Computation and Intentionality: A Recipe for Epistemic Impasse

Minds and Machines (2005) 15: 207–228 DOI 10.1007/s11023-005-2004-2
Springer 2005 Computation and Intentionality: A Recipe for Epistemic Impasse I. SHANI The University of Haifa, Mount Carmel 31905, Haifa, Israel; E-mail: ishani479@hotmail.com Abstract. Searle’s celebrated Chinese room thought experiment was devised as an attempted refutation of the view that appropriately programmed digital computers literally are the possessors of genuine mental states. A standard reply to Searle, known as the ‘‘robot reply’’ (which, I argue, reflects the dominant approach to the problem of content in contemporary philosophy of mind), consists of the claim that the problem he raises can be solved by supplementing the computational device with some ‘‘appropriate’’ environmental hookups. I argue that not only does Searle himself casts doubt on the adequacy of this idea by applying to it a slightly revised version of his original argument, but that the weakness of this encodingbased approach to the problem of intentionality can also be exposed from a somewhat different angle. Capitalizing on the work of several authors and, in particular, on that of psychologist Mark Bickhard, I argue that the existence of symbol-world correspondence is not a property that the cognitive system itself can appreciate, from its own perspective, by interacting with the symbol and therefore, not a property that can constitute intrinsic content. The foundational crisis to which Searle alluded is, I conclude, very much alive. Key words: Bickhard, computational theory of the mind, encoding, intrinsic intentionality, Searle, the Chinese room, the robot reply 1. Introduction John Searle’s Chinese room thought experiment (1980) is, without a doubt, one of the most celebrated, and most pointed and divisive, criticisms of the computational (symbol-manipulation, information processing) theory of the mind. From the moment of its inception, Searle’s Gedankenexperiment provoked strong reactions. Many have dismissed the argument, giving it short ‘‘refutations’’ (see for example, Abelson, 1980; Block, 1980), yet its import continues to reverberate in the milieu of cognitive science, artificial intelligence, and the philosophy of mind.1 As Searle himself made quite clear, the aim of the Chinese room argument was to show that digital computers do not, and cannot, exhibit intrinsic intentionality and hence that the intentionality of intelligent beings, humans included, does not, and cannot, consist of computer-like symbol manipulation. In what follows, I examine Searle’s argument in some detail. I proceed to discuss one of the major rebuttals to Searle’s charge – the robot reply (RR). 208 I. SHANI I argue that RR reflects the most popular approach to the naturalization of mental content, an approach inspired by the computational theory of the mind (CTM). According to RR, what constitute mental content are symbolworld correspondence relations – in complete independence from, yet in supplementation to, the system’s cognitive architecture. I argue that not only does Searle himself cast doubt on the adequacy of this idea by applying to it a slightly revised version of his original argument, but that the weakness of this encoding-based approach to the problem of content can also be exposed from a somewhat different angle. Capitalizing on the work of several authors, and, in particular, on that of psychologist Mark Bickhard, I argue that the existence of symbol-world correspondence is not a property that the cognitive system itself can appreciate, from its own perspective, by interacting with the symbol and therefore not a property that can constitute intrinsic content. Thus, attempts to deal with the problem of content within the dominant information-processing approach fail to live up to the expectation (shared by the majority of its own practitioners) of articulating a notion of mental content that can make epistemic sense from the first-person perspective of a psychological agent. The result, I submit, is a failure to naturalize content. 2. Part One: Intrinsic Intentionality and the Chinese Room 2.1. SEARLE’S AARGUMENT AS A QUA ARGUMENT What is the Chinese room argument an argument for: The Chinese room thought experiment is devised as an attempted refutation of the view that appropriately programmed digital computers literally are the possessors of genuine mental states. As such, the argument is a direct assault on the CTM, often described as ‘‘cognitivism’’ (Harnad, 1990; Searle, 1992) or ‘‘the cognitive science view’’ (Fodor, 1975; Sayre, 1987). According to CTM, mental processes are computational processes performed on internal symbol strings; intelligent beings are, therefore, digital computers of the appropriate sort, and the challenge, from the point of view of theoretical cognitive research, is to identify, understand and (eventually) be able to construct such intelligent formal automata. The Chinese room argument aims to show that computerlike information processing is not only insufficient for the having of genuine mental states but that, in all likelihood, it plays no constitutive role in the making of intentional agency (mental life). If appropriately programmed computers really are minds, Searle reminds us, then they ‘‘can be literally said to understand and have other cognitive states’’ (1980, 282). Correspondingly, Searle constructs his argument as a counterexample to the claim that (a) appropriately programmed digital COMPUTATION AND INTENTIONALITY 209 computers literally are the owners of intrinsic intentional states and (b) the modus operandi of such computational machines explains human cognition. To test the claim that all minds are computational machines in CTM’s sense, Searle suggests, we need to ask ourselves what it would be like if our own minds operated in a similar manner to such machines. If computer-like symbol manipulation constitutes understanding and other intrinsically intentional states, then we, too, ought to understand (and to have other cognitive states) in virtue of performing such computations. The Chinese room argument is devised as a modus tollens to this conditional. It is meant to cast serious doubt on the idea that human cognizers understand (and possess other intentional states) in virtue of performing computational procedures on symbol structures, and by doing so the aim is to undermine CTM’s aspirations to provide the key to the understanding of intelligent behavior. In the scenario Searle imagines, he is situated in a room, manipulating Chinese characters, which he does not understand, according to written instructions in English, which he does understand. As it turns out, some of the input in Chinese that Searle receives is interpreted by outsiders as questions, and the Chinese characters Searle issues as an output are interpreted as answers to those questions. Occasionally, Searle is also asked questions in English, and he issues answers as he sees fit.2 Now, Searle maintains that in the Chinese case he himself operates like a computational machine, he manipulates symbols according to rules, and the output he issues is interpreted by outsiders as intelligent answers to specific questions. Yet all the while Searle understands nothing. In the English case, on the other hand, Searle understands perfectly well, but nothing in the example suggests that his understanding has anything to do with computer-like information processing. From his lack of understanding of the Chinese symbols he manipulates, Searle concludes that a computing machine which operates along similar lines of symbol manipulation does not understand either, and does not literally provide answers to the questions it is being asked. Since (leaving aside practical considerations such as the amount of time required for accomplishing complex computations) Searle is capable of performing all the computations the machine is capable of performing, if what the machine does were to generate understanding, then he, too, would have understood the