ETHICA EX MACHINA Exploring Artificial
ETHICA EX MACHINA Exploring Artificial
ETHICA EX MACHINA Exploring Artificial
Exploring
artificial moral agency or the possibility of
computable ethics
Rodrigo Sanz
ISSN 2522-0063
ZEMO
DOI 10.1007/s42048-020-00064-6
1 23
Your article is protected by copyright and
all rights are held exclusively by Springer-
Verlag GmbH Deutschland, ein Teil von
Springer Nature. This e-offprint is for personal
use only and shall not be self-archived
in electronic repositories. If you wish to
self-archive your article, please use the
accepted manuscript version for posting on
your own website. You may further deposit
the accepted manuscript version in any
repository, provided it is only made publicly
available 12 months after official publication
or later and provided acknowledgement is
given to the original source of publication
and a link is inserted to the published article
on Springer's website. The link must be
accompanied by the following text: "The final
publication is available at link.springer.com”.
1 23
Author's personal copy
ZEMO
https://doi.org/10.1007/s42048-020-00064-6
SUBMITTED PAPER
Rodrigo Sanz
Abstract Since the automation revolution of our technological era, diverse machines
or robots have gradually begun to reconfigure our lives. With this expansion, it seems
that those machines are now faced with a new challenge: more autonomous decision-
making involving life or death consequences. This paper explores the philosophical
possibility of artificial moral agency through the following question: could a machine
obtain the cognitive capacities needed to be a moral agent? In this regard, I propose
to expose, under a normative-cognitive perspective, the minimum criteria through
which we could recognize an artificial entity as a genuine moral entity. Although my
proposal should be considered from a reasonable level of abstraction, I will critically
analyze and identify how an artificial agent could integrate those cognitive features.
Finally, I intend to discuss their limitations or possibilities.
R. Sanz ()
Facultad de Humanidades (FHUCE), University of the Republic of Uruguay (UdelaR), Montevideo,
Uruguay
E-Mail: contact@rodrigosanz.com
Author's personal copy
R. Sanz
1 Introduction
Since the automation revolution of our technological era, diverse machines or robots1
have gradually begun to reconfigure our lives. Nowadays, no longer restricted to
factories’ mechanized process, they are quickly turning to new fields: social health
or commercial services, stock market management, manoeuvres of search and rescue,
entertainment, social accompaniment, transport (self-driving cars or trains), labor or
military matters. With this expansion, it seems that those machines are now faced
with a new challenge: more autonomous decision-making involving life or death
consequences. As machines continue to increase in capacity and autonomy, human-
machine interaction will be implicated by moral circumstances, inevitably. Hence,
how consistently have we thought about machine’s moral behavior?
One of the most difficult challenges in the new field of “Machine Ethics” (Allen
et al. 2000, 2006; Anderson and Anderson 2011; Sullins 2006) is to confront the
moral dilemmas we have been trying to answer for over 2500 years of Western
philosophical thought. Thinking about this matter has been, and still is, a very
difficult job, enflaming diverse perspectives or discrepancies about their fundamental
foundations: do moral entities exist? If so, what kind of entities are we talking about
or how do we get knowledge from them?
In fact, the same occurs with the exploration of a possible ethical theory for
robots: what challenges or limitations are imposed since the development of an
intelligence called “artificial”? Is a computational model of ethics really viable?
Could a machine obtain the cognitive capacities needed to be a moral agent?
To answer these matters, I will expose a moral philosophical perspective focused
on agency (on the cognitive characteristics of the subjects who are behind a moral
action). At first, I would bring to light when we, humans, are morally considered.
I will focus on a normative ethical perspective to identify when an artificial agent
could be recognized as a moral entity disassociated from any biological necessity.
In other words, to find the necessary and sufficient conditions in order to establish
the existence of an artificial moral agent (AMA). Thereafter, I will discuss if this
model could be implemented in the robotics realm.
In the last decades, diverse opinions have been raised for the projection of an
AMA. On one side, some thinkers believe robots will never have the faculty of
being morals, neither today nor in the future (Bringsjord 2008). Others, conversely,
1 “Robot” is a term commonly used today and spread mostly by science fiction literature. We are quite
informal about how we use the word “robot”. Its etymological origin is Czech and derives from the word
“robota” (labor) but could also be related to the term “rob” (slave) of the ancient Slavic. The invention
of the term is attributed to Karel Capek in his play R.U.R. from 1921. Thus, the term can be applied to
a variety of artefacts: android, automaton, machine, among others. In this article, I will use indifferently
“robot” or “machine” to refer to any software packaged in a hardware with a certain degree of artificial
intelligence (AI).
