Machine Ethics: Creating an Ethical Intelligent Agent.

Article (PDF Available)inAi Magazine 28(4):15-26 · December 2007with499 Reads
Source: DBLP
The newly emerging field of machine ethics (Anderson and Anderson 2006) is concerned with adding an ethical dimension to machines. Unlike computer ethics—which has traditional- ly focused on ethical issues surrounding humans’ use of machines—machine ethics is concerned with ensuring that the behavior of machines toward human users, and perhaps other machines as well, is ethically acceptable. In this article we discuss the importance of machine ethics, the need for machines that rep- resent ethical principles explicitly, and the chal- lenges facing those working on machine ethics. We also give an example of current research in the field that shows that it is possible, at least in a limited domain, for a machine to abstract an ethical principle from examples of correct ethi- cal judgments and use that principle to guide its own behavior.

Full-text (PDF)

Available from: Michael Anderson
The newly emerging field of machine ethics
(Anderson and Anderson 2006) is concerned
with adding an ethical dimension to machines.
Unlike computer ethics—which has traditional-
ly focused on ethical issues surrounding
humans’ use of machines—machine ethics is
concerned with ensuring that the behavior of
machines toward human users, and perhaps
other machines as well, is ethically acceptable.
In this article we discuss the importance of
machine ethics, the need for machines that rep-
resent ethical principles explicitly, and the chal-
lenges facing those working on machine ethics.
We also give an example of current research in
the field that shows that it is possible, at least in
a limited domain, for a machine to abstract an
ethical principle from examples of correct ethi-
cal judgments and use that principle to guide its
own behavior.
The ultimate goal of machine ethics, we
believe, is to create a machine that itself
follows an ideal ethical principle or set of
principles; that is to say, it is guided by this
principle or these principles in decisions it
makes about possible courses of action it could
take. We need to make a distinction between
what James Moor has called an “implicit ethical
agent” and an “explicit ethical agent” (Moor
2006). According to Moor, a machine that is an
implicit ethical agent is one that has been pro-
grammed to behave ethically, or at least avoid
unethical behavior, without an explicit repre-
sentation of ethical principles. It is constrained
in its behavior by its designer who is following
ethical principles. A machine that is an explic-
it ethical agent, on the other hand, is able to
calculate the best action in ethical dilemmas
using ethical principles. It can “represent ethics
explicitly and then operate effectively on the
basis of this knowledge.” Using Moor’s termi-
nology, most of those working on machine
ethics would say that the ultimate goal is to
create a machine that is an explicit ethical
We are, here, primarily concerned with the
ethical decision making itself, rather than how
a machine would gather the information need-
ed to make the decision and incorporate it into
its general behavior. It is important to see this
as a separate and considerable challenge. It is
separate because having all the information
and facility in the world won’t, by itself, gen-
erate ethical behavior in a machine. One needs
to turn to the branch of philosophy that is con-
cerned with ethics for insight into what is con-
sidered to be ethically acceptable behavior. It is
a considerable challenge because, even among
experts, ethics has not been completely codi-
fied. It is a field that is still evolving. We shall
argue that one of the advantages of working on
machine ethics is that it might lead to break-
throughs in ethical theory, since machines are
well-suited for testing the results of consistent-
ly following a particular ethical theory.
One other point should be made in intro-
ducing the subject of machine ethics. Ethics
can be seen as both easy and hard. It appears
easy because we all make ethical decisions on a
daily basis. But that doesn’t mean that we are
all experts in ethics. It is a field that requires
much study and experience. AI researchers
must have respect for the expertise of ethicists
just as ethicists must appreciate the expertise of
AI researchers. Machine ethics is an inherently
interdisciplinary field.
WINTER 2007 15
Copyright © 2007, American Association for Artificial Intelligence. All rights reserved. ISSN 0738-4602
Machine Ethics:
Creating an Ethical
Intelligent Agent
Michael Anderson and Susan Leigh Anderson
AI Magazine Volume 28 Number 4 (2007) (© AAAI)
The Importance of
Machine Ethics
Why is the field of machine ethics important?
There are at least three reasons that can be giv-
en. First, there are ethical ramifications to what
machines currently do and are projected to do
in the future. To neglect this aspect of machine
behavior could have serious repercussions.
South Korea has recently mustered more than
30 companies and 1000 scientists to the end of
putting “a robot in every home by 2010”
(Onishi 2006). DARPAs grand challenge to
have a vehicle drive itself across 132 miles of
desert terrain has been met, and a new grand
challenge is in the works that will have vehi-
cles maneuvering in an urban setting. The
United States Army’s Future Combat Systems
program is developing armed robotic vehicles
that will support ground troops with “direct-
fire” and antitank weapons. From family cars
that drive themselves and machines that dis-
charge our daily chores with little or no assis-
tance from us, to fully autonomous robotic
entities that will begin to challenge our
notions of the very nature of intelligence, it is
clear that machines such as these will be capa-
ble of causing harm to human beings unless
this is prevented by adding an ethical compo-
nent to them.
Second, it could be argued that humans’ fear
of the possibility of autonomous intelligent
machines stems from their concern about
whether these machines will behave ethically,
so the future of AI may be at stake. Whether
society allows AI researchers to develop any-
thing like autonomous intelligent machines
may hinge on whether they are able to build in
safeguards against unethical behavior. From
the murderous robot uprising in the 1920 play
R.U.R. (Capek 1921) and the deadly coup d’état
perpetrated by the HAL 9000 computer in
2001: A Space Odyssey (Clarke 1968), to The
Matrix virtual reality simulation for the pacifi-
cation and subjugation of human beings by
machines, popular culture is rife with images of
machines devoid of any ethical code mistreat-
ing their makers. In his widely circulated trea-
tise, “Why the future doesn’t need us,” Bill Joy
(2000) argues that the only antidote to such
fates and worse is to “relinquish dangerous
technologies.” We believe that machine ethics
research may offer a viable, more realistic solu-
Finally, we believe that it’s possible that
research in machine ethics will advance the
study of ethical theory. Ethics, by its very
nature, is the most practical branch of philoso-
phy. It is concerned with how agents ought to
behave when faced with ethical dilemmas.