questions he is being asked. Yet, as the argument shows, no such understanding ensues. The intended upshot of the Chinese room example, then, is to argue that digital computers lack understanding in any literal sense of the word. Since, qua computational machine, Searle understands nothing, and since, qua intentional agent, his understanding seems to have nothing to do with computerized symbol manipulation, the argument suggests that computer-like 210 I. SHANI information-processing adds nothing to the amount of intentional content in the room and, by way of extrapolation, that it adds nothing to the sum total of intrinsic intentionality in nature. Moreover, a significant feature of the argument is that while Searle himself does not understand Chinese, the Chinese characters he issues as output in reaction to the questions given to him are treated by people outside the room as meaningful verbal responses. This fact marks a distinction that is crucial for Searle, the distinction between the first-person perspective and the thirdperson perspective. Genuine intentional content, he maintains, must be manifested as a first-person property – it must be intrinsic to the system itself and cannot be derived from the intentionality of observers, users, or programmers. Unfortunately for CTM, Searle argues, ‘‘such intentionality as computers appear to have is solely in the minds of those who program them and those who use them, those who send in the input and who interpret the output’’ (Ibid., 301). 2.2. NATURALIZED CONTENT AND INTRINSIC INTENTIONALITY Interestingly, while sympathizers with CTM are almost unanimously in large disagreement with Searle’s conclusion, the majority amongst them agree with him on two of his most basic assumptions: first, that the distinction between intrinsic and derived intentionality is theoretically sound and important; second, that the success of a theory of psychological content hinges on its ability to explain intentionality as a system-intrinsic phenomenon (Block, 1990; Dretske, 1988; Fodor [in Dennett, 1987, 288], Millikan, 1989; see, however, Dennett, 1987). Indeed, one of the attractive features of CTM was that it carried a promise for the reconciliation of naturalism with realism about intentional phenomena. On the premise of intentional realism, however, the distinction between intrinsic and derived intentionality is mandatory. Realists believe that intentional states are not just in the eye of the beholder. If representation is real, if there are positive facts of the matter regarding the possession of intentional states, there must be some systems that possess such states, and the possession of such states in such systems cannot be explained by reference to the intentional states of still other systems, on pain of regress.3 It transpires, then, that a failure to respect the distinction between derived intentionality and intrinsic intentionality, and to explain the possibility of the latter, should count heavily against a theory – to the point of constituting a reductio of the theory (again, at least insofar as intentional realism is accepted as an axiom...). It is precisely because it challenges the appropriateness of CTM as a framework on which to found an explanation of intentionality as a system- COMPUTATION AND INTENTIONALITY 211 intrinsic phenomenon that Searle’s argument has provoked such strong reactions. If Searle is right, then ‘‘the basic idea of modern cognitive theory’’ (Fodor, 1981) suffers from a vividly serious theoretical incoherence. Despite being considered by many as the key to unraveling the mystery of cognitive phenomena, it would seem that, on the premise that minds are digital computers, it becomes impossible to explain the real thing about aboutness, the intrinsic intentional contents of the cognitive states of intelligent beings. Put differently, if Searle is right, CTM fails to sustain a notion of mental content that can make explanatory sense as a first-person property of psychological agents, without presupposing the epistemic contribution of observers, programmers, users, interpreters, or designers. 2.3. THE CHINESE ROOM AGAIN: THE ROBOT REPLY Many objections have been raised against the conclusions Searle draws from his thought experiment, and Searle himself addresses some of them in his original paper as well as in his replies to peer commentaries. In the context of the present discussion, I would like to concentrate on one of these objections. One of the major lines of criticism addressed against Searle is an argument that Searle calls ‘‘the robot reply‘’. The main point of RR is this. While Searle may be right that a stationary computational machine of the sort targeted in his argument lacks genuine understanding, the problem can be solved if we allow the computer to be placed inside a robot in such a way that (i) The computer receives information input from peripheral ‘‘sensory’’ processes within the robot, which in turn encode environmental events in the form of transduced energy patterns. (ii) The computer issues information output that controls the robot’s ‘‘motor’’ apparatus. Such a computer, the idea goes, would be hooked up to the environment and would be in charge of operating the robot’s seemingly intelligent behavior, and this is all that may be required to sustain intrinsic intentionality. Two points need to be made with regard to RR. First, as Searle himself emphasizes, ‘‘it tacitly concedes that cognition is not solely a matter of formal symbol manipulation’’ (1980, 293). Second, while conceding that cognition and computation are not strictly identical, RR nevertheless persists in presupposing the adequacy of CTM as a theory of cognitive architecture. That is to say, according to RR there is nothing wrong with the idea that mental processes are formal, symbol manipulation, processes; it is just that such processes must be supplemented with content-generating environmental hookups (see below). 212 I. SHANI In reaction to RR, Searle’s claims that the reply does nothing to alleviate the problem. The addition of such ‘‘sensory’’ and ‘‘motor’’ capacities’’, he argues, ‘‘adds nothing by way of understanding, in particular, or intentionality, in general, to Schank’s original program’’ (Ibid.). In order to show that this is, indeed, the case, Searle offers an updated version of his original Chinese room argument. Suppose, he says, that, as in the original scenario, I am situated in a room wherein I perform blind computations on Chinese characters. Only this time, unbeknownst to me, the room is placed inside a robot in such a way that some of the input I receive comes from a television camera in the robot’s head that ‘‘perceives’’ external events, and the output I issue contains instructions to the robot’s motor apparatus. Still, Searle maintains, he himself – the robot’s cognitive processor as it were – has no grasp whatsoever of the meaning of the ‘‘information’’ he receives and of the motor ‘‘instructions’’ he issues, while the rest of the robot’s story consists of nothing but non-intelligent receptors and effectors (mechanical and electric processes of energy transduction, limb movement etc.). There is nothing here to sustain intrinsic intentionality, let alone a sophisticated human-like intelligence. In reply to Searle’s reply, Fodor has argued that Searle’s treatment of the robot reply is quite unconvincing. Given that there are the right kinds of causal linkages between the symbols that the device manipulates and things in the world – including the afferent and efferent transducers of the device – it is quite unclear that intuition rejects ascribing propositional attitudes to it. All that Searle’s example shows is that the kind of causal linkage he imagines – one that is in effect mediated by a man sitting in the head of a robot – is, unsurprisingly, not the right kind (1980, 520; for a similar argument see Hayes et al., 1992). From the point of view of our present concern, Fodor’s reply is of particular interest since it summarizes in a nutshell the most popular approach to the