregulative, 25-26, 59. See also Epis-
evaluation temic evaluation; Value
Direct skepticism. See Skepticism Evidence, 17, 145-146, 147, 339; eviden-
Discrimination, 47, 371-373 tial evaluation, 21, 22; evidential gaps,
Discount model, 347, 348 32; and best explanation, 37; theories of,
Dispositional belief. See Belief 37; evidential strength and J-rule right-
Doing the best one can, 104,3941122 ness, 66, 69; weight of, 89; proportion-
Domain of evaluation. See Epistemic eval- alism, 89-93, 97; retrieval of, from
uation memory, 133-134, 204-207; availability
Doxastic attitudes, 14: and J-rules, 77-78 of, 204; total evidence principle, 204;
Dreams, 239 and inference, 204-207; and belief up-
Duplex theory of memory, 200 dating, 344-358. See also Justification;
Dutch Book argument, 316,327428 Skepticism
Evolutionary epistemology, 2
Elaboration, 209-214, 217, 218-219. See Explanation: and theory of evidence, 37,
also Memory 249-250; as theory of justificational
Elenchus, 8 value, 98, 100-101, 140; and question
Empirical science, 2, 299,300. See also answering, 125; psychological, 172-173;
Science and perseverance, 217-218
Empiricism, 39-40 Externalism, 24, 72
Epistemic approach to truth. See External world, problem of, 31. See also
Truth Skepticism
Epistemic boundedness, 228,245
Epistemic decision theory, 101-103 Fallacy, 295-298,308410,311. See also
Epistemic evaluation, 20, 74; domain of, Conjunction fallacy
21; terms and dimensions of, 21-22: Fallibility: of cognitive faculties, 31; of
standards of, 22-25.26-27; styles reason, 31; of the senses, 31. See also
of, 25-26; multiple standards of, 122- Error; Skepticism
125. See also Evaluation, factual Fantasizer, 99, 100, 101
basis of First-order processes, 94, 369, 373
Epistemics, 9,28, 181-184,378-380. See Fittingness, 152- 156
also Primary epistemology Folk psychology, 14, 164-169
Epistemic utility, 101-102 Forgetting, 222-223,225
Epistemologically Original Position. See Fonnalism: mathematical, 34; in theory of
Original Position rationality, 353
Epistemology: as multidisciplinary affair, Formality condition, 170
1-3,8-9; and psychology, 1-5,6-7,37, Foundationalism: and basic beliefs, 79,
53,57,66, 163, 181-184,249,378480, 196; core doctrine, 79; and truth, 79;
383n4; history of, 6-8,299300,304. and coherentism, 80, 194-198; and relia-
378; and skepticism, 28-41; belief vs. bilism, 80; strong vs. weak, 196, 198
product, 138. See also Epistemics; Evo- Foundational questions, 8-9
lutionary epistemology; Individual Framework principle (for justification),
epistemology; Naturalistic epistemology; 59-63; rejection of iterative, 61; no-un-
Primary epistemology; Secondary dermining clause in, 62-63: historical
epistemology; Skepticism; Social episte- reformulation of, 83
mology Functionalism, 164
432 Subject Index
Generality problem, 49-51 Intelligence, 22, 122-124, 282, 283-284.