Despite the obvious applied nature of the field
of ethics, too often work in ethical theory is
done with little thought to actual application.
When examples are discussed, they are typical-
ly artificial examples. Research in machine
ethics has the potential to discover problems
with current theories, perhaps even leading to
the development of better theories, as AI
researchers force scrutiny of the details
involved in actually applying an ethical theory
to particular cases. As Daniel Dennett (2006)
recently stated, AI “makes philosophy honest.”
Ethics must be made computable in order to
make it clear exactly how agents ought to
behave in ethical dilemmas.
An exception to the general rule that ethi-
cists don’t spend enough time discussing actu-
al cases occurs in the field of biomedical ethics,
a field that has arisen out of a need to resolve
pressing problems faced by health-care work-
ers, insurers, hospital ethics boards, and bio-
medical researchers. As a result of there having
been more discussion of actual cases in the
field of biomedical ethics, a consensus is begin-
ning to emerge as to how to evaluate ethical
dilemmas in this domain, leading to the ethi-
cally correct action in many dilemmas. A rea-
son there might be more of a consensus in this
domain than in others is that in the area of bio-
medical ethics there is an ethically defensible
goal (the best possible health of the patient),
whereas in other areas (such as business and
law) the goal may not be ethically defensible
(make as much money as possible, serve the
client’s interest even if he or she is guilty of an
offense or doesn’t deserve a settlement) and
ethics enters the picture as a limiting factor
(the goal must be achieved within certain eth-
ical boundaries).
AI researchers working with ethicists might
find it helpful to begin with this domain, dis-
covering a general approach to computing
ethics that not only works in this domain, but
could be applied to other domains as well.
Explicit Ethical Machines
It does seem clear, to those who have thought
about the issue, that some sort of safeguard
should be in place to prevent unethical machine
behavior (and that work in this area may pro-
vide benefits for the study of ethical theory as
well). This shows the need for creating at least
implicit ethical machines; but why must we cre-
ate explicit ethical machines, which would seem
to be a much greater (perhaps even an impossi-
ble) challenge for AI researchers? Furthermore,
many fear handing over the job of ethical over-
seer to machines themselves. How could we feel
confident that a machine would make the right
decision in situations that were not anticipated?
Finally, what if the machine starts out behaving
in an ethical fashion but then morphs into one
that decides to behave unethically in order to
secure advantages for itself?
On the need for explicit, rather than just
implicit, ethical machines: What is critical in
the “explicit ethical agent” versus “implicit
ethical agent” distinction, in our view, lies not
only in who is making the ethical judgments
(the machine versus the human programmer),
but also in the ability to justify ethical judg-
ments that only an explicit representation of
ethical principles allows. An explicit ethical
agent is able to explain why a particular action
is either right or wrong by appealing to an eth-
ical principle. A machine that has learned, or
been programmed, to make correct ethical
judgments, but does not have principles to
which it can appeal to justify or explain its
judgments, is lacking something essential to
being accepted as an ethical agent. Immanuel
Kant (1785) made a similar point when he dis-
tinguished between an agent that acts from a
sense of duty (consciously following an ethical
principle), rather than merely in accordance
with duty, having praise only for the former.
If we believe that machines could play a role
in improving the lives of human beings—that
this is a worthy goal of AI research—then, since
it is likely that there will be ethical ramifica-
tions to their behavior, we must feel confident
that these machines will act in a way that is
ethically acceptable. It will be essential that
they be able to justify their actions by appeal-
ing to acceptable ethical principles that they
are following, in order to satisfy humans who
will question their ability to act ethically. The
ethical component of machines that affect
humans’ lives must be transparent, and princi-
ples that seem reasonable to human beings
provide that transparency. Furthermore, the
concern about how machines will behave in
situations that were not anticipated also sup-
ports the need for explicit ethical machines.
The virtue of having principles to follow, rather
than being programmed in an ad hoc fashion
to behave correctly in specific situations, is that
it allows machines to have a way to determine
the ethically correct action in new situations,
even in new domains. Finally, Marcello Guari-
ni (2006), who is working on a neural network
model of machine ethics, where there is a pre-
disposition to eliminate principles, argues that
principles seem to play an important role in
revising ethical beliefs, which is essential to eth-
ical agency. He contends, for instance, that
they are necessary to discern morally relevant
differences in similar cases.
The concern that machines that start out
behaving ethically will end up behaving
unethically, perhaps favoring their own inter-
ests, may stem from fears derived from legiti-
mate concerns about human behavior. Most
human beings are far from ideal models of eth-
ical agents, despite having been taught ethical
principles; and humans do, in particular, tend
to favor themselves. Machines, though, might
have an advantage over human beings in terms
of behaving ethically. As Eric Dietrich (2006)
has recently argued, human beings, as biologi-
cal entities in competition with others, may
have evolved into beings with a genetic predis-
position toward unethical behavior as a sur-
vival mechanism. Now, though, we have the
chance to create entities that lack this predis-
position, entities that might even inspire us to
behave more ethically. Consider, for example,
Andrew, the robot hero of Isaac Asimov’s story
“The Bicentennial Man” (1976), who was far
more ethical than the humans with whom he
came in contact. Dietrich maintained that the
machines we fashion to have the good quali-
ties of human beings, and that also follow prin-
ciples derived from ethicists who are the excep-
tion to the general rule of unethical human
beings, could be viewed as “humans 2.0”—a
better version of human beings.
This may not completely satisfy those who
are concerned about a future in which human
beings share an existence with intelligent,
autonomous machines. We face a choice, then,
between allowing AI researchers to continue in
their quest to develop intelligent, autonomous
machines—which will have to involve adding
an ethical component to them—or stifling this
research. The likely benefits and possible harms
of each option will have to be weighed. In any
case, there are certain benefits to continuing to
work on machine ethics. It is important to find
a clear, objective basis for ethics—making
ethics in principle computable—if only to rein
in unethical human behavior; and AI
researchers, working with ethicists, have a bet-
ter chance of achieving breakthroughs in ethi-
cal theory than theoretical ethicists working
alone. There is also the possibility that society
would not be able to prevent some researchers
from continuing to develop intelligent,
autonomous machines, even if society decides
that it is too dangerous to support such work.