Author's personal copy
ETHICA EX MACHINA. Exploring artificial moral agency or the possibility of computable...
consider this could be a fact even though robots are not moral agents yet (Dennett
1997). Nevertheless, skeptical pessimism aside, many people speculate their real
emergence (Moor 2006; Wallach and Allen 2008; Floridi and Sanders 2004).
Technology, in some sense, has always been normative. Artefacts are made for
different purposes and evaluated according to how well they perform their pre-
assigned tasks; they are conceived taking into account their repercussions or ethical
consequences. My goal here is not to evaluate an AMA by designed standards, but
to examine the chances he2 has to acquire some ethical and moral principles by
himself. Now, determining which model could be implemented within an AMA in
order to adjust to the variety of moral and social practices becomes crucial.
Certainly, the idea that a moral decision could be a matter of calculation is not new,
in fact, it had a profound influence on the development of some traditional ethical
theories3. Until now, engineers have handled ethics strictly as an additional set of
restrictions (as other software condition). Yet, over the last decades, new studies
have emerged on the subject. In 2005, a group of researchers (Anderson and Armen
2005) started a “utilitarian robotic project”. Their first artificial agent, Jeremy (after
Jeremy Bentham), deployed the “Hedonist Law of Utilitarianism” and calculates
the probability of pleasure or displeasure caused on people affected by a particular
action. The “right action” would be the one resulting from the best consequence
amongst all possible options. But, was it really a good approach? This type of
“moral mathematics” omits the strong link between ethics and epistemology, an
oversight that brings enormous problems. Imagine the number of infinite potential
consequences to be considered. Utilitarianism will be only partially computable due
to the limitations of the amount of actions or the number of possible consequences
that a machine can compute in order not to get stuck in an infinite process. That
brings down the basics of this theory. Also, as pointed out by critics of utilitarianism,
sacrificing a person’s life for a major good is not always the best option. This type
of ethical model would permanently incur in unsolvable dilemmas like the famous
“trolley problem” of Philippa Foot.
As an alternative, other researchers propose to install “deontological models”
(Wiegel 2006) into the machines, but again this perspective seems to forget the
problematic consequences of actions followed by immovable rules in life. The model
falls into paradoxes that are not easy to unravel. When a machine has specific
2 Due to the subtle philosophical notion of “person”, I will continue to use ‘he’ as a gender-neutral pronoun
Manner of Computing the Morality of Actions” (Hutcheson 1753). Hutcheson’s theory is a precursor to
a type of ethical theory that thrives in calculus: utilitarianism.
Author's personal copy
R. Sanz
restrictions like Isaac Asimov’s “Three Laws of Robotics” (for instance, “not injuring
a human being”), given the scenario where it is only possible to save one person of
a group of humans, who should be saved? A computable ethical approach based on
fixed and strict rules will always be incomplete due to the struggle to anticipate all
potential situations that may happen. Creating a moral robot is not merely finding
the number of restrictions or right formulas to resolve conflicts, this would be
a “restricted” morality and an artificial agent would result doomed by the limitations
that someone conceived a priori. To reach the moral sphere, a robot must come
up with ethical principles that could be applied to a wide range of situations and
contexts instead of trying to decode them case by case: ethics as a reflection of
human decision-making. But, could ethics be resistant to computational science?
To solve this question, I would like to present an ethical cognitivist approach that
reduces many of the previous problematics when reflected inside the informatics
realm. This view of normativity depends on the agent’s position and it constitutes
a path to sustain moral objectiveness, without the necessity to assume (unlike “moral
realism”) the ontological existence of “moral facts”. One main theory is proposed
by philosopher Christine Korsgaard, let’s see shortly what it is about.
4 The next paragraphs combine the thoughts of philosopher Harry Frankfurt through Korsgaard’s view.
5 The desires of higher or lower order do not have to be dissonant, this only shows the sinuous and complex
labyrinths of the mind.
Author's personal copy
ETHICA EX MACHINA. Exploring artificial moral agency or the possibility of computable...
desires that exert a causal but not normative pressure on his behavior. When we act,
there is something that is above our desires, something that constitutes our identity
and allows us to choose which desire to endorse. Hence, according to Korsgaard
(2009, 2014), only through some practical identity we can determine which desires
become our reasons to act, and, therefore, unconditional obligations. We are our
own normative law, and this means that the principles that guide our actions express
who we are.
Social life implies diversity of practical identities. An individual can think of
himself as a citizen of a country, a father, an academic, or even as a gangster, and
what will determine how he acts is the conception he has of himself. This self-
conception is not purely descriptive or contemplative but practical, it is a conception
according to which we value ourselves or believe that certain actions are valuable.