problem of content. Indeed, understood as it is understood within the dominant cognitivist tradition, the entire project of naturalizing mental content can be viewed as a variant of RR. It consists of an attempt to supplement the basic cognitive architecture of CTM with causal or informational hookups, or with teleological evolutionary history, premised on the assumption that such addition is necessary for a complete account of mental content but that, nevertheless, CTM is an adequate theory of mental processing. Opinions are widely diverged regarding the merits of specific proposals, but none of the leading theories of mental content currently in vogue – information semantics, teleosemantics, or conceptual role semantics – challenges the validity of CTM as a theory of cognitive processing. Thus, the basic supposition is that the symbolic approach to cognition provides a more or less adequate account of information (or ‘cognitive’) processing, and that COMPUTATION AND INTENTIONALITY 213 the project of naturalizing content consists in articulating a supplementary story, couched in naturalistic terms, of the conditions under which the symbols that partake in the processing acquire semantic values. On this view, then, cognition simply is computation plus symbol-world linkage, from which it follows that computation plus symbol-world linkage, is sufficient for intrinsic intentionality. As a result, it is believed that there must be some symbol-world linkages that, when joined with the appropriate computing device, simply are content-constitutive. This is, precisely, the issue between Searle and the dominant cognitivist approach to the problem of intentionality of which Fodor is such an avid representative. In response to Fodor, Searle argues that to rule out conscious implementation as an illegitimate agency of causal mediation would be arbitrary and unjustified. Precisely because symbol systems are formal, such implementation details need not matter (see Section 3.1 below); and, as Searle reminds us, some of Turing’s own examples of Turing machines involved conscious agents going through the steps of the program. Moreover, even if we grant Fodor’s point that CTM need not imply that every instantiation of a Turing machine is constitutive of intrinsic intentionality (see 1980, 525), we would be hard-pressed defending the claim that the class of implementations that do constitute such intentionality necessarily does not include conscious implementations. In order to demonstrate the arbitrariness of such exclusion, Searle introduces another thought experiment. Suppose, he says, that we found that the quickest way and most efficient way to make computers work was by utilizing tiny creatures imported from Mars who were the size of a few molecules and who implemented the program by consciously going through all the steps in the symbol manipulation. . . We can suppose that they find it enormous fun to do this, so they enjoy the work and are willing to work for no pay, and that they compute faster than any known electronic circuits. They are better, cheaper, and easier to program than silicon chips. (1980, 526). ‘‘Now’’, Searle concludes, ‘‘it would not be even remotely plausible to say that in this eventuality they were not actually going through the steps of the program.’’ (Ibid.). Plausibility notwithstanding, the point of this little thought experiment is to show that there is no well-motivated reason to exclude conscious rule following from being a potential machine-table implementation of intelligent agency (if such implementation is to be considered possible at all). Nothing in the logic of the mind-as-computer hypothesis suggests a thesis so strong as the thesis that the implementational infrastructure must, of necessity, be nonconscious, precisely because nothing in this logic hinges heavily on any implementation details. 214 I. SHANI More importantly, Searle insists that the fact that he can iterate his original thought experiment under the new (robot) circumstances, shows that no amount of causal hookups can help generate intentional content as long as the thinking-qua-computation hypothesis is retained. The robot reply consists of adding externalist causal connections to an otherwise solipsistic process of symbol manipulation, yet Searle maintains that the fact that a modified version of his argument is applicable to the new circumstances shows that ‘‘no matter what outside causal impacts there are on the formal tokens, these are not by themselves sufficient to give the tokens any intentional content.’’ (1980, 522–523). There is, of course, a causal story that accounts for the having of intrinsically intentional states (or else, naturalism is plain false), but Searle denies that such a story has anything in particular to do with computer-like symbol manipulation (for more on that, see Section 3 below). In short, Searle believes, and he believes that his thought experiment shows, that as long as the computational picture is presupposed, there is no intrinsic intentionality, and that no supplementation in the form of causal transductions can compensate for this basic fact. 3. Part Two: Cognitivism, Content and the Impasse 3.1. THE BASIC IDEA OF MODERN COGNITIVE THEORY: SEMANTIC ENGINES AND THE PROBLEM OF CONTENT In the remaining parts of this paper, I would like to argue in support of Searle’s skeptical conclusion, and against the adequacy of RR. A natural thing to do is to go back to the basic notion underlying CTM, to what many refer to as ‘‘the basic idea of cognitive science’’ (Haugeland, 1981, 31), namely, the idea that intelligent beings are semantic engines (Dennett, 1978), to wit, ‘‘automatic formal systems with interpretations under which they consistently make sense’’ (Haugeland, 1981, 31). To understand this idea, we need to consider briefly the notions of a formal system and of an automatic formal system. Formal systems, whether axiomatic logico-deductive systems or rule-based games such as chess, are systems in which physical tokens are manipulated according to explicit rules. A crucial feature of formal systems is the fact that they are digital, namely (as Haugeland puts it), self-contained, perfectly definite, and finitely checkable.4 • Self-containment: The system is immune to external influence in the sense that only its own tokens, positions, and rule-governed moves make any difference to it, and that they do so only insofar as they exhibit those aspects that are relevant for legal (formally defined) state-transition of the system. COMPUTATION AND INTENTIONALITY 215 • Perfect definiteness: Bearing outright mistakes or breakdowns, there are no ambiguities, approximations, or ‘‘judgement calls’’ in determining what the position is or whether a certain move is legal. • Finite checkability (effective procedure): For each state and each candidate state transition, only a finite number of steps have to be checked to see whether this state-transition could be legal in that state. One of the implications of the fact that formal systems are digital in the above sense is that they are multiply realizable and intersubstitutable. Multiple realizability follows from the fact that what determines whether or not two tokens are tokens of the same (formal) type is whether or not they are treated equally by the rules for token (symbol) manipulation; in short, formally equivalent tokens are intersubstitutable, regardless of their extraneous, non-formal, properties. Formal equivalence, however, applies not only to tokens of formal systems, but to formal systems as wholes. Formal systems that are isomorphic, and thus, can be translated to one another, are equivalent, and, therefore, intersubstitutable (i.e., bearing practical considerations, the substitution of a formal system A for an equivalent system B makes no difference). The intersubstitutability of formal systems is of the utmost importance to CTM, since it is in virtue of this property that one formal system can model (or ‘‘mimic’’) another and, ultimately, that a universal Turing machine can model