If this research should be successful, it will be
important that we have ethical principles that
we insist should be incorporated into such
machines. The one thing that society should
fear more than sharing an existence with intel-
WINTER 2007 17
human to consider the effects of each
of those actions on all those affected.
Finally, for some individuals’
actions—actions of the president of
the United States or the CEO of a large
international corporation—their
impact can be so great that the calcu-
lation of the greatest net pleasure may
be very time consuming, and the
speed of today’s machines gives them
an advantage.
We conclude, then, that machines
can follow the theory of act utilitari-
anism at least as well as human beings
and, perhaps, even better, given the
data that human beings would need, as
well, to follow the theory. The theory
of act utilitarianism has, however,
been questioned as not entirely agree-
ing with intuition. It is certainly a
good starting point in programming a
machine to be ethically sensitive—it
would probably be more ethically sen-
sitive than many human beings—but,
perhaps, a better ethical theory can be
Critics of act utilitarianism have
pointed out that it can violate human
beings’ rights, sacrificing one person
for the greater net good. It can also
conflict with our notion of justice—
what people deserve—because the
rightness and wrongness of actions is
determined entirely by the future con-
sequences of actions, whereas what
people deserve is a result of past
behavior. A deontological approach to
ethics (where the rightness and
wrongness of actions depends on
something other than the conse-
quences), such as Kant’s categorical
imperative, can emphasize the impor-
tance of rights and justice, but this
approach can be accused of ignoring
consequences. We believe, along with
W. D. Ross (1930), that the best
approach to ethical theory is one that
combines elements of both teleologi-
cal and deontological theories. A the-
ory with several prima facie duties
(obligations that we should try to sat-
isfy, but which can be overridden on
occasion by stronger obligations)—
some concerned with the conse-
quences of actions and others con-
cerned with justice and rights—better
acknowledges the complexities of eth-
ical decision making than a single
absolute duty theory. This approach
ligent, autonomous machines is shar-
ing an existence with machines like
these without an ethical component.
Challenges Facing
Those Working on
Machine Ethics
The challenges facing those working
on machine ethics can be divided into
two main categories: philosophical
concerns about the feasibility of com-
puting ethics and challenges from the
AI perspective. In the first category, we
need to ask whether ethics is the sort
of thing that can be computed. One
well-known ethical theory that sup-
ports an affirmative answer to this
question is “act utilitarianism.” Ac -
cording to this teleological theory (a
theory that maintains that the right-
ness and wrongness of actions is deter-
mined entirely by the consequences of
the actions) that act is right which, of
all the actions open to the agent, is
likely to result in the greatest net good
consequences, taking all those affect-
ed by the action equally into account.
Essentially, as Jeremy Bentham (1781)
long ago pointed out, the theory in -
volves performing “moral arithmetic.”
Of course, before doing the arith-
metic, one needs to know what counts
as a “good” and “bad” consequence.
The most popular version of act utili-
tarianism—hedonistic act utilitarian-
ism—would have us consider the
pleasure and displeasure that those
affected by each possible action are
likely to receive. And, as Bentham
pointed out, we would probably need
some sort of scale to account for such
things as the intensity and duration of
the pleasure or displeasure that each
individual affected is likely to receive.
This is information that a human
being would need to have as well to
follow the theory. Getting this infor-
mation has been and will continue to
be a challenge for artificial intelligence
research in general, but it can be sepa-
rated from the challenge of computing
the ethically correct action, given this
information. With the requisite infor-
mation, a machine could be developed
that is just as able to follow the theory
as a human being.
Hedonistic act utilitarianism can be
implemented in a straightforward
manner. The algorithm is to compute
the best action, that which derives the
greatest net pleasure, from all alterna-
tive actions. It requires as input the
number of people affected and, for
each person, the intensity of the pleas-
ure/displeasure (for example, on a
scale of 2 to –2), the duration of the
pleasure/displeasure (for example, in
days), and the probability that this
pleasure or displeasure will occur, for
each possible action. For each person,
the algorithm computes the product
of the intensity, the duration, and the
probability, to obtain the net pleasure
for that person. It then adds the indi-
vidual net pleasures to obtain the total
net pleasure:
Total net pleasure = (intensity ×
duration ×probability) for each af -
fected individual
This computation would be performed
for each alternative action. The action
with the highest total net pleasure is
the right action (Anderson, Anderson,
and Armen 2005b).
A machine might very well have an
advantage over a human being in fol-
lowing the theory of act utilitarianism
for several reasons: First, human
beings tend not to do the arithmetic
strictly, but just estimate that a certain
action is likely to result in the greatest
net good consequences, and so a
human being might make a mistake,
whereas such error by a machine
would be less likely. Second, as has
already been noted, human beings
tend toward partiality (favoring them-
selves, or those near and dear to them,
over others who might be affected by
their actions or inactions), whereas an
impartial machine could be devised.
Since the theory of act utilitarianism
was developed to introduce objectivi-
ty into ethical decision making, this is
important. Third, humans tend not to
consider all of the possible actions
that they could perform in a particu-
lar situation, whereas a more thor-
ough machine could be developed.
Imagine a machine that acts as an
advisor to human beings and “thinks”
like an act utilitarian. It will prompt
the human user to consider alterna-
tive actions that might result in
greater net good consequences than
the action the human being is consid-
ering doing, and it will prompt the
has one major drawback, however. It
needs to be supplemented with a deci-
sion procedure for cases where the pri-
ma facie duties give conflicting advice.
This is a problem that we have worked
on and will be discussed later on.