Generalization, 371373 See also Power
Geometry, 268,272 Interference, 223-226
Gestalt rules, 230-233,m,237, 246 Internal codes. See Codes
Global reliability. See Reliability Internalism, 24, 72
Global skepticism. See Skepticism Internal realism, 35, 144
Goal-responsiveness (directedness):rela- Interpretation, principles of, 120, 164,
tion to power and intelligence, 123-124, 174-177
287,288; and deductive reasoning, Intuition, sensuous, 268, 269
287-288, 303; and production systems, Irrationality. See Rationality
368-369 Iterative skepticism. See Skepticism
Grammar, 228, 235, 238, 246
J-rules, 60; permissive vs. obligatory, 60,
Hierarchical structuring, 229-235, 74; systems of, 60, 78, 83, 113, 115, 277,
243-245 339; direct and indirect, 74-75; inter-
Holism, 197-198; inferential, 206- personal vs. intrapersonal, 75-76: as
207 cognitive-state transition rules, 77-78,
Hypothesis formation, 21, 131, 239, 85-88; and logic, 81-89; belief-depen-
250-251 dent and belief-independent, 83; right J-
rules as cognitive process rules, 85-89;
Idealism, 34 and psychology, 87-89; as proportioning
Ideally logical being (ILB), 279-283 degree of belief to weight of evidence,
Ignorance, 26-27, 122-123 89-91. See also Criterion; Justification;
Illusion, 322 Rightness
Imagery, 253-268; and perception, Justification, 3, 21,s-24, 181; vs. discov-
256-259; and problem solving, 263-268; ery, 6; social conception of, 8, 18, 68,
and reliability, 263-268; and mathemat- 75-77; and skepticism, 29,40,61-62;
ics, 268-273. See also Codes; Imagina- and knowledge, 42,53-55, 76; reasons
tion; Perception-similitudethesis; for interest in, 58; and the rule frame-
Representations work, 59; framework principle for,
Imagination, 139-140, 248, 252-253. See 59-63; and no-undermining clause,
also Imagery; Originality; Preferences; 62-63; levels of theory of, 63, 66; and
Representations psychology, 66, 96, 306-307; nihilist,
Incoherence, probabilistic, 312, 315-318. monist, and pluralist views, 70-71; sub-
See a h Probability; Subjective proba- jectivist and objectivist views, 71-73;
bility real vs. apparent, 73; historical compo-
Inconsistency, 73,84,88,313-315 nent in, 76, 78,83; and evidence
Incorrigibility, 195-196 proportionalism, 89-91; primary, sec-
Indexical subjectivism, 71 ondary, and full, 93-95; and normal
Individual epistemology, 1 , 9 worlds, 107-109; ex post vs. ex ante,
Induction: Hurnean problem, 32,393n21; 112; and rival hypotheses, 250-251. See
new riddle, 385n10 also Criterion; Framework principle; J-
Inductive inference. See Inference rules; Justifiedness
Inference, 4; rules of, 82; and memory, Justifiedness: mixed conception of, 73,95;
204-207.214. See also Logic; Natural full, 93-95; primary (P-justifiedness),
deduction; Reasoning 93-95, 159,318,320,369; secondary (S
Informational encapsulation, 187, 192 justifiedness), 93-95, 159,318,320,369
Information pcmessing unit, 329-334, 336,
338339. See a h Computation; Neural Kinematics of belief, 99, 349. See also Up-
networks dating, of belief
Input systems, 187 Knowledge: loose vs. strict sense of, 19;
Intellectual products. See Products and realism, 19; and justification, 42,
53-55; and causal processes, 42-44; and 303304; working, declarative, and pro-
reliability, 43-53; and Gettier problems, duction memories, 363364, 369-373.
46-47.54-55; and no relevant alterna- See also Credal residue; Depth of pro-
tives, 4 7 4 5 4 - 5 5 ; of necessities, cessing; Elaboration; Reconstruction, in
48-49; and second-order processes, memory
51-53; and skepticism, 55-57; first-order Mental logic, 288,287-290.293-298,
vs. higher-order, 56-57; declarative vs. 303304
procedural, 369. See also Causal theory Mental models, 287,290-293,298-299,
of knowledge; Justification;Reliability 304
Mental rotation, 255, 258, 265
Language acquisition, 228,246-247 Metainduction, "disastrous", 158-160
Language games, 8,66 Metaphor, 239-240
Legal positivism, 65 Metaphysical realism, 144
Legal realism, 65 Metaphysics, as a response to skepticism,
Levels: in the theory of justification, 63, 34-35
96-97; of the theory of justice, 64-65; of Metareliability. See Second-order proper-
justifiedness, primary and secondary, 93 ties
Local reliability. See Reliability Methodology, 2, 91-93, 148; of psychol-
Local skepticism. See Skepticism ogy, 7,184
Loci, method of, 264 Methods, 3, 21, 52, 91-93; acquisition of,
Logic, 2, 5, 6, 7, 82; and rightness criteria 93-94,369373; contrasted with pro-
for J-rules, 66,67-68,81-89; and con- cesses, 93-95, 366; selection of, 94,
tent, 174-175; and justifiedness, 364-366:rightness of, 94-95. See also
278-279; and rationality, 279-284 Productions
bgicism, 34 Mind-object relation, 31
Long-term memory. See Memory Models. See Mental models
Lottery paradox, 115,325 Modularity, 187, 239,341
Modus ponens, 67,89,285
Matching, 50-51, 321423, 364366,of Monism, 70
productions, 365, 367-368; and reliabil- Moral and ethical philosophy, 8, 73, 74
ity, 367; and power, 368. See also Pro- Morality: rightness criterion of, 64; moral
ductions; Similarity codes, validity of, 69
Mathematics, 34-35, 141, 143, 145; and Motion, representation of, 237
sense-linked representation, 268-273; Music, representation of, 233-235
and proof, 269-270; and concept forma- Mutual inhibition, 330, 333
tion, 271-272
Maximizing: true belief, 66; value, 97; cri- Natural deduction, 287
terion of rightness, 104-105 Naturalistic epistemology, 2, 163,378,
Meaning, 38-39, 150-151, 155,385n19 383115
Medium, representational, 257,258-259, Necessary and sufficient conditions, 38,
261-264 385n19
Memory, 199-226; long-term vs. short- Necessary truths, 299; knowledge of,
term, 86, 200-204;retrieval, 133-134, 48-49
137; power, 199,208,212-214,219, Negative coherence. See Coherence
221-223; duplex theory of, 200; acti- Neural networks, 329-337
vated vs. unactivated, 201-204; and in- New connectionism, 328338,339
ference, 204-207; and rationality, Nihilism, 70
204-207; and reliability, 208, 210-214, Normal worlds, 107-109, 113
219; and forgetting, 222-223; and inter- Normative judgments. See Evaluation
ference, 223-226; span, 244; and sense-
linked codes, 266-267,274277; and Objectivism, 3, 69, 71-73; objectivist crite-
logic tasks, 287-289, 291, 292,301, 302, rion of justificational rightness, 94-95
434 Subject Index
Obligation, J-rules as rules of, 60, 388n3 Pragmatism, 8