every formal system. This brings us to the notion of an automatic formal system – a computational machine or digital computer. An automatic formal system is a physical device that models a formal system, that is, which automatically manipulates its own physical tokens in structural isomorphism to that formal system. Can automatic formal systems be endowed with intrinsic intentionality? An obvious difficulty comes to mind. In standard formal systems (whether automatic or not), the manipulation of tokens is purely syntactic; it is based solely on the formal properties of symbol tokens and is insensitive to their semantic properties. It transpires, then, that to the extent that automatic formal systems, computational machines, are cognitive ‘‘engines’’ at all, they are syntactic, not semantic ‘‘engines’’ (cf. Fodor, 1981). How, then, can such systems operate as semantic engines? The standard reply is that even though the manipulation of tokens in formal systems is purely syntactic, it is nevertheless such that it may respect and preserve the semantic properties of these tokens qua interpreted symbols, and that this is all that is needed to ensure the ‘‘two lives’’ – syntactic and semantic – of such tokens. As we know from axiomatic formal deductive systems, tokens with semantic values can be the objects of purely formal manipulations that nevertheless respect their semantic identity. In other words, if the formal 216 I. SHANI manipulations are validly pursued, the corresponding semantic relations follow as a matter of course, or as Haugeland puts it: ‘‘if you take care of the syntax, the semantics takes care of itself’’ (1981, 23). Semantic engines, then, are syntactic engines plus interpretation – syntactically driven machines that ‘‘can handle meaning’’, as it were. Yet, at this joint we are facing another problem. CTM accounts for semantic content by recourse to the notion of ‘interpretation’, and ‘interpretation’, at least in the canonical use of the term, is a user-relative construct – a paradigmatic example of derived intentionality. Interpretations, whether of maps, texts, logical symbols or data structures, require interpreters. Since, ultimately, at the end of the line of this hermeneutic process there must be interpreters already equipped with intelligence, the problem of intrinsic (or original) intentionality is begged or regressed.5 In other words, interpretation-based semantics creates an unsolvable explanatory backlog for the simple fact that interpretation is a process of content transmission, not of content generation. In standard home computers, for example, it is we as interpreters, programmers, and users who transform the systems’ syntactically driven operations into semantically meaningful operations. The system itself does not really ‘‘handle meaning’’ by itself and for itself. Advocates of CTM are usually aware of this fact, yet the standard response to the problem is that the problem of intrinsic intentionality may be solved if something very much like an interpretation can be secured naturally – without the mediation of intelligent agents. Roughly, the idea is that the problem of intrinsic intentionality can be solved if it can be shown that, by dint of natural processes alone, formal automata can be constructed that ‘‘mimic’’ artificially constructed semantic engines in the sense that (a) They are syntactically driven computational machines. (b) The symbol tokens on which the computations are performed are mapped onto the environment (and perhaps also onto each other) in a way that parallels the mapping relations that result from the assignment of semantic values in standard interpretations (thereby bestowing semantic values onto the symbol tokens). And, (c) There is a correlation between the syntactically driven operations the system undergoes and rational relations among the meanings of the symbolic structures on which computations are performed (Fodor, 1981; Block, 1993). In short, cognitivist theories of content deal with the problem of the observer-dependency of interpretation by assuming that the problem can be solved if we eliminate the interpreters from the picture and let nature take care of the interpretation. COMPUTATION AND INTENTIONALITY 217 I would like to conclude by arguing that this strategy is ill fated, and that it is ill equipped to deal with the problem of intrinsic intentionality. While the import of the argument to be presented converges remarkably with Searle’s argument, it illuminates the problem from a somewhat different angle, thereby providing additional support to Searle’s provocative conclusion. 3.2. ON THE INCOHERENCE OF ENCODING-BASED REPRESENTATION A crucial step in the ‘‘interpretation without interpreters’’ solution to the problem of intrinsic intentionality is the assumption that (some) naturally constructed mapping relations are content-constitutive (see clause (b) above). Namely, the assumption is that it is in virtue of ‘‘linkages between symbols that the device manipulates and things in the world’’ (as Fodor, (1980, 520) puts it) that mental states become endowed with content. The underlying idea, then, is that what accounts for the intrinsic intentionality of genuine intelligent beings is a relation of correspondence, or encoding, between internal symbol structures and external conditions. Like interpretation, however, ‘encoding’ – at least in the canonical use of the term – is an exemplar of derived intentionality. Paradigmatically (for example, in communication theory), encodings are symbols the meaning of which is made by convention to correspond to (stand for) other symbols whose meaning is already fixed. Thus, Morse codes, Braille characters, or binary computer digits are all systems of symbolization into which the previously established meanings of natural language (e.g., English) symbol structures can be transposed – often with great practical yield. For example, the Morse code ‘‘...’’ is a stand-in for the character ‘‘S’’, and it is useful because, unlike characters and numerals, dots and dashes can be sent over telegraph wires.6 But the dots and dashes are only meaningful due to the fact that they borrow their meaning from the already fixed meanings of the characters and numerals they encode. Such use of the term ‘encoding’ is non-problematic as far as it goes, but it is also irrelevant for the purpose of explicating intrinsic content. Conventional encodings do not generate new representational contents: they are merely transformations of already existing contents into new representational forms (and, often, new physical media), and they presuppose interpreters. By contrast, the processes that constitute the intentionality of intrinsic mental states must be content generating. On pain of regress they cannot rely on prior meanings, and cannot presuppose the intervention, or authority, of intelligent interpreters. Nevertheless, the accepted wisdom in cognitivist theories of content is that something much akin to standard encodings is precisely what gives intrinsic representational structures their meaning. Recall that what makes standard 218 I. SHANI encodings work is the fact that they are made to correspond, in a reliably controlled manner, to meaning-laden symbols. In a strikingly similar fashion, it is believed that the problem of explaining intrinsic intentionality can be solved if it can be shown that basic-level (original) representational structures correspond in a definite, unaided, natural way (the ‘‘right’’ way) to specific external conditions. The implicit assumption, then, is that what really matters is the existence of some sort of correspondence relations, whereby the external correspondents are ‘‘specified’’. This, the specification, is what gives representational structures their content. Since such specification does not require that the specified objects are themselves meaning-laden symbols, and since it can obtain by dint of natural processes alone, that is, without presupposing the intervention of interpreters (designers, users), it is, the idea goes, naturalistically