Among those who maintain that
ethics cannot be computed, there are
those who question the action-based
approach to ethics that is assumed by
defenders of act utilitarianism, Kant’s
categorical imperative, and other
well-known ethical theories. Accord-
ing to the “virtue” approach to ethics,
we should not be asking what we
ought to do in ethical dilemmas, but
rather what sort of persons we should
be. We should be talking about the
sort of qualities—virtues—that a per-
son should possess; actions should be
viewed as secondary. Given that we
are concerned only with the actions
of machines, it is appropriate, howev-
er, that we adopt the action-based
approach to ethical theory and focus
on the sort of principles that
machines should follow in order to
behave ethically.
Another philosophical concern
with the machine ethics project is
whether machines are the type of enti-
ties that can behave ethically. It is
commonly thought that an entity
must be capable of acting intentional-
ly, which requires that it be conscious,
and that it have free will, in order to
be a moral agent. Many would, also,
add that sentience or emotionality is
important, since only a being that has
feelings would be capable of appreci-
ating the feelings of others, a critical
factor in the moral assessment of pos-
sible actions that could be performed
in a given situation. Since many doubt
that machines will ever be conscious,
have free will, or emotions, this would
seem to rule them out as being moral
This type of objection, however,
shows that the critic has not recog-
nized an important distinction
between performing the morally cor-
rect action in a given situation, includ-
ing being able to justify it by appeal-
ing to an acceptable ethical principle,
and being held morally responsible for
the action. Yes, intentionality and free
will in some sense are necessary to
hold a being morally responsible for
WINTER 2007 19
its actions, and it would be difficult to
establish that a machine possesses
these qualities; but neither attribute is
necessary to do the morally correct
action in an ethical dilemma and jus-
tify it. All that is required is that the
machine act in a way that conforms
with what would be considered to be
the morally correct action in that situ-
ation and be able to justify its action
by citing an acceptable ethical princi-
ple that it is following (S. L. Anderson
The connection between emotion-
ality and being able to perform the
morally correct action in an ethical
dilemma is more complicated. Cer-
tainly one has to be sensitive to the
suffering of others to act morally. This,
for human beings, means that one
must have empathy, which, in turn,
requires that one have experienced
similar emotions oneself. It is not
clear, however, that a machine could
not be trained to take into account the
suffering of others in calculating how
it should behave in an ethical dilem-
ma, without having emotions itself. It
is important to recognize, further-
more, that having emotions can actu-
ally interfere with a being’s ability to
determine, and perform, the right
action in an ethical dilemma. Humans
are prone to getting “carried away” by
their emotions to the point where
they are incapable of following moral
principles. So emotionality can even
be viewed as a weakness of human
beings that often prevents them from
doing the “right thing.”
The necessity of emotions in ration-
al decision making in computers has
been championed by Rosalind Picard
(1997), citing the work of Damasio
(1994), which concludes that human
beings lacking emotion repeatedly
make the same bad decisions or are
unable to make decisions in due time.
We believe that, although evolution
may have taken this circuitous path to
decision making in human beings,
irrational control of rational processes
is not a necessary condition for all
rational systems—in particular, those
specifically designed to learn from
errors, heuristically prune search
spaces, and make decisions in the face
of bounded time and knowledge.
A final philosophical concern with
the feasibility of computing ethics has
to do with whether there is a single
correct action in ethical dilemmas.
Many believe that ethics is relative
either to the society in which one
lives—“when in Rome, one should do
what Romans do”—or, a more extreme
version of relativism, to individuals—
whatever you think is right is right for
you. Most ethicists reject ethical rela-
tivism (for example, see Mappes and
DeGrazia [2001, p. 38] and Gazzaniga
[2006, p. 178]), in both forms, prima-
rily because this view entails that one
cannot criticize the actions of soci-
eties, as long as they are approved by
the majority in those societies, or indi-
viduals who act according to their
beliefs, no matter how heinous they
are. There certainly do seem to be
actions that experts in ethics, and
most of us, believe are absolutely
wrong (for example, torturing a baby
and slavery), even if there are societies,
or individuals, who approve of the
actions. Against those who say that
ethical relativism is a more tolerant
view than ethical absolutism, it has
been pointed out that ethical rela-
tivists cannot say that anything is
absolutely good—even tolerance (Poj-
man [1996, p. 13]).
What defenders of ethical relativism
may be recognizing—that causes them
to support this view—are two truths,
neither of which entails the accept-
ance of ethical relativism: (1) Different
societies have their own customs that
we must acknowledge, and (2) there
are difficult ethical issues about which
even experts in ethics cannot agree, at
the present time, on the ethically cor-
rect action. Concerning the first truth,
we must distinguish between an ethi-
cal issue and customs or practices that
fall outside the area of ethical concern.
Customs or practices that are not a
matter of ethical concern can be
respected, but in areas of ethical con-
cern we should not be tolerant of
unethical practices.
Concerning the second truth, that
some ethical issues are difficult to
resolve (for example, abortion)—and
so, at this time, there may not be
agreement by ethicists as to the correct
action—it does not follow that all
views on these issues are equally cor-
rect. It will take more time to resolve
these issues, but most ethicists believe
that we should strive for a single cor-
rect position on these issues. What
needs to happen is to see that a certain
position follows from basic principles
that all ethicists accept, or that a cer-
tain position is more consistent with
other beliefs that they all accept.
From this last point, we should see
that we may not be able to give
machines principles that resolve all
ethical disputes at this time. (Hopeful-
ly, the machine behavior that we are
concerned about won’t fall in too
many of the disputed areas.) The
implementation of ethics can’t be
more complete than is accepted ethi-
cal theory. Completeness is an ideal
for which to strive but may not be pos-
sible at this time. The ethical theory,
or framework for resolving ethical dis-
putes, should allow for updates, as
issues that once were considered con-
tentious are resolved. What is more
important than having a complete
ethical theory to implement is to have
one that is consistent. This is where
machines may actually help to
advance the study of ethical theory, by
pointing out inconsistencies in the
theory that one attempts to imple-
ment, forcing ethical theoreticians to
resolve those inconsistencies.
Considering challenges from an AI
perspective, foremost for the nascent
field of machine ethics may be con-
vincing the AI community of the
necessity and advisability of incorpo-
rating ethical principles into ma -
chines. Some critics maintain that
machine ethics is the stuff of science
fiction—machines are not yet (and
may never be) sophisticated enough to
require ethical restraint. Others won-
der who would deploy such systems
given the possible liability involved.