Occurrent belief. See Belief Praise and blame, 74-75,384116
Operations, representational, 238-242, Preferences, among representational oper-
248-251. See also Preferences; Repre- ations, 238, 242, 246, 248-249
sentations Primacy effect, 352-354
Order effects, 350354,355456 Primary epistemology, 1, 4-5, 9, 93,
Ordinary-language philosophy, 6 181-184,271,273,283-284,300,306,
Originality, 22,247-249 369, 378-380. See also Epistemics
Original Position, 65, 67-68 Probability: theory of, 2,5,326,328; am-
biguity of, 24; and criterion of J-rule
Paradox: Russell's, 93; liar, 144 rightness, 66, 67, 90, 99; judgments, 305,
Parallel processing, 329 307411; and rationality, 311-318; o b
Partial belief. See Belief; Subjective proba- jective and subjective interpretations of
bility probability judgments. 312,315-316,
Partial cognitive states, 78 328,336, 337,338, 342; logical, 312-
Partitions, representational, 228-234, 313; epistemic, 313; and justifiedness,
235-236,243-244,247 318423. See also Subjective probability
Pattern recognition, 184-185 Problems: solving, 27, 123-128; definition
Payoffs, 342 of, 126; dehition(s) of solution to,
Perception, 150, 154, 184-194, 197-198, 127-128; and cognitive activity,
230, -237.335-336; and reliability, 131-136, 240-241; representation of,
189-191, 193; and speed, 191-193; and 132-133, 229; problem solving as social
pragmatic values, 193; and power, enterprise, 136-138. See also Power, in-
193-194; and the a priori/a posteriori tellectual; Questions
distinction, 301-302. See also Visual Procedural learning, 389373. See also
representation Productions
Perception-similitude thesis, 256-259 Processes: psychological, 3,4-5, 21, 26,
Permission, J-rules as rules of, 60-61 391n7; second-order, 21,51-53,94,367;
Perseverance, of belief, 214-221 and right ]-rules, 85-89, 90-95; under-
Phenomenalism, 34,35 stood as basic processes, 92,366,
Philosophy of science. See Science 391nll; contrasted with methods,
Phoneme-restoration effect, 186, 189- 93-95,293,366
190 Productions, 288,359377: application of,
Phrase structure, 229 362-363; as methods rather than pro-
Pictorial representations. See Representa- cesses, 366; acquisition and adjustment
tions of (procedural learning), 369-373
Plans, 229,368-369 Products, intellectual, 137-138
Platonism, mathematical, 34 Proof: in mathematics, 268, 269-271,
Pluralism, 3, 70-71 272-273; in mental logic, 287-289,
Positive coherence. See Coherence 295
Positivism, 6; legal, 65 Proof theory. See Syntax
Power, intellectual: as antidote to igno- Proportionalism, evidence. See Evidence
rance, 26-27; as capacity to solve prob- Propositional attitudes, 14-17, 164-166
lems, 27, 122-125; and perception, Propositions, 14, 15-16, 168
193-194; and representational con- Prototype theory, 321422
straints, 244, 245-246; and imagery, Psychologism, 6, 7
263-288,271-273: and WTA mecha- Psychology: and epistemology, 1-5, 6-7,
nisms, 341; p d partial matching, 368. 37,53, 57, 66, 181-184,306307,
See also Power, memory 378380; and right J-rules, 87-89,96;
Power, memory, 199,208,212214,219, and subjective probabilities, 114,
221-223 325-328; and belief content, 162-177.
Pragmatic values, 98, 193,341-343 See also Cognitive science; Processes
Subject Index 435
Questions: answering, 27, 123-124, Knowledge; Second-order properties;