adequate. From this perspective, the disanalogies with standard encoding, the fact that the ‘‘encoded’’ external conditions are not themselves (at any rate, need not be) meaning-laden symbols, and the fact that the interpreters are eliminated, may be seen as unproblematic, advantageous modifications. After all, these modifications are necessary in order to meet naturalistic standards of explanation, and, in any case, they are inessential insofar as the epistemic function of specification, the ‘‘true essence’’ of encoding, is concerned. Thus, representation is reconstructed as a natural ‘‘encoding’’ process, a naturally designed correspondence, as it were. Such correspondence retains the structure of standard encodings in the sense that they both presuppose covariance between the encoding and the encoded elements. At the same time, the interpreters are eliminated from the scene, leaving only natural processes to bear on the covariance. In short, the assumption is that nature can ‘‘take care of the situation’’ by creating correspondence relations semi-analogous to artificial encodings, and that the real challenge is to find how this is being done, namely, to find the ‘‘right’’ kind of correspondence (recall the citation from Fodor in Section 2.3). As Mark Bickhard has argued extensively, however (e.g., 1980, 1993, 1996, 2000, 2003; Bickhard and Terveen, 1995; see also Edelman and Tononi, 2000; Shannon 1993 for some converging arguments), the idea that such relations of natural correspondence are constitutive of intrinsic intentionality leads to an epistemic, or theoretical, impasse. The problem, Bickhard argues, is that the information that such correspondence relations are taken to generate is senseless once appeals to the epistemic authority of interpreters is banned and the first person perspective of the psychological agent itself is considered on its own epistemic terms. Mental states, he says, do not wear their denotational values on their sleeves, and the fact that they correspond to external sources does not explain how knowledge of the correspondence, knowledge of the encoded elements, is attained at the ‘‘destination’’ (i.e., by the agent). Consequently, mental states cannot be intrinsically informative, COMPUTATION AND INTENTIONALITY 219 cannot serve their owners as representations, in virtue of the fact that they ‘‘encode’’ external conditions.7 Bickhard’s argument can be presented in two versions, each of which illustrates the same problem, but with different emphases. Epistemic vacuity (from the first-person perspective, symbol-object correspondences are epistemically vacuous): Conventional correspondences are informative for those who are familiar with the convention since the latter have access to both sides of the correspondence (and to the rules that connect them together). By contrast, when a mental state R is said to represent an external condition F in virtue of the fact that it corresponds to it in some naturally constructed fashion, the system – call it S – that is supposed to make use of R has no equally available access to both sides of the correspondence. Rather, all S has to work with are its own mental states. S can only access F via R (or some other mental states); it cannot observe the correspondence from both ends and use the fact that it obtains as an independent source of knowledge. In the absence of such knowledge, however, the situation is analogous to having an access to the symbol string ‘‘-.’’ without knowing that it is the Morse code correspondent of ‘‘N’’: no such knowledge can be miraculously gained merely in virtue of the fact that the correspondence obtains. To put it differently, if correspondence relations were content-constitutive, then the mere fact that the correspondence obtains ought to have been sufficient for making R a representation of F; but, as the Morse code example suggests, from the first-person perspective, this fact, considered in itself, is epistemically vacuous – it does nothing by way of informing S what’s on the other end of the presumed correspondence.8 If the fact that R corresponds to F is epistemically vacuous, however, if it yields no intrinsically available information to the effect that it is F that R stands for, then it cannot be taken as constitutive of making R a representation of F. For, if R is, indeed, a representation of F, then, whatever else it might be, it must be capable of being used (used by S itself, that is) as a representation of the sort. But, what Bickhard’s argument shows, I think, is that the one thing that the postulation of a symbol-world correspondence relation does not explain is how any symbol could function, intrinsically, as a representation in virtue of the fact that it stands in such a correspondence relation to some external item. In short, the fact that inner structures may encode external conditions does not, in itself, explain how such structures can come to serve their owners as representations and, consequently, it offers no satisfactory solution to the problem of intrinsic intentionality.9 Extrinsicality (the property of ‘corresponding to...’ is an extrinsic property of mental states and, as such, it is, again, epistemically vacuous): The fact that R encodes F (that it correlates with, or is caused by, F) is an extrinsic fact about R in the sense that it makes no difference to the internal causal structure of the representation. That is, in principle, R could have been 220 I. SHANI exactly the same even if instead of F it corresponded to F¢, or even to nothing at all (for more detail, see Bickhard, 2003). If R’s content makes no difference to its internal causal structure, however, then the fact that R possesses this particular content has no unique impact on the manner in which it interacts with other mental structures (R¢, R¢¢ ...) in S’s cognitive economy. But then the difference between R  F and, say, R  F¢ (where ‘’ stands for ‘corresponds to’) is not a difference that can be detected elsewhere in the system, and hence not a difference that makes a difference to the ongoing flow of the system’s cognitive activity.10 In other words, from the first-person perspective of the cognitive agent itself, the encoded value of R is simply nontransparent. S cannot learn that R encodes F, nor can it use R as a representation of F, merely by virtue of the fact that the encoding relation obtains (i.e., that R carries information – in the technical sense used in communication theory – about F, or that it is caused by F). In short, since encoding is an extrinsic (or, external) relation, the fact that R encodes F does not explain how, relying on nothing but its own cognitive resources, the system itself, S, can use R as a representation of F. The upshot of both versions, then, is that mental structures cannot function as representations, cannot be intrinsically informative, in virtue of the fact that they encode whatever it is they encode. True, the elimination of interpreters from the scene renders the encoding process naturalistically legitimate, but, as Bickhard argues, this is being done at the price of epistemic coherence. The only coherent sense in which encodings can be said to represent is observer-dependent, and eliminating the observer does not yield a coherent account of intrinsic intentionality – rather, it yields the total elimination of all forms of content, derived or intrinsic. 