We contend that machines with a lev-
el of autonomy requiring ethical delib-
eration are here and both their num-
ber and level of autonomy are likely to
increase. The liability already exists;
machine ethics is necessary as a means
to mitigate it. In the following section,
we will detail a system that helps
establish this claim.
Another challenge facing those con-
cerned with machine ethics is how to
proceed in such an inherently inter-
disciplinary endeavor. Artificial Intel-
ligence researchers and philosophers,
although generally on speaking terms,
do not always hear what the other is
saying. It is clear that, for substantive
advancement of the field of machine
ethics, both are going to have to listen
to each other intently. AI researchers
will need to admit their naiveté in the
field of ethics and convince philoso-
phers that there is a pressing need for
their services; philosophers will need
to be a bit more pragmatic than many
are wont to be and make an effort to
sharpen ethical theory in domains
where machines will be active. Both
will have to come to terms with this
newly spawned relationship and,
together, forge a common language
and research methodology.
The machine ethics research agenda
will involve testing the feasibility of a
variety of approaches to capturing eth-
ical reasoning, with differing ethical
bases and implementation forma lisms,
and applying this reasoning in systems
engaged in ethically sensitive activi-
ties. This research will investigate how
to determine and represent ethical
principles, incorporate ethical princi-
ples into a system’s decision procedure,
make ethical decisions with incom-
plete and uncertain knowledge, pro-
vide explanations for decisions made
using ethical principles, and evaluate
systems that act based upon ethical
System implementation work is
already underway. A range of ma -
chine-learning techniques are being
employed in an attempt to codify eth-
ical reasoning from examples of par-
ticular ethical dilemmas. As such, this
work is based, to a greater or lesser
degree, upon casuistry—the branch of
applied ethics that, eschewing princi-
ple-based approaches to ethics,
attempts to determine correct respons-
es to new ethical dilemmas by drawing
conclusions based on parallels with
previous cases in which there is agree-
ment concerning the correct response.
Rafal Rzepka and Kenji Araki (2005),
at what might be considered the most
extreme degree of casuistry, explore
how statistics learned from examples
of ethical intuition drawn from the
full spectrum of the world wide web
might be useful in furthering machine
ethics. Working in the domain of safe-
ty assurance for household robots,
they question whether machines
should be obeying some set of rules
decided by ethicists, concerned that
these rules may not in fact be truly
universal. They suggest that it might
be safer to have machines “imitating
millions, not a few,” believing in such
“democracy-dependent algorithms”
because, they contend, “most people
behave ethically without learning
ethics.” They propose an extension to
their web-based knowledge discovery
system GENTA (General Belief Retriev-
ing Agent) that would search the web
for opinions, usual behaviors, com-
mon consequences, and exceptions,
by counting ethically relevant neigh-
boring words and phrases, aligning
these along a continuum from posi-
tive to negative behaviors, and sub-
jecting this information to statistical
analysis. They suggest that this analy-
sis, in turn, would be helpful in the
development of a sort of majority-rule
ethics useful in guiding the behavior
of autonomous systems. An important
open question is whether users will be
comfortable with such behavior or
will, as might be expected, demand
better than average ethical conduct
from autonomous systems.
A neural network approach is
offered by Marcello Guarini (2006). At
what might be considered a less
extreme degree of casuistry, particular
actions concerning killing and allow-
ing to die are classified as acceptable or
unacceptable depending upon differ-
ent motives and consequences. After
training a simple recurrent network on
a number of such cases, it is capable of
providing plausible responses to a
variety of previously unseen cases.
This work attempts to shed light on
the philosophical debate concerning
generalism (principle-based approaches
to moral reasoning) versus particular-
ism (case-based approaches to moral
reasoning). Guarini finds that,
although some of the concerns per-
taining to learning and generalizing
from ethical dilemmas without resort-
ing to principles can be mitigated with
a neural network model of cognition,
“important considerations suggest
that it cannot be the whole story
about moral reasoning—principles are
needed.” He argues that “to build an
WINTER 2007 21
artificially intelligent agent without
the ability to question and revise its
own initial instruction on cases is to
assume a kind of moral and engineer-
ing perfection on the part of the
designer.” He argues, further, that such
perfection is unlikely and principles
seem to play an important role in the
required subsequent revision—“at
least some reflection in humans does
appear to require the explicit repre-
sentation or consultation of…rules,”
for instance, in discerning morally rel-
evant differences in similar cases. Con-
cerns about this approach are those
attributable to neural networks in gen-
eral, including oversensitivity to train-
ing cases and the inability to generate
reasoned arguments for system re -
Bruce McLaren (2003), in the spirit
of a more pure form of casuistry, pro-
motes a case-based reasoning ap -
proach (in the artificial intelligence
sense) for developing systems that
provide guidance in ethical dilemmas.
His first such system, Truth-Teller,
compares pairs of cases presenting eth-
ical dilemmas about whether or not to
tell the truth.
The Truth-Teller program marshals
ethically relevant similarities and dif-
ferences between two given cases from
the perspective of the “truth teller”
(that is, the person faced with the
dilemma) and reports them to the
user. In particular, it points out reasons
for telling the truth (or not) that (1)
apply to both cases, (2) apply more
strongly in one case than another, or
(3) apply to only one case.
The System for Intelligent Retrieval
of Operationalized Cases and Codes
(SIROCCO), McLaren’s second pro-
gram, leverages information concern-
ing a new ethical dilemma to predict
which previously stored principles and
cases are relevant to it in the domain
of professional engineering ethics.