Humans are constituted as such when they embrace their own desires since non-
compliance would imply to lose their identity, to cease to be who they are. That is
why our reasons become unconditional obligations to act.
This moral perspective is not ethical relativism or subjectivism. Our reasons to
act are not just private, they are rather defined in the public realm. Korsgaard‘s idea
is that although in principle we could describe our identity from individual aspects
and thereby ground how we act, in the end, the identity pilaster will always belong
to the “Kingdom of Ends”: we act as we would like others to recognize us (i.e. if
an individual sees himself as a feminist he will never act by denying his feminist
identity).
I will now briefly venture to answer if this ethical model could be computable in
some type of computer system.
From this cognitivist moral perspective, I will enumerate the minimum conditions
under which an artificial entity would be accepted as a genuine moral agent. It is
essential to examine the core of the theoretical model and subtract its fundamental
requirements. I plan to expose them from a general view, applicable to any entity.
In sum, the Korsgaardian requirements for moral agency could be:
(i) Intelligence
(ii) Autonomy (as self-governance)
(iii) Self-reflection
(iv) Having at least one practical identity
In the following sections, although my approach should be viewed from a reason-
able level of abstraction, I plan to distinguish, step by step, if an artificial agent could
integrate these characteristics to be recognized as a moral agent, without distorting
them. Let’s see if it is possible to get there.
Author's personal copy
R. Sanz
3.1 Intelligence
Despite the word “intelligence” has different connotations, there is a common ground
to all of them. The notion of intelligence is defined as a characterization in degrees of
acquisition, that is, not all entities qualify as “intelligent”. Some reach the threshold
needed to be more or less intelligent, but not others (like a stone or a vacuum
cleaner). Throughout history, the standard criterion turns out to be the human faculty
for information processing. Human beings are the ones who evaluate and classify
intelligent entities. But, how do we do that? What are the pragmatic principles we
use to categorize them?
Inside the realm of morality, a valid answer lies in the analysis of behavior on
the search of intelligence related to contextual practices – the ability to execute and
plan behavior as socio-spatial activities. But still, how is it possible to characterize
a mental state as a “belief” or “desire” on an entity with artificial intelligence?
Initially, I will focus my attention on a theory of the mind called “functionalism”6.
According to it, what makes some particular mental state is not its internal consti-
tution, but the way it works or the role it accomplishes within the system it is a part
of. A mental state is a functional state, the causal role of a certain state determines
what kind of mental state it is. Therefore, if a computer is reduced to processes that
implement computable functions, mental states could be understood as software
states.
The idea is to pay attention to how data is processed. An AMA will demonstrate
intelligent behavior when he acts “intelligently” according to an observer’s point
of view. An intelligent machine does not have to be “identical” to an intelligent
human, it should only exhibit the same type of behavior in similar circumstances.
This type of behavioral criteria could facilitate the conditions demanded from robots
when they irrupt into the field of morality7. Without a doubt, such task, even when
exhaustively detailed, will always be in conflict with skeptics. However, it could
be gradually alleviated if, in the near future, it weren’t easy to find behavioral
differences between a human being and his new existential partner with artificial
intelligence.
In previous sections, I exposed how humans are able to endorse or question their
own beliefs and desires, and that moral agency implies a reflective structure of
consciousness to legislate oneself. This self-reflection is a form of rationality that
has to do with the intelligent choice of ends. But, does an entity have to be self-
6 Hilary Putnam started the debate about the nature of mental states through a functionalist lens (Putnam
1980), but, later, he also argued against it. Last Putnam’s arguments try to show that functionalism cannot
give an adequate account of the content of propositional attitudes (Putnam 1988). In this article, function-
alism is intended to be a middle-ground approach between behaviorism and neuron-artificial-brain-specific
theories.
7 In the history of science, this type of procedure has been linked to Alan Turing, a mathematician who
proposed in the 1950s an “imitation game” to determine the achievement of an AI (called the “Turing
Test”). In the recent past, Allen et al. (2000) also suggested a “Moral Turing Test”.
Author's personal copy
ETHICA EX MACHINA. Exploring artificial moral agency or the possibility of computable...
aware to be a moral agent? And if so, when do we know with certainty that an entity
is really conscious?