4. Recapitulations, a Conclusion, and Some Afterthoughts Bickhard’s argument against encoding-based representation converges on, and provides additional support to, Searle’s contention that RR offers no adequate response to the original charge of the Chinese room argument, namely, to the claim that CTM yields a notion of mental content that is inherently observer-dependent. As mentioned before, the Chinese room argument was devised as a refutation of the view that appropriately programmed digital computers literally are the possessors of genuine mental states. The robot’s reply concedes this point but, at the same time, attempts to sustain a close relative of the position targeted in Searle’s original argument. It consists of the idea that intrinsic intentionality is, or at any rate can be, constituted by appropriately programmed digital computers whose internal symbol structures appropriately encode external conditions. COMPUTATION AND INTENTIONALITY 221 As we have seen, Searle’s main objection to RR is that it offers no real improvement over the problematic situation illustrated by the original Chinese room argument. He argues that the same old argument can be adapted, with little modification and with equal force, to the new, robot, circumstances; and the moral, he reasons, is that computation plus symbol-world linkages is no more intrinsically intentional than computation without such linkages. All the reply does is to add blind (non-intelligent) causal linkages to blind computations, and nothing in the addition of such linkages can compensate for the epistemic (cognitive, intentional) vacuity of the original situation. But one person’s modus ponens is another’s modus tollens. Challenging Searle’s intuition, Fodor considers the formula ‘‘cognition equals computation plus causal linkage’’ to be firmly intact and he concludes, instead, that there must be something wrong with Searle’s example. No machine table instantiation constitutive of intrinsic intentionality, he argues, employs a little man to bring about its state transitions. Yet, as Searle points out (see Section 2.3 above), it is unclear just on what ground conscious implementation ought to be universally excluded. In an attempt to dissolve Searle’s argument, Fodor claims that the machine statetransitions at the instantiation level must consist of effective (immediate, proximal) causes only (1980, 525), yet he offers no argument just why this must be the case. While intuitions may diverge when it comes to the dispute between Searle and Fodor and the implications it bears for the status of RR, Bickhard’s argument against encoding-based representation illuminates Searle’s position from a somewhat different angle and provides powerful reasons to side with Searle. What Bickhard’s argument shows, if correct, is that encoding is not, and cannot be, constitutive of intrinsic content. In other words, it is not in virtue of encoding that an otherwise non-intentional system can become intentional. Since RR concedes that computation alone does not generate cognition (intrinsic intentionality), and since it identifies the missing piece in the puzzle with an ‘‘appropriate’’ relation of encoding (symbol-world linkage), Bickhard’s argument is, in effect, an indirect assault on the adequacy of the reply. Recall that by applying his thought experiment to the circumstances decreed by RR, and by illustrating his own total lack of comprehension of the meaning of the ‘‘information’’ he gains under these new circumstances, Searle attempts to show the futility of the reply. By arguing against the possibility of content-generating encodings, Bickhard’s argument converges onto the same conclusion. The imports of the arguments are remarkably similar. In fact, in both cases, the conclusion is identical: Symbol-world linkages do not generate content and cannot fill the theoretical void in the explanation of intentionality as a 222 I. SHANI system-intrinsic phenomenon. Even the language is sometimes similar. Consider the following passage from Searle’s reply to Fodor. [l]et the egg foo yung symbol be causally connected to egg foo yung in any way you like, that connection by itself will never enable the agent to interpret the symbol as meaning egg foo yung. To do that he would have to have, for example, some awareness of the causal relation between the symbol and the referent; but now we are no longer explaining intentionality in terms of symbols and causes but in terms of symbols, causes, and intentionality... (1980b, 522). Compare Bickhard The basic point here is simply that informational correspondence, even should it exist, does not announce on its sleeve that it exists, or what it is in correspondence with. Some state or event in a brain or machine that is in informational correspondence with something in the world must in addition have content about what that correspondence is with in order to function as a representation for that system. (2000, 68) Searle and Bickhard agree, then, that the kind of information that the ‘‘egg foo yung’’ symbol is said to carry is just not the kind of information that can be appreciated from the first-person perspective of a psychological agent.11 Where they differ is in the way they arrive at their conclusions. Searle’s conclusion is based, primarily, on his persuasion that the revised version of the Chinese room argument shows that the allegedly content-constitutive relations between the symbol ‘‘egg foo yung’’ and the object egg foo yung fail to generate any understanding of the meaning of the symbol. Bickhard arrives at the same conclusion by arguing that the fact that the symbol encodes the object does nothing by way of explaining how a psychological subject could use the former as a representation of the latter. Where Searle demonstrates his point phenomenologically (by imagining himself performing the computations and by observing that he does not understand a thing), Bickhard bases his argument on functional considerations (by showing that there is nothing in encoding per se to guarantee the capacity of mental states to inform their owners of the world around them); but these are simply alternative ways of making the same point. If symbol-world correspondences are incapable of generating content, then, not only does RR fail, but, as was mentioned before, the entire project of explaining intentionality in cognitivist (computational, information-processing) terms seems to be facing a blind alley. The CTM is premised on a categorical dichotomy between symbol manipulation and symbol significance (between form and content, syntax and COMPUTATION AND INTENTIONALITY 223 semantics, information processing and information pickup, computation and interpretation, etc.). The logic of the dichotomy urges a piecemeal approach in which the task of providing an account of mental content can be sharply distinct from, yet at the same time be supplementary to, the task of accounting for mental (cognitive) processing. CTM being the dominant theory of cognitive ‘‘architecture’’ (that is, of mental processing), it was only natural that attempts to explain intentionality, to naturalize content, were equally dominated by this piecemeal, complementary approach. An encoding-based approach to the explanation of content fits perfectly well in this picture, since it respects the dichotomy between content and process and, in particular, is compatible with CTM. Yet, if Searle and Bickhard are right, the process-content dualism inherent in ‘‘the basic idea of modern cognitive theory’’, rather than being the solution to the problem of intentionality, is a veritable source of trouble. The assumption that mental processing is a purely formal procedure of symbol-manipulation implies that mental processing, qua mental processing, has nothing to do with content. Computation, in other words, is nonintentional, as the original Chinese room argument so vividly illustrates. Similarly, the assumption that mental content is purely a matter of symbolworld linkage implies that, qua content, it has nothing to do with mental processing. Now, both Searle’s criticism of the robot’s reply and Bickhard’s argument against encoding-based representation demonstrate that the same problem – lack of intentionality – arises again. The existence of symbol-world linkage is not a property that the cognitive system itself can appreciate, from its own perspective, by interacting with the symbol and therefore not a property that can constitute intrinsic content (cf. Edelman and Tononi, 2000, chapter 11). Indeed, given the assumption that mental content is extrinsic to mental processing, we need not be surprised at this failure to explicate intentionality (content) as a system-intrinsic phenomenon! Either