Cases are exhaustively formalized and
this formalism is used to index similar
cases in a database of formalized, pre-
viously solved cases that include prin-
ciples used in their solution. SIROC-
CO’s goal, given a new case to analyze,
is “to provide the basic information
with which a human reasoner …
could answer an ethical question and
then build an argument or rationale
for that conclusion.” SIROCCO is suc-
cessful at retrieving relevant cases but
performed beneath the level of an eth-
ical review board presented with the
same task. Deductive techniques, as
well as any attempt at decision mak-
ing, are eschewed by McLaren due to
“the ill-defined nature of problem
solving in ethics.” Critics might con-
tend that this “ill-defined nature” may
not make problem solving in ethics
completely indefinable, and attempts
at just such a definition may be possi-
ble in constrained domains. Further, it
might be argued that decisions offered
by a system that are consistent with
decisions made in previous cases have
merit and will be useful to those seek-
ing ethical advice.
We (Anderson, Anderson, and
Armen 2006a) have developed a deci-
sion procedure for an ethical theory in
a constrained domain that has multi-
ple prima facie duties, using inductive
logic programming (ILP) (Lavrec and
Dzeroski 1997) to learn the relation-
ships between these duties. In agree-
ment with Marcello Guarini and
Baruch Brody (1988) that casuistry
alone is not sufficient, we begin with
prima facie duties that often give con-
flicting advice in ethical dilemmas and
then abstract a decision principle,
when conflicts do arise, from cases of
ethical dilemmas where ethicists are in
agreement as to the correct action. We
have adopted a multiple prima facie
duty approach to ethical decision
making because we believe it is more
likely to capture the complexities of
ethical decision making than a single,
absolute duty ethical theory. In an
attempt to develop a decision proce-
dure for determining the ethically cor-
rect action when the duties give con-
flicting advice, we use ILP to abstract
information leading to a general deci-
sion principle from ethical experts’
intuitions about particular ethical
dilemmas. A common criticism is
whether the relatively straightforward
representation scheme used to repre-
sent ethical dilemmas will be suffi-
cient to represent a wider variety of
cases in different domains.
Deontic logic’s formalization of the
notions of obligation, permission, and
related concepts1make it a prime can-
didate as a language for the expression
of machine ethics principles. Selmer
Bringsjord, Konstantine Arkoudas,
and Paul Bello (2006) show how for-
mal logics of action, obligation, and
permissibility might be used to incor-
porate a given set of ethical principles
into the decision procedure of an
autonomous system. They contend
that such logics would allow for proofs
establishing that (1) robots only take
permissible actions, and (2) all actions
that are obligatory for robots are actu-
ally performed by them, subject to ties
and conflicts among available actions.
They further argue that, while some
may object to the wisdom of logic-
based AI in general, they believe that
in this case a logic-based approach is
promising because one of the central
issues in machine ethics is trust and
“mechanized formal proofs are per-
haps the single most effective tool at
our disposal for establishing trust.”
Making no commitment as to the eth-
ical content, their objective is to arrive
at a methodology that maximizes the
probability that an artificial intelligent
agent behaves in a certifiably ethical
fashion, subject to proof explainable
in ordinary English. They propose a
general methodology for implement-
ing deontic logics in their logical
framework, Athena, and illustrate the
feasibility of this approach by encod-
ing a natural deduction system for a
deontic logic for reasoning about what
agents ought to do. Concerns remain
regarding the practical relevance of
the formal logics they are investigat-
ing and efficiency issues in their
The work of Bringjord, Arkoudas,
and Bello is based on research that
investigates, from perspectives other
than artificial intelligence, how deon-
tic logic’s concern with what ought to
be the case might be extended to repre-
sent and reason about what agents
ought to do. It has been argued that the
implied assumption that the latter will
simply follow from investigation of
the former is not the case. In this con-
text, John Horty (2001) proposes an
extension of deontic logic, incorporat-
ing a formal theory of agency that
describes what agents ought to do
under various conditions over extend-
ed periods of time. In particular, he
adapts preference ordering from deci-
ers or external circumstances, such as a
lack of funds) and sufficiently free of
internal constraints (for example, pain
or discomfort, the effects of medica-
tion, irrational fears, or values that are
likely to change over time). The prin-
ciple of nonmaleficence requires that
the health-care professional not harm
the patient, while the principle of
beneficence states that the health-care
professional should promote patient
welfare. Finally, the principle of justice
states that health-care services and
burdens should be distributed in a just
Step Two
The domain we selected was medical
ethics, consistent with our choice of
prima facie duties, and, in particular, a
representative type of ethical dilemma
that involves three of the four princi-
ples of biomedical ethics: respect for
autonomy, nonmaleficence, and bene -
fi cence. The type of dilemma is one
that health-care workers often face: A
health-care worker has recommended
a particular treatment for her compe-
tent adult patient, and the patient has
rejected that treatment option. Should
the health-care worker try again to
change the patient’s mind or accept
the patient’s decision as final? The
dilemma arises because, on the one
hand, the health-care professional
shouldn’t challenge the patient’s
autonomy unnecessarily; on the other
hand, the health-care worker may
have concerns about why the patient
is refusing the treatment.
In this type of dilemma, the options
for the health-care worker are just two,
either to accept the patient’s decision
or not, by trying again to change the
patient’s mind. For this proof of con-
cept test of attempting to make a pri-
ma facie duty ethical theory com-
putable, we have a single type of
dilemma that encompasses a finite
number of specific cases, just three
duties, and only two possible actions
in each case. We have abstracted, from
a discussion of similar types of cases
given by Buchanan and Brock (1989),
the correct answers to the specific cas-
es of the type of dilemma we consider.
We have made the assumption that
there is a consensus among bioethi-
cists that these are the correct answers.
sion theory to “both define optimal
actions that an agent should perform
and the propositions whose truth the
agent should guarantee.” This frame-
work permits the uniform formaliza-
tion of a variety of issues of ethical
theory and, hence, facilitates the dis-
cussion of these issues.
Tom Powers (2006) assesses the fea-
sibility of using deontic and default
logics to implement Kant’s categorical
Act only according to that maxim
whereby you can at the same time
will that it should become a universal
law… If contradiction and contrast
arise, the action is rejected; if harmo-
ny and concord arise, it is accepted.
From this comes the ability to take
moral positions as a heuristic means.