The notion of “consciousness” is difficult to apprehend: there is no exact defini-
tion or a real understanding of the totality of its processes. From neurosciences to
psychological or philosophical approaches, there is a vast history of diverse theo-
ries about the structure of consciousness (Gennaro 2018). A predominant common
ground is to identify them with the subjective character of experience (Husserl
1962; Nagel 1974), however, to turn a subjective attribute into an objective science
seems almost impossible. Epistemologically, the mental state of another “I” is never
cognoscible since there are no direct means to connect with others’ subjective ex-
perience, and as a result, it is only communicable through language or behavior. So,
what are we trying to depict with the term “self-consciousness” then?
Imagine the mental life of an animal, for example, a squirrel. Presumably, squir-
rels have a wide variety of experiences: they feel pain or pleasure, they have a par-
ticular visual experience, they have beliefs or desires about the world around them.
But even so, it is hard to think of a squirrel “aware” of itself as humans are. Is the
squirrel critically endorsing its internal mental states?
According to the “Higher Order Thought” theory (Carruthers 2003; Rosenthal
2005), an experience is phenomenally conscious in virtue of other mental state that
is about the same experience. The idea is that a conscious experience of, for example,
a color, consists of a representation of such color in the visual cortex, accompanied
by a high order thought in the same subject to the effect that the person is having
with such color experience (Block 2009).
This type of proposal implies the ability to be aware of what is external to oneself
and the faculty to access our internal states and question them. Humans are aware
of their mental contents not only as an “internal experience” but as an exercise to
reflect critically on them. When I are aware of our mental activity (and therefore
of our behavior), it is plausible that, after thinking about it, I decide either that
a behavior must change, or an action is morally reprehensible, or I made a mistake
in my reasoning, or I have an unjustified belief that I should abandon.
From this perspective, to continue with the goals outlined at the beginning, human
self-consciousness does not seem to be a tacit need for a pragmatic development
of an AMA. On the other hand, “self-reflection” does, since a conscious mental
state must have the possibility of being questioned (or endorsed) from a hierarchical
level of abstraction. What I propose is that this condition sought by Korsgaard’s
theory (and that I am trying to mirror here within the computational field) should
not be understood as “self-consciousness” but rather as a requirement for “self-
reflection”. If a machine manages to behave morally in a consistent manner and can
respond for his actions and is able to learn from his previous behaviors, he could be
a moral agent in a very real sense, temporarily avoiding the philosophical “problem
of consciousness”. Hence, how can we define, in formal terms, this self-reflection?
At a computational level, self-reflection could consist of the faculty to monitor
lower level’s activities, so when a behavior becomes unproductive or unsuccessful,
the action is not persistently repeated; it is identified as useless and the agent must
try something new (Hofstadter 1982). Often, a robot has a set of first-order programs
that govern his most basic behaviors. Perhaps, one way of making a self-reflective
Author's personal copy
R. Sanz
robot is to write programs whose main job is to keep track of those first-order
behaviors to ensure they do not get stuck in not perceived circumstances. What
is sought is a general “self-observation” program: a program that controls other
programs but also maintains a critical eye towards its own behavior8.
This seems to be a general rule even for us. When we are not attentive, we are
vulnerable to get stuck in repetitive routines or conditioned actions. Think about
daily life, how many times do we make the same pointless action before correcting
it? No one is immune to that.
Clearly, a notion such as self-consciousness is essential for our human condition,
but from a computational point of view, for an AMA, it should be seen as self-
reflection. Isn’t a computer program with “self-observation” algorithms an idea very
close to it? I do not claim it is the same, in the end, only the future will reveal
whether our ability to know ourselves is equivalent to a complex set of computer
programs. In any case, we may think that self-observation programs will play a key
role in acquiring this requisite of self-reflection, crucial for moral beings.
An old common criticism in AI development is the idea that robots only act on how
they have been precoded. For sure, some programmatic coded doses are unavoidable
for the construction of an operative AMA. Nevertheless, the simple follow-up of
incontestable rules does not constitute a moral sovereignty but computational slavery
instead. Now, the question is whether we can grant enough autonomy to artificial
agents. An AMA governed only by fixed rules is not a genuine moral agent at all,
he must build principles that guide his decision-making even when the rules come
into conflict (and also to back down from them when certain circumstances require
it).
Among engineers, autonomy is often addressed from its technical aspects, not
from the metaphysical point of view. In many cases, they define “autonomy” as the
ability to be under no direct control of any other agent or user. The more competently
the robot independently achieves its objectives, the more agency is attributed to him.