way, then, the system itself, the psychological agent, is left without intentionality. The only sort of intentionality that CTM sustains, Bickhard and Searle argue, is derived intentionality – the ascription of content made from the third-person perspective of programmers, users, observers, and interpreters.12 As mentioned in Section 2.1, standard (artificially constructed) digital computers are semantic processors only because they are given semantic interpretation. The underlying idea of cognitivism, however, is that computational machines can come to possess intrinsic intentionality by virtue of processes that mimic the assignment of content by way of an interpretation in that they specify mapping relations that parallel the mapping relations that result from interpretation. The problem with this strategy, as I think both Searle’s and Bickhard’s arguments show, is that the structure of such semantic allocation retains its original, observer-dependent character. The naturally constituted mapping relations that the theory envisions do not 224 I. SHANI generate intrinsic content; they are merely such that they could make (epistemic) sense to an interpreter if only there were one around. It transpires, then, that there is no coherent reconciliation between computation and intentionality (or between CTM and the problem of intrinsic content) and that the foundational crisis pointed out by Searle almost a quarter of a century ago is very much alive. 5. Afterthoughts 5.1. ON THE AD HOMINEM, AND IGNORANCE, FALLACIES OF THE ‘‘SO, WHAT DO YOU SUGGEST INSTEAD?’’ OBJECTION One of the most puzzling features of the controversy that surrounds the Chinese room argument is the fact that so many able people tend to behave as if Searle’s failure to provide a convincing alternative to the computational story somehow implies that his criticism need not be taken seriously. To be sure, it would be nice to have a viable positive account of intrinsic intentionality. But the cogency of Searle’s critique of cognitivism in no way depends on the availability of a ready-made alternative, and the shortcomings of his own account of intentionality (developed in Searle, 1983, 1992) give no reason to dismiss what he has to say. Clearly, the reasoning behind such rhetorical device involves an ad hominem fallacy, but it also involves an interesting fallacy of ignorance: Since my opponent has not shown that his recipe for intentionality is correct, or that he has one, my own recipe remains intact. This being said, it may be mentioned that, unlike Searle, Bickhard has developed, over the years, a powerful systematic alternative to CTM and to encoding-based models of representation. Bickhard labels his alternative to the computational approach ‘interactivism’. This is not the time and place to explore interactivism in detail (the interested reader may consult Bickhard, 1980, 1982, 1993, 1996, 2000, 2003; Bickhard and Richie, 1983; Bickhard and Terveen, 1995; Christensen and Hooker, 2001). Suffice it to mention here that (a) The theory strives to explain representation on the basis of the notion of self-organization in far-from-equilibrium thermodynamics and is a pioneering, and one of the theoretically most developed, accounts framed within the emerging dynamic framework in the cognitive sciences. (b) It explains mental content without presupposing, neither computational mental processing, nor an encoding-based account of meaning. Thus, aligning Bickhard’s critique of encoding models of representation with Searle’s critique of CTM may also yield the advantageous side product COMPUTATION AND INTENTIONALITY 225 of giving no pretext to an immature dismissal of Searle’s argument on pain of there being no alternative to the formula ‘cognition equals computation plus symbol-world linkage’. 5.2. ON THE FALLACY OF EQUATING FIRST-PERSON SIGNIFICANCE WITH CONSCIOUS APPREHENSION Discussions of the problem of intrinsic intentionality are frequently burdened with the inculcated false impression that first-person intentionality is necessarily (that is, in all its manifestations) conscious. Often, what nourishes this commonly held assumption is the fact that the debate between Searle and his critics is marked by Searle’s own emphasis on conscious processes versus the tendency of his cognitive adversaries to reply by pointing to the prevalence of non-conscious cognitive processes. On a self-organization-based account such as Bickhard’s, however, the requirement for first-person significance is not tantamount to a presupposition of conscious accessibility (for a similar, highly detailed, view see also Damasio, 1994, 1999). Rather, it is a requirement that the information will function as information for the agent, namely, that it will take an integral part in the agent’s self-organizing (or self-maintaining) action flow – consciously or not. On this view, consciousness is nested in a larger embodied context that is already perspectival; to wit, the functional self is broader than the conscious self is. Thus, the first-person perspective begins with the functional, homeostatic, ‘‘infrastructure’’ of the embodied agent, which predates consciousness, and, consequently, the reduction of this perspective to something that entails conscious experience is fallacious. When the possibility that intrinsic intentionality need not be conscious is acknowledged, it becomes clear that the fact that many, perhaps most, of the processes that CTM strives to explain are non-conscious (informationally encapsulated, cognitively impenetrable, or however you prefer to put it) cannot be advanced as a counterargument to the charge that the theory suffers from a chronic disability to make good of intrinsic intentionality. Simply put, if all cognitive processes, conscious or not, are perspectival, then the fact that advocates of the computational approach are typically preoccupied with non-conscious cognitive processes does not imply that they need not worry about first-person significance. Acknowledgements I would like to thank Ausonio Marras, Mark Bickhard, Robert Stainton, Liam Dempsey, Marnina Norris, Daniel Vaillancourt and Zara McDoom, John Kearns and an anonymous referee at Minds and Machines. 226 I. SHANI Notes 1 Consider, in this context, the following quotations from Harnad and from Hayes. ‘‘Searle’s celebrated Chinese Room Argument has shaken the foundations of Artificial Inteligence. Many refutations have been attempted, but none seem convincing.’’ (Harnad, 1989, 1). ‘‘My own notion of what constitutes [the core of cognitive science] could be summed up in the following way: it consists of a careful and detailed explanation of what’s really silly about Searle’s Chinese room argument’’ (Hayes, p. 2 in Lucas and Hayes, 1982). 2 Searle’s original exposition of the Chinese room argument is modeled after a machine designed by Schank and his co-workers for the purpose of understanding a story (Schank and Abelson, 1977). 3 It is worth mentioning that to insist on the intrinsicality of mental content is not tantamount to espousing internalism (though this is a common mistake; see for example, Dennett, 1987, chap. 8). Granted that representation is a relational phenomenon that involves a relation to (a portion of) the environment, there is still an internal component to it that is unique to the subject – my representational states are internal states had be me, and not by someone else. It is my representations, which function as representations for me, whereby I become aware of the world around me and that are causally efficacious in my thought processes and in my actions; and it is in this sense that they are intrinsic to me, as opposed to being intrinsic to someone else, or to be ascribed to me from someone else’s perspective (cf. Searle, 1992, 78–82). 4 There are problems with, and disagreements about, the correct technical characterization of ‘digital system’ and ‘digital representation’ (see for example, Demopolous, 1987; Pylyshyn, 1984, 200–202), but none of them bears heavily on the issues I discuss below. 