125-129: types of questions, 128. See Truth ratio
also Power, intellectual; Problems Representations, 164, 170-171, 172,
227-251; operations on, 238,241-242,
Rationalism, 39-40 248-249; pictorial vs. descriptional, 255-
Rationality, 21, 27, 40, 70; and content, 259; deep vs. surface, 262. See also Codes;
173-175; and memory, 204-207; and Operations; Partitions; Preferences
psychology, 279; and logic, 279-284; Representation theorem, 327
basic vs. acquired, 283-284; and proba- Representativeness heuristic, 307,
bility judgments, 305318; and belief 310311,320323
updating, 353 Retrieval, 133-134, 137; and reasoning,
Ratio, of true beliefs. See Truth ratio 204-207; and spreading activation,
Realism, 17, 142-161; internal, 35, 144; 2Q5-U)6. See also Activation
legal, 65; scientific, 131, 158-161; meta- Rightness: of J-rules, 60; criteria of, 63-69;
physical, 144-156: convergent, 157-160 in morality, 64; in justice, 64-65; of pro-
Real justification, 73-74 cesses and methods, 93-95. See also Cri-
Reasoning: and formal logic, 82; and acti- terion; Justification
vated belief, 204-207; deductive, Rival hypotheses: and skepticism, 32-33;
278-299. See also Inference; Proof and justification, 250-251, 338339,340,
Recency effect, 349351,352-354 341. See also Competition
Reconstruction, in memory, 209-214. See Rules, 59. See also J-rules
also Memory
Reduction: metaphysical, 3435; inter- Satisficing, 97, 104-106, 116
theoretic, 164, 165-166. See also Behav- Science, 8, 136, 140-141, 149, 156-161;
iorism; Phenomenalism philosophy of, 6, 21, 40, 96, 125,
Reference, 35, 155, 157 129-130, 135, 184,357-358; history of,
Reflective equilibrium, 66,81 8, 241; sociology of, 8; and imagery, 265
Relativism, 69, 70-71, 95 Search strategies, 21, 134-135
Relevant alternatives, 45, 46, 54-55. See Secondary epistemology, 7, 93, 277, 300,
also Knowledge; Reliability, local 369
Reliability: as antidote to error, 26; defini- Second-order processes, 21,51-53, 92,
tion of, 26; metareliability (second-order 94-95, 113, 115-116,369,373477. See
reliability), 27, 52-53, 115-1 16, also Processes
374377; approach to knowledge, Second-order properties: second-order reli-
43-53; causal reliability approach, ability (metareliability),27, 52-53,
43-44; reliable indicator approach, 115-1 16,374-377; second-order power,
43-44; reliable process approach, 44-51; 27,375; second-order speed, 27,375
global vs. local, 4445, 47, 48; in actual Semantics: model theory, 81-82, 279; se-
vs. counterfactual applications, 45. mantic properties, 162, 170-171, 260,
47-48; in pure subjunctive vs. relevant 261. See also Logic
alternative situations, 45-47, 48; as a Sensory codes. See Codes; Imagery; Repre-
propensity, 49; and the generality prob- sentations
lem, 49-51; as criterion of justificational Similarity, 239-242, 310, 321-323. See also
rightness, 94-95, 103-109; criterion, ob- Analogy; Matching
jections to, 109-113; and content, 173, Simplicity, 37
174, 175-177; and perception, 189-191; Skepticism, 28-41; Pyrrhonian, 28-29;
and memory, 208,210-214.219; and global vs. local, 29, 33, 40; about justifi-
representational constraints, 246, 251; cation, 29, 40, 61-62, 97, 105-106;
and imagery, 264-268; and deductive about knowledge, 29,40,55-57;
reasoning, 293-299; and representative- strength of, 2940; direct vs. iterative,
ness, 320-323: and WTA mechanisms, 30,56-57; and human cognitive powers,
338-340, 342-343. See also Justification; 30, 37; and error possibilities, 30-32;
436 Subject Index
Skepticism (cont.) Trade-offs, 124, 183, 192-193, 212,214,