For we are social beings by nature,
and what we do not accept in others,
we cannot sincerely accept in our-
Powers suggests that a machine
might itself construct a theory of
ethics by applying a universalization
step to individual maxims, mapping
them into the deontic categories of
forbidden, permissible, or obligatory
actions. Further, for consistency, these
universalized maxims need to be test-
ed for contradictions with an already
established base of principles, and
these contradictions resolved. Powers
suggests, further, that such a system
will require support from a theory of
commonsense reasoning in which
postulates must “survive the occasion-
al defeat,” thus producing a nonmo-
notonic theory whose implementa-
tion will require some form of default
reasoning. It has been noted (Ganascia
2007) that answer set programming
(ASP) (Baral 2003) may serve as an effi-
cient formalism for modeling such
ethical reasoning. An open question is
what reason, other than temporal pri-
ority, can be given for keeping the
whole set of prior maxims and disal-
lowing a new contradictory one. Pow-
ers offers that “if we are to construe
Kant’s test as a way to build a set of
maxims, we must establish rules of pri-
ority for accepting each additional
maxim.” The question remains as to
what will constitute this moral epis-
temic commitment.
Creating a Machine That
Is an Explicit Ethical Agent
To demonstrate the possibility of cre-
ating a machine that is an explicit eth-
ical agent, we have attempted in our
research to complete the following six
Step One
We have adopted the prima facie duty
approach to ethical theory, which, as
we have argued, better reveals the com-
plexity of ethical decision making than
single, absolute duty theories. It incor-
porates the good aspects of the teleo-
logical and deontological approaches
to ethics, while allowing for needed
exceptions to adopting one or the oth-
er approach exclusively. It also has the
advantage of being better able to adapt
to the specific concerns of ethical
dilemmas in different domains. There
may be slightly different sets of prima
facie duties for biomedical ethics, legal
ethics, business ethics, and journalistic
ethics, for example.
There are two well-known prima
facie duty theories: Ross’s theory, deal-
ing with general ethical dilemmas,
that has seven duties; and Beauchamp
and Childress’s four principles of bio-
medical ethics (1979) (three of which
are derived from Ross’s theory) that
are intended to cover ethical dilem-
mas specific to the field of biomedi-
cine. Because there is more agreement
between ethicists working on biomed-
ical ethics than in other areas, and
because there are fewer duties, we
decided to begin to develop our prima
facie duty approach to computing
ethics using Beauchamp and Chil-
dress’s principles of biomedical ethics.
Beauchamp and Childress’s princi-
ples of biomedical ethics include the
principle of respect for autonomy that
states that the health-care profession-
al should not interfere with the effec-
tive exercise of patient autonomy. For
a decision by a patient concerning his
or her care to be considered fully
autonomous, it must be intentional,
based on sufficient understanding of
his or her medical situation and the
likely consequences of forgoing treat-
ment, sufficiently free of external con-
straints (for example, pressure by oth-
Step Three
The major philosophical problem with
the prima facie duty approach to ethi-
cal decision making is the lack of a
decision procedure when the duties
give conflicting advice. What is need-
ed, in our view, are ethical principles
that balance the level of satisfaction or
violation of these duties and an algo-
rithm that takes case profiles and out-
puts that action that is consistent with
these principles. A profile of an ethical
dilemma consists of an ordered set of
numbers for each of the possible
actions that could be performed,
where the numbers reflect whether
particular duties are satisfied or violat-
ed and, if so, to what degree. John
Rawls’s “reflective equilibrium” (1951)
approach to creating and refining eth-
ical principles has inspired our solu-
tion to the problem of a lack of a deci-
sion procedure. We abstract a principle
from the profiles of specific cases of
ethical dilemmas where experts in
ethics have clear intuitions about the
correct action and then test the prin-
ciple on other cases, refining the prin-
ciple as needed.
The selection of the range of possi-
ble satisfaction or violation levels of a
particular duty should, ideally, depend
upon how many gradations are need-
ed to distinguish between cases that
are ethically distinguishable. Further,
it is possible that new duties may need
to be added in order to make distinc-
tions between ethically distinguish-
able cases that would otherwise have
the same profiles. There is a clear
advantage to our approach to ethical
decision making in that it can accom-
modate changes to the range of inten-
sities of the satisfaction or violation of
duties, as well as adding duties as
Step Four
Implementing the algorithm for the
theory required formulation of a prin-
ciple to determine the correct action
when the duties give conflicting
advice. We developed a system (Ander-
son, Anderson, and Armen 2006a)
that uses machine-learning tech-
niques to abstract relationships be -
tween the prima facie duties from par-
ticular ethical dilemmas where there is
an agreed-upon correct action. Our
WINTER 2007 23
chosen type of dilemma, detailed pre-
viously, has only 18 possible cases
(given a range of +2 to –2 for the level
of satisfaction or violation of the
duties) where, given the two possible
actions, the first action supersedes the
second (that is, was ethically prefer-
able). Four of these cases were provid-
ed to the system as examples of when
the target predicate (supersedes) is
true. Four examples of when the target
predicate is false were provided by
simply reversing the order of the
actions. The system discovered a prin-
ciple that provides the correct answer
for the remaining 14 positive cases, as
verified by the consensus of ethicists.
ILP was used as the method of learn-
ing in this system. ILP is concerned
with inductively learning relations
represented as first-order Horn clauses
(that is, universally quantified con-
junctions of positive literals Liimply-
ing a positive literal H: H (L1
Ln)). ILP is used to learn the relation
supersedes (A1, A2), which states that
action A1 is preferred over action A2
in an ethical dilemma involving these
choices. Actions are represented as
ordered sets of integer values in the
range of +2 to –2 where each value
denotes the satisfaction (positive val-
ues) or violation (negative values) of
each duty involved in that action.
Clauses in the supersedes predicate are
represented as disjunctions of lower
bounds for differentials of these val-
ILP was chosen to learn this relation
for a number of reasons. The poten-
tially nonclassical relationships that
might exist between prima facie duties
are more likely to be expressible in the
rich representation language provided
by ILP than in less expressive repre-
sentations. Further, the consistency of
a hypothesis regarding the relation-
ships between prima facie duties can
be automatically confirmed across all
cases when represented as Horn claus-
es. Finally, commonsense background
knowledge regarding the supersedes
relationship is more readily expressed
and consulted in ILP’s declarative rep-
resentation language.