But autonomy thus described is not enough to obtain moral agency. Many entities
such as bacteria, viruses, or even simple programs in artificial life, already exhibit
such autonomy, and, of course, they should not be seen as moral agents. Freedom of
action and will must arise as an emergent property in a sophisticated level of systemic
organization, as it emerges in humans. Every human being is limited by multiple
circumstances, specific to their cognitive capacity or external constraints imposed
on them. People are conditioned by their past actions, their present life choices and
future decisions that translate into the real capacity to implement them. Certainly, all
human freedom is “restricted freedom”, and so, being a moral agent does not imply
“radical freedom” (as compatibilists have defended). Something similar happens in
8 I am aware that this approach is related to a classic problem in computability theory: the halting problem.
But, even if a mathematical theorem guarantees that no software will ever be a perfect observer of itself, it
could simply be deduced that a perfect artificial intelligence is unattainable; something that should excite
us instead of worrying us (Hofstadter 1982).
Author's personal copy
ETHICA EX MACHINA. Exploring artificial moral agency or the possibility of computable...
the field of the AMAs: they will have freedom within the deterministic limitations
of their design.
So, how can a robot emancipate from its programmers? An interesting answer
comes with the notion of “psychological autonomy” (Frankfurt 1971) since it shows
that our human practices and actions can, in fact, be compatible with the belief of
determinism.
Two main ideas underlie this type of model of free will: the psychological “di-
vision” of a person into first-and-second order attitudes, and the characterization
of autonomy as a hierarchical mesh between second-order volitions and first-order
desires. Then, where is the will exactly? According to Frankfurt (1988), a “second-
order volition” is an “effective desire” that moves to act. A person can figure out any
desire, but if he yearns for one to truly become his will, then he will act specifically
on such first order desire rather than another. Ergo, personal autonomy, from this
psychological point of view, resides in the hierarchical structure of desires. First, the
person is constituted by identifying with some of his lower-order desires, second, he
governs himself by evaluating his desires according to his own higher-order desires.
Autonomy consists in the exercise of his reflexive capacity for self-identification
with (some of) his desires.
In terms of the agent’s own behavior, volition can be understood as the will to
move by an intentional act of the mind because it concerns the agent’s motivational
states. Here, the will is not a mysterious act but a process that can be attributed to
explain people’s behavior. Yet, if free will could be guaranteed from this type of
hierarchical theory, what would it take to compute it?
An initial approach would have to involve the distinction between different order
desires. Artificial agent architectures have design methodologies. In AI research,
there have been diverse perspectives for the construction of artificial intelligent
agents: top-down (deliberative or cognitive), bottom-up (behavior-based or reactive),
hybrid (a combination of both). Bottom-up approaches achieve their goals basically
through reactions on a changing environment, which often implies nor symbolic nor
representational knowledge. In fact, there is a universal acceptance that at least some
amount of intelligence is best modelled with behavior-based models (Brooks 1991),
but not all cognition can be systematized this way, let alone moral actions or social
practices.
I do not intent to expose a full review of artificial agent architectures9. Yet, the
combination of both methodologies increases the effectiveness of artificial systems
in solving ethical problems (Bryson 2000).
A powerful modelling framework in the field of robotics engineering is BDI
architecture (Rao and Georgeff 1995)10. Its success resides in the combination of
several valuable elements: a philosophical model for deliberative reason (Bratman
9 I am aware that there was controversy over whether behavior should be systematized using hierarchical
structures (Maes 1991; Tyrrell 1993; Vereijken and Whiting 1998) or if it emerges from a more dynamic
process (Van Gelder 1998).
10 In the recent past, Honarvar and Ghasem-Aghaee (2009a, 2009b) proposed a “Casuist BDI-Agent ar-
11 I would like to sound a philosophical alarm here. No matter how much I try to focus on the condition of
self-governance, the notion of autonomy is very vast. Perhaps, a critique may emerge from the conception
of autonomy regarding the capacity for reflective self-identification. I will work this aspect in subsequent
sections.
Author's personal copy
ETHICA EX MACHINA. Exploring artificial moral agency or the possibility of computable...
with information about the environment and about himself (beliefs), proattitudes
that guide him to action (desires and intentions) and decision-making mechanisms
using modal logic reasoning with the purpose of reaching circumstantial objectives
(Molina et al. 2005; Maes 1990).
I am not suggesting that robots could achieve “radical autonomy”; in fact, humans
do not need to live in a non-deterministic world to assume moral decisions. It doesn’t
matter if we are radically free or determined in some respects (or “programmed” for
the case of robots). Being free implies psychological freedom to act. Machines will
have “free will” in a similar (though not identical) way to us if their computational
architecture is the “correct” one.
Now, I need to raise a self-criticism here, how can an AMA be free if he doesn’t
understand what “freedom” means? Furthermore, how could robots possibly endorse
a desire to become a “normative obligation to act”? Or, even, how can an AMA attain
an artificial practical identity?