grounds for, 3033; about the external 247
world, 31; about other minds, 31; and Transcendental arguments, 37-38
circularity, 32; and evidential gaps, 32; Transformations, 261-263. See also Opera-
and induction, 32.3931121; and rival hy- tions
potheses, 3233; responses to, 33-39; Triangulation, 149-150
and metaphysical reconstruction, 34-35; Truth, 3,5, 142-156; -linked standards, 3,
and conceptual analysis, 35-38; and reli- 69,72, 116-121; -values, 17.20; con-
able processes, 39, 40; and justificational trasted with evidence or justification,
nihilism, 70, 92-93; and scientific 17-18; contrasted with knowledge,
theories, 160 18-19; true belief as valued conse-
Social epistemology, 1, 5-6, 7-8, 136-138 quence, 98, 100-103, 138-141; and
Social philosophy, 8. See also Moral phi- problem solving, 125, 127, 130-131:
losophy epistemic vs. nonepisternic conceptions
Source nodes, 364 of, 144-151; and correspondence,
Spacing effect, 221-222 151-156; contrasted with truth-condi-
Speed, 27, 122, 124-125; and perception, tions, 155-156; truths of reason vs. mat-
191-193; and representational con- ters of fact, 299300. See also Realism;
straints, 246-247; and complexity the- Reliability; Truth ratio
ory, 282-283; and WTA mechanisms, Truth ratio, 26, 103-104, 113-115,395n23;
340-341: and partial matching, 368 maximizing vs. satisficing of, 104-
Spreading activation, 205-206 105: and scientific realism, 159-160.
Standards of evaluation. See Epistemic See also Criterion; Justification; Relia-
evaluation; Evaluation, factual bility
basis of Typicality, 321
Statistics, 2,s. 72-73, 90-92, 276-277,
311-312,31!3-320 Unactivated belief, 201-204. See also Acti-
Stimulus theory, 190-191 vation; Belief; Memory; Retrieval
Story understanding, 229 Uncertainty, 326,328,335,336-338
Strength, of productions, 356-366, Undermining, and justification, 62-63,
373-377 111-112
Subjective probability, 14, 40, 90, 99, Unit, information processing. See Informa-
101-102,113-115,312,324-326,344; tion processing unit
psychological reality questioned, 114, Universal cognitive state, 78
326-328,336337. See also Belief, par- Updating, of belief, 344-358
tial; Probability; Uncertainty Utilitarianism, 23, 102
Subjectivism, 3, 24, 69, 71-73, 95 Utilities, 343
Supervenience,22,23
Synthetic, 299,300 Validity: in behavioral sciences' sense, 26;
Syntax: proof theory, 81-82,279; syntactic of arguments, 67; of a sentence, 81; and
approach to the mind, 170-173; syntac- right J-rules, 83-88
tic properties, 261. See also Logic Value, 20; justificationally relevant,
Systems, of rules. See J-rules 97-103; monism vs. pluralism, 138-141;
adjunct and specialized, 139-141. See
Template, 50 also Evaluation; Pragmatic values; Veri-
Temporal strings, 229, 233-235,239 fic
Theory-ladenness, 188-189 Verific, 5; value, 98. See also Truth
Topdown pryessing, 184-188, 189-192, Verification, 24-25; and transcendental ar-
368-369 guments, 38; irrelevant in a criterion;
Total cognitive states, 78 64-65; and legal theory, 65; and truth,
Total evidence principle. See Evidence 143-151. See also Evidence; Justifica-
Tracking, 45 tion; Truth
Subject Index 437
Virtues and vices, epistemic, 25. See also Weight, of evidence. See Evidence
Epistemic evaluation Winner-take-all (WTA) network, 329,
Visual buffer, 262, 264,265, 267 333-334,338-343,417n19
Visual representation, 229-233, 235-237
Vividness, 268, 273-277