5 As a referee of the paper has kindly pointed out, it could be the case that one formal system, say a Lisp interpreter, interprets another formal system, say a Lisp program. It follows, then, that not every interpreter has to be intelligent (i.e., to possess intentionality). However, a Lisp interpreter is only an interpreter to the effect that it is assigned this role by a human observer. In other words, it is only an interpreter in a derived sense: though not intelligent in itself, it presupposes an intelligent human interpreter capable of taking what it does to be an interpretation – hence the circularity problem. 6 The example is borrowed from Bickhard (1996). 7 The argument presented here against encoding is, in fact, only one among a host of arguments Bickhard has developed over the years. I chose to focus on this particular argument, however, because it seems to be one of the more compelling ways of demonstrating the problem and because it converges nicely with Searle’s rebuttal to the robot reply. 8 In this context, it is instructive to mention an old argument by Piaget against what he called ‘‘copy’’ theories of Knowledge. Compare: ‘‘ [I] find myself opposed to the view of knowledge as a copy, a passive copy, of reality. In point of fact, this notion is based on a vicious circle: in order to make a copy we have to know the model that we are copying, but according to this theory of knowledge the only way to know the model is by copying it, until we are caught in a circle, unable ever to know whether our copy of the model is like the model or not.’’ (1970, 15). 9 It is worth emphasizing that the point made here about the inability of encoding relations to constitute intrinsically functioning representations have nothing to do with the question whether or not the representation in question is being used consciously, or explicitly. Rather, if the argument applies at all, it applies with equal force to non-conscious, or implicit, use of mental representations (see also Section 5.2). 10 For an extended analysis of this problem of internal communication see Edelman and Tononi (2000, chap. 11). COMPUTATION AND INTENTIONALITY 227 11 In this context, it is useful to distinguish between two notions of information that are often being conflated: one, information in the technical sense of communication theory, and another, which roughly parallels (or at any rate implies) propositional content and is therefore necessarily semantic. Sayre (1987) suggests the convention info(t) and info(s) to distinguish the technical from the semantic sense (see also Bickhard, 2000, Bunge, 1974, Searle, 1980). Using this convention, we can express Searle’s and Bickhard’s criticism in the following way: symbolworld linkages may be sufficient for info(t), but (information semantics notwithstanding) they do not constitute info(s). 12 Notably, a very similar conclusion is also reached by Dennett (e.g., 1987) implied as it is by his influential theory of the intentional stance. There is, however, an important difference too. Unlike Searle and Bickhard, who are manifest intentional realists, Dennett is an instrumentalist who believes that there simply is no such thing as intrinsic intentionality. Consequently, Dennett is not disturbed a tiny bit by the conclusion that CTM entails the absence of intrinsic intentionality. As I have argued in Section 2.2, however, this result should be worrisome to a large squad of philosophers who are at once realists and computationalists (e.g., Block, Dretske, Fodor and Millikan). References Abelson, R.P. (1980), Searle’s Argument is Just a Set of Chinese Symbols, Behavioral and Brain Sciences 3, pp. 424–425. Bickhard, M.H. (1980), Cognition, Convention and Communication, NY: Praeger. Bickhard, M.H. (1982), Automata Theory, Artificial Intelligence, and Genetic Epistemology, Revue Internationale de Philosophie 36, pp. 549–566. Bickhard, M.H. (1993), Representational Content in Humans and Machines, Journal of Experimental and Theoretical Artificial Intelligence 5, pp. 285–333. Bickhard, M.H. (1996), Troubles with Computationalism, in W. O’Donohue and R.F. Kitchener, eds., The Philosophy of Psychology, London: Sage, pp. 173–183. Bickhard, M.H. (2000), Information and Representation in Autonomous Agents, Journal of Cognitive System Research 1, pp. 65–75. Bickhard, M.H. (2003), Process and Emergence: Normative Function and Representation, in J. Seibt, ed., Process Theories: Crossdisciplinary Studies in Dynamic Categories, Dordrecht: Kluwer Academic. Bickhard, M.H. and Richie, D.M. (1983), On the Nature of Representation: A Case Study of James Gibson’s Theory of Perception, NY: Praeger. Bickhard, M.H. and Terveen, L. (1995), Foundational Issues in Artificial Intelligence and Cognitive Science, Amsterdam: Elsevier. Block, N. (1980), What Intuitions about Homonculi Don’t Show, Behavioral and Brain Sciences 3, pp. 425–426. Block, N. (1990), Can the Mind Change the World?, in G. Boolos, ed., Meaning and Method: Essays in Honor of Hilary Putnam, Cambridge: Cambridge University Press, pp. 137–170. Block, N. (1993), The Computer Model of the Mind, in A.I. Goldman, ed., Readings in Philosophy and Cognitive Science., Cambridge MA: MIT Press, pp. 818–831. Bunge, M.A. (1974), Treatise on Basic Philosophy Sense and Reference Vol. 1 Dordrecht: Reidel. Christensen, W.D. and Hooker, C.A. (2001), ‘Self-Directed Agents’, in J. McIntosh, ed., Naturalism, Evolution and Intentionality. Canadian Journal of Philosophy, Special Supplementary Volume. 228 I. SHANI Damasio, A.R. (1994), Descarte’s Error: Emotion, Reason and the Human Brain, NY: GrossetPutnam. Damasio, A.R. (1999), The Feeling of What Happens: Body and Emotion in the Making of Consciousness, NY: Harcourt Brace. Demopolous, W. (1987), On Some Fundamental Distinctions of Computationalism, Synthese 70, pp. 79–96. Dennett, D.C. (1978), Brainstorms: Philosophical Essays on Mind and Psychology, Montgomery VT: Bradford Books. Dennett, D.C. (1987), The Intentional Stance, Cambridge MA: MIT. Dretske, F. (1988), Explaining Behavior: Reasons in a World of Causes, Cambridge, MA: MIT/ Bradford. Edelman, G.M. and Tononi, G. (2000), Consciousness: How Matter Becomes Imagination, London: Penguin Books. Fodor, J.A. (1975), The Language of Thought, Hassocs Sussex/Scranton PA: Harvester Press/ Crowell. Fodor, J.A. (1980), Searle on What Only Brains Can Do. Open Peer Commentary. Behavioral and Brain Sciences, 3. Reprinted in D. Rosenthal (ed.). Fodor, J.A. (1981), Representations, Montgomery: Bradford. Harnad, S. (1989), Minds Machines and Searle, Journal of Theoretical and Experimental Artificial Intelligence 1, pp. 5–25. Harnad, S. (1990), The Symbol Grounding Problem, Physica D 42, pp. 335–346. J. Haugeland (ed.) (1981), Mind Design, Cambridge MA: MIT. Hayes, P.J., Harnad, S., Perlis, D. and Block, N. (1992), Virtual Symposium on Virtual Mind, Minds and Machines 2(3), pp. 217–238. Lucas M.M. Hayes P.J. (ed.) (1982), Proceedings of the Cognitive Curricula Conference, Rochester NY: University of Rochester. Millikan, R.G. (1989), Biosemantics, Journal of Philosophy 86(6), pp. 281–297. Piaget, J. (1970), Genetic Epistemology, NY: Columbia. Pylyshyn, Z. (1987), Computation and Cognition: Toward a Foundation for Cognitive Science, Cambridge MA: Bradford/MIT. Rosenthal, D. (1991), The Nature of Mind, Oxford: Oxford University Press. Sayre, K. (1987), Cognitive Science and the Problem of Semantic Content, Synthese 70, pp. 247–269. Searle, J.R. (1980), Minds, Brains, and Programs, Behavioral and Brain Sciences 3, pp. 417– 424. Searle, J.R. (1980), From Author Response. Behavioral and Brain Sciences, 3. Reprinted in D. Rosenthal (ed.). Searle, J.R. (1983), Intentionality: An Essay in the Philosophy of Mind, Cambridge: Cambridge University Press. Searle, J.R. (1992), The Rediscovery of Mind, Cambridge MA: MIT/Bradford. Schank, R.C. and Abelson, R.P. (1977), Scripts, Plans, Goals and Understanding, Hillsdale, N.J.: Laurence Erlbaum Associates. Shannon, B. (1993), The Representational and the Presentational Hertfordshire, England: Havester Wheatsheaf.