This whole issue arises here to understand what relationship a notion as “episodic
memory”12 (or “Lockean consciousness”) could have with practical reason. What
consequences can memory have on our reasons to act? And what requirements are
linked to it in the computable development of an artificial practical identity?
Memory is constitutive of the self-assessment processes that are intrinsically im-
portant for human beings. One of the goals of practical reason is to evaluate our
existence: how our past experiences were like (positive or negative), what relation-
ships we have with others, and so on. Memory forms an important part in the process
of personal identity construction from a narrative and historical sense. It involves
a self-biographical memory (Christman 2009). Without episodic memory it is not
possible to get involved in an identity project of self-understanding or to highlight
what could be meaningful for us in the light of the past actions’ evaluation. In fact,
my current condition has no meaning whatsoever outside the historical trajectory I
am part of. Knowing who I am means remembering the actions and decisions with
which I am involved in terms of normative obligations; it is a self-constitution over
time. Therefore, identity is formed from this introspective evaluation of beliefs and
desires that constitute the practical root of the individual, his system of values. I
identify myself with the promises I made because I am committed to be guided by
them. Without the ability to remember certain social relations or commitments, I
cannot assure that my current life plans form a coherent narrative. I understand my
worldly experience when I actively remember and interpret my past events as mine.
It is a historical identity13: a continuous process that involves a project of self-
identification in relation to others. We store our social experiences similarly to how
they have developed but perceived or interpreted from our perspective. These life
representations are not kept in our memory simply as fixed episodes, rather, they
are impregnated with meaning: it is an active reconstruction endowed with certain
coherence, relevance and planning, enclosing the normative values that we stamp
on them.
Surely, to program a robot to have a narrative history of “himself” or to worry
about certain things seems viable. But, would it really be possible for a robot to
question what “matters” to him as a moral reason?
Let’s go back to the initial objective. From the Korsgaardian conception, a practi-
cal identity generates epistemic commitments: it represents a set of beliefs that imply
morally relevant restrictions for a given identity. For that reason, individuals with
different identities often disagree on their beliefs. Perhaps, a computational model
for an artificial identity could be conceived as a precoded values system on which
moral actions are justified as circumstances come up. However, as previously stated,
this would not support moral principles as “normative obligations”, it would only be
an instrumentalization of a means to an end. Another tentative solution, as seen with
BDI architecture, would be to develop the normative endorsement as hierarchical
levels of “mental states” within an AMA embedded software module. Although this
12 From a psychological-cognitive perspective, I realize that human memory cannot be simplified into
a single concept. But, given the brevity of my article, and that I am linking it to personal (conscious)
identity, I have decided to focus on the categorization of episodic memory.
13 In this perspective, the influence of Paul Ricoeur (1990) should be recognized.
Author's personal copy
ETHICA EX MACHINA. Exploring artificial moral agency or the possibility of computable...
4 Conclusions
I would like to present now two major difficulties that make me conclude that an
AMA cannot currently become a genuine moral agent. Firstly, the notion of “artificial
identity” carries many problems (DiGiovanna 2017). Even if an AMA could acquire
all features of personal identity, this notion could be extended so largely that its
original meaning would be lost. It would force us to rethink the limits of what is
meant by it, because our perception of psychological and physical continuity, that
adapts to our concept of identity, would vanish. In robotics, the norm is to reprogram,
update and replace each of its parts. For an AMA, a radical change would be possible
both physically and “psychologically”. This would lead to the creation of entities
that are, to a large extent, capable of rewriting and reconstituting their identities or
bodies, that is, their whole being. An “artificial person” becomes too volatile: he
could, by his own will, or by his designers, get rid of the fundamental characteristics
that form his identity. In our social exchanges, we are used to identifing people over
time, can we say the same about a being that is capable of modifying ad libitum? We
presuppose a continuous ontic stability and we assign responsibility by identifying
people’s actions and consequences. If all the characteristics that constitute a person’s
identity are temporary or unstable, how can we trace an “artificial person” over time?
This new form of identity, capable of altering his own principles or beliefs, creates
enormous difficulties not only for the moral realm but also on a metaphysical level
(unimaginable until today). The possibility of access and free modification of their
“mental” contents entail moral dilemmas not foreseen in traditional ethical theories
and alters the conditions under which moral agents operate. How can we judge them,
hold them accountable, create a bond of friendship or trust them?
Even if an AMA had all the characteristics to be a person and passed all the tests
philosophers have established (self-reflection, ethical cognition, autonomy, second-
order volitions, etc.), he would also have a devastating feature: the ability to change
instantly and without effort. If this occurs, our ethical and moral theories are con-
ceived to blame or praise people, so which moral theory could be implemented to
AMAs?