Epistemology and Congnition - Alvin I. Goldman

Uploaded by

Copyright:

Available Formats

Epistemology and Congnition - Alvin I. Goldman

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Epistemology and Congnition - Alvin I. Goldman

Uploaded by

Copyright:

Available Formats

EPISTEMOLOGY A N D COGNITION

HARVARD UNIVERSITY PRESS

Publication of this book has been aided by a

Library of Congress Catabging in Publication Data

Writing of the book began in earnest in 1981-82, when I was a fellow of

I. 1. Epistemology as a Multidisciplinary Aflair

If epistemology is a branch of psychology, or evolutionary theory, the

12. Objects of Evaluation

evaluated? Epistemologistshave often held that the prime objects of evalu-

and to devise new instruments. On this conception logic, probability theory,

of channel controllers--such as editors, media owners, or boards of educa-

sages be cultivated? Which principles of selectivity are suitable for differ-

1.3. Continuities and Discontinuities with Historical Epistemology

paradigms of epistemology became the logic of confirmation, the analysis of

The same holds for epistemology. Although foundational questions may

The Elements of Epistemology

1.1. Beliefs, Assertions, and Propositions

They urge us to do epistemology in terms of degrees of belief, or perhaps

I favor a combination of the neutrality and refinement approaches. I do

1.2. Truth-Values and Knowledge

1.3. Epistemic Evaluation

( 1 ) Domain of ewalwltion. Which 'things' are evaluated in epistemic eval-

Many terms of widespread intellectual evaluation have not figured pro-

to answer by using other terms of epistemic appraisal, such as 'rational',

of epistemic evaluation where substantive conditions are sought. (This

seeking action-guiding principles. Indeed, when it comes to fixed or auto-

1.4. Reliability, Power, and Speed

lectual) power. Power is the capacity of a process, method, system, or what

2.1. Varieties and Grounds of Skepticism

grades of knowledge or justification. It might hold that we are not the

The difficulty of knowing about an abstract entity may be assigned to the

can mere behavioral access to a humanoid creature preclude this possibility

2.2. Responses to Skepticism: Metaphysics, Analysis, and Psychology

The remaining categories of response to skepticism are all one or another

Hume's skeptical tendencies). At any rate, twentieth-century Humeans who

thereof. Recent transcendental arguments have tried to show the absurd or

gical terms. This can be regarded as a first approximation to a better way of

2.3. How Important Is Skepticism?

justifiedness for partial beliefs. This would be of interest even if skepti-

3.1. Knowledge and Causal Processes

reported). In this variation, the cause of your belief is suitable for

3.2. Alternative Reliable-Process Theories

forming processes, or methods, are not globally, or generally, reliable. Gen-

instance, namely, the instance in question. Since this instance by hypothesis

3.3. Second-Order Processes

warned that Fraud has no credentials in mathematics. Humperdink hears

high. However, this requirement might be too strong. A weaker require-

3.4. Knowledge and Justif;catwn

3.5. Knowledge and Skepticism

Justification: A Rule Framework

4.1. Justification and Rules

there seems to be a core idea of justifiedness, which my theory will seek to

of theorizing are needed to get a complete account. I will return to these

because of some appropriate antecedent or current state. Thus, someone

(P2) S's believing p at time t is justified if and only if

for a correct account of justifiedness. We must not assume the skeptic to be

4.2. Rightness Criteria for ./-Rules

correct principles of justice. This parallels principle (PI). Second, he offers a

L is a law if and only if