Secondly, constructing an artificial personal identity in a narrative sense, as a con-
tinuous process of socially mediated self-interpretation, does not seem viable yet.
On the way to building artificial architectures, the research field of AI was initially
divided into two paradigms: symbolic (representational knowledge, logic rules) and
subsymbolic (artificial neural networks, machine learning). Nowadays, some inte-
grated systems propose that symbolic processing functionalities could emerge from
Author's personal copy
R. Sanz
References
Allen, Colin, Gary Varner, and Jason Zinser. 2000. Prolegomena to any future artificial moral agent. Jour-
nal of Experimental & Theoretical Artificial Intelligence 12.3: 251-261.
Allen, Colin, Wendell Wallach, and Iva Smit. 2006. Why machine ethics? IEEE Intelligent Systems 21.4:
12-17.
Anderson, Michael, and Susan Leigh Anderson, eds. 2011. Machine ethics. Cambridge University Press.
Anderson, Michael, S. Anderson, and Chris Armen. 2005. Towards machine ethics: Implementing two
action-based ethical theories. Proceedings of the AAAI 2005 Fall Symposium on Machine Ethics.
Audi, Robert and Paul Audi (Eds). 1999. The Cambridge dictionary of philosophy. Vol. 584. Cambridge:
Cambridge university press.
Block, Ned. 2009. Comparing the major theories of consciousness. The cognitive neurosciences.
1111–1122. Cambridge, MA: MIT Press.
Borst, Jelmer P. and John R. Anderson. 2015. Using the ACT-R Cognitive Architecture in combination
with fMRI data. An introduction to model-based cognitive neuroscience. Springer, New York, NY,
2015. 339-352.
Bratman, Michael. 1987. Intention, plans, and practical reason. Vol. 10. Cambridge, MA: Harvard Uni-
versity Press.
Bringsjord, Selmer. 2008. Ethical robots: the future can heed us. Ai & Society 22.4: 539-550.
Brooks, R. A. 1991. Intelligence without representation. Artificial intelligence, 47(1-3), 139-159.
Bryson, Joanna. 2000. Cross-paradigm analysis of autonomous agent architecture. Journal of Experimental
& Theoretical Artificial Intelligence 12.2. 165-189.
Carruthers, Peter. 2003. Phenomenal consciousness: A naturalistic theory. Cambridge University Press.
Choi, Dongkyu, and Pat Langley. 2018. Evolution of the ICARUS cognitive architecture. Cognitive Sys-
tems Research, 48: 25-38.
Christman, John. 2009. The politics of persons: Individual autonomy and socio-historical selves. Cam-
bridge University Press.
Dennett, Daniel C. 1997. When Hal Kills, Who’s to Blame? Computer Ethics. Hal’s Legacy: 2001’s Com-
puter as Dream and Reality, 351-365. Cambridge, MA: MIT Press.
DiGiovanna, James. 2017. Artificial Identity. Robot Ethics 2.0: From Autonomous Cars to Artificial Intel-
ligence.
Floridi, Luciano. 2004. Open problems in the philosophy of information. Metaphilosophy, 35, 554–582.
Floridi, Luciano and Jeff W. Sanders. 2004. On the morality of artificial agents. Minds and machines 14.3:
349-379.
Frankfurt, Harry. 1971. Free Will and the Concept of a Person. Journal of Philosophy 68.1: 5-20.
Frankfurt, Harry G. 1988. The importance of what we care about: Philosophical essays. Cambridge Uni-
versity Press.
Author's personal copy
R. Sanz
Van Gelder, Tim. 1998. The dynamical hypothesis in cognitive science. Behavioral and Brain Sciences
21(5), 615-628.
Gennaro, R. J. (Ed.). 2018. The Routledge Handbook of Consciousness. New York: Routledge.
Georgeff, M. and A. Rao. 1998. Rational software agents: from theory to practice. Agent technology.
Springer, Berlin, Heidelberg. 139-160.
Haddadi, Afsaneh, and Kurt Sundermeyer. 1996. Belief-desire-intention agent architectures. Foundations
of distributed artificial intelligence. 169-185.
Hassabis, Demis, D. Kumaran, C. Summerfeld, and M. Botvinick. 2017. Neuroscience-inspired artificial
intelligence. Neuron, 95(2), 245–258.
Hofstadter, D. R. 1982. Can creativity be mechanized. Scientific American 247.3: 18-34.
Honarvar, A. R. and N. Ghasem-Aghaee. 2009a. An artificial neural network approach for creating an eth-
ical artificial agent. 2009 IEEE International Symposium on Computational Intelligence in Robotics
and Automation-(CIRA). 290-295. New Jersey: IEEE.
Honarvar, A. R. and N. Ghasem-Aghaee. 2009b. Casuist BDI-agent: a new extended BDI architecture with
the capability of ethical reasoning. International conference on artificial intelligence and computa-
tional intelligence. 86-95. Springer, Berlin, Heidelberg.
Husserl, Edmund. 1962. Ideas: General introduction to pure phenomenology. 1913. Trans. WR Boyce
Gibson. New York: Collier.
Korsgaard, Christine M. 1996. The sources of normativity. Cambridge University Press.
Korsgaard, Christine M. 2009. Self-constitution: Agency, identity, and integrity. Oxford University Press:
USA.
Korsgaard, Christine M. 2014. The constitution of agency: Essays on practical reason and moral psychol-
ogy. Oxford University Press: USA.
Laird, J. E. 2012. The Soar cognitive architecture. Cambridge, MA: MIT Press.
Laird, J. E., C. Lebiere and P. S. Rosenbloom. 2017. A standard model of the mind: Toward a common
computational framework across artificial intelligence, cognitive science, neuroscience, and robotics.
AI Magazine, 38(4), 13–26.
Locke, John. 1841. An essay concerning human understanding.
Maes, Pattie. 1990. Situated agents can have goals. Robotics and autonomous systems 6.1-2: 49-70.
Maes, Pattie. 1991. A bottom-up mechanism for behavior selection in artificial creature. Proceedings of
First International Conference on Simulation of Adaptive Behavior (SAB’90). Cambridge: The MIT
Press
Molina, José Manuel, Juan Manuel Corchado Rodríguez, and Juan Pavón Mestras. 2005. Modelos y arqui-
tecturas de agente. Agentes software y sistemas multiagente: conceptos, arquitecturas y aplicaciones.
New Jersey: Prentice Hall.
Moor, James H. 2006. The nature, importance, and difficulty of machine ethics. IEEE intelligent systems
21.4: 18-21.
Nagel, Thomas. 1974. What is it like to be a bat? The philosophical review, 83(4), 435-450.
Parfit, Derek. 1984. Reasons and persons. New York: Oxford University Press.
Putnam, Hilary. 1980. The nature of mental states. Readings in philosophy of psychology 1: 223-231
Putnam, Hilary. 1988. Representation and reality. Cambridge, MA: MIT Press.
Rao, Anand S. and Michael P. Georgeff. 1995. BDI agents: from theory to practice. ICMAS. Vol. 95.
Ricoeur, Paul. 1990. Soi-même comme un autre. Paris: Le Seuil.
Rosenthal, David. 2005. Consciousness and mind. UK: Oxford University Press.
Schild, Klaus. 2000. On the Relationship Between BDI Logics and Standard Logics of Concurency. Au-
tonomous Agents and Multi-Agent Systems 3, 259–283 https://doi.org/10.1023/A:101007602770.
Schneider, Howard. 2019. Emergence of belief systems and the future of artificial intelligence. Biologically
Inspired Cognitive Architectures Meeting (pp. 485-494). Springer, Cham.
Stahl, Bernd C. 2004. Information, ethics, and computers: The problem of autonomous moral agents.
Minds and Machines, 14(1), 67-83.
Sullins, John. P. 2006. When is a robot a moral agent. Machine ethics, 151-160.
Taddeo, M. and Floridi, L. 2007. A praxical solution of the symbol grounding problem. Minds and Ma-
chines, 17(4), 369-389.
Trafton, J. G., L. M. Hiatt, A. M. Harrison, F. P. Tamborello II, S. S. Khemlani and A. C. Schultz (2013).
Act-r/e: An embodied cognitive architecture for human-robot interaction. Journal of Human-Robot
Interaction 2(1), 30-55.
Tyrrell, Toby. 1993. The Use of Hierarchies for Action Selection. Adaptive Behavior 1(4), 387–420.
Vereijken, Beatrix and H. Whiting. 1998. Hoist by their own petard: The constraints of hierarchical models.
Behavioral and Brain Sciences 21(5), 705-705.
Author's personal copy
ETHICA EX MACHINA. Exploring artificial moral agency or the possibility of computable...
Wallach, Wendell and C. Allen. 2008. Moral machines: Teaching robots right from wrong. Oxford Univer-
sity Press.
Wiegel, Vincent. 2006. Building blocks for artificial moral agents. Proc. Artificial Life X.