The object of training is to learn a
new hypothesis that is, in relation to
all input cases, complete and consis-
tent. Defining a positive example as a
case in which the first action super-
sedes the second action and a negative
example as one in which this is not
the case, a complete hypothesis is one
that covers all positive cases, and a
consistent hypothesis covers no nega-
tive cases. Negative training examples
are generated from positive training
examples by inverting the order of
these actions, causing the first action
to be the incorrect choice. The system
starts with the most general hypothe-
sis stating that all actions supersede
each other and, thus, covers all posi-
tive and negative cases. The system is
then provided with positive cases (and
their negatives) and modifies its
hypothesis, by adding or refining
clauses, such that it covers given posi-
tive cases and does not cover given
negative cases.
The decision principle that the sys-
tem discovered can be stated as fol-
lows: A health-care worker should
challenge a patient’s decision if it isn’t
fully autonomous and there’s either
any violation of nonmaleficence or a
severe violation of beneficence.
Although, clearly, this rule is implicit
in the judgments of the consensus of
ethicists, to our knowledge this prin-
ciple has never before been stated
explicitly. Ethical theory has not yet
advanced to the point where princi-
ples like this one—that correctly bal-
ance potentially conflicting duties
with differing levels of satisfaction or
violation—have been formulated. It is
a significant result that machine-
learning techniques can discover a
principle such as this and help
advance the field of ethics. We offer it
as evidence that making the ethics
more precise will permit machine-
learning techniques to discover philo-
sophically novel and interesting prin-
ciples in ethics because the learning
system is general enough that it can
be used to learn relationships between
any set of prima facie duties where
there is a consensus among ethicists
as to the correct answer in particular
Once the principle was discovered,
the needed decision procedure could
be fashioned. Given a profile repre-
senting the satisfaction/violation lev-
els of the duties involved in each pos-
sible action, values of corresponding
duties are subtracted (those of the sec-
ond action from those of the first).
The principle is then consulted to see
if the resulting differentials satisfy any
of its clauses. If so, the first action of
the profile is deemed ethically prefer-
able to the second.
Step Five
We have explored two prototype
applications of the discovered princi-
ple governing Beauchamp and Chil-
dress’s principles of biomedical ethics.
In both prototypes, we created a pro-
gram where a machine could use the
principle to determine the correct
answer in ethical dilemmas. The first,
MedEthEx (Anderson, Anderson, and
Armen 2006b), is a medical ethical
advisor system; the second, EthEl, is a
system in the domain of elder care
that determines when a patient should
be reminded to take medication and
when a refusal to do so is serious
enough to contact an overseer. EthEl is
more autonomous than MedEthEx in
that, whereas MedEthEx gives the eth-
ically correct answer (that is, that
which is consistent with its training)
to a human user who will act on it or
not, EthEl herself acts on what she
determines to be the ethically correct
MedEthEx is an expert system that
uses the discovered principle and deci-
sion procedure to give advice to a user
faced with a case of the dilemma type
previously described. In order to per-
mit use by someone unfamiliar with
the representation details required by
the decision procedure, a user inter-
face was developed that (1) asks ethi-
cally relevant questions of the user
regarding the particular case at hand,
(2) transforms the answers to these
questions into the appropriate pro-
files, (3) sends these profiles to the
decision procedure, (4) presents the
answer provided by the decision pro-
cedure, and (5) provides a justification
for this answer.2
The principle discovered can be
used by other systems, as well, to pro-
vide ethical guidance for their actions.
Our current research uses the principle
to elicit ethically sensitive behavior
from an elder-care system, EthEl, faced
with a different but analogous ethical
dilemma. EthEl must remind the
patient to take his or her medication
and decide when to accept a patient's
refusal to take a medication that might
prevent harm or provide benefit to the
patient and when to notify an over-
seer. This dilemma is analogous to the
original dilemma in that the same
duties are involved (nonmaleficence,
beneficence, and respect for autono-
my) and “notifying the overseer” in
the new dilemma corresponds to “try-
ing again” in the original.
Machines are currently in use that
face this dilemma.3The state of the art
in these reminder systems entails pro-
viding “context-awareness” (that is, a
characterization of the current situa-
tion of a person) to make reminders
more efficient and natural. Unfortu-
nately, this awareness does not
include consideration of ethical duties
that such a system should adhere to
when interacting with its patient. In
an ethically sensitive elder-care sys-
tem, both the timing of reminders and
responses to a patient’s disregard of
them should be tied to the duties
involved. The system should chal-
lenge patient autonomy only when
necessary, as well as minimize harm
and loss of benefit to the patient. The
principle discovered from the original
dilemma can be used to achieve these
goals by directing the system to
remind the patient only at ethically
justifiable times and notifying the
overseer only when the harm or loss of
benefit reaches a critical level.
In the implementation, EthEl
receives input from an overseer (most
likely a doctor), including: the pre-
scribed time to take a medication, the
maximum amount of harm that could
occur if this medication is not taken
(for example, none, some, or consider-
able), the number of hours it would
take for this maximum harm to occur,
the maximum amount of expected
good to be derived from taking this
medication, and the number of hours
it would take for this benefit to be lost.
The system then determines from this
input the change in duty satisfaction
and violation levels over time, a func-
tion of the maximum amount of harm
or good and the number of hours for
this effect to take place. This value is
used to increment duty satisfaction and
violation levels for the remind action
and, when a patient disregards a
reminder, the notify action. It is used to
decrement don’t remind and don’t notify
actions as well. A reminder is issued
when, according to the principle, the
duty satisfaction or violation levels
have reached the point where remind-
ing is ethically preferable to not
reminding. Similarly, the overseer is
notified when a patient has disregarded
reminders to take medication and the
duty satisfaction or violation levels
have reached the point where notify-
ing the overseer is ethically preferable
to not notifying the overseer.
EthEl uses an ethical principle dis-
covered by a machine to determine
reminders and notifications in a way
that is proportional to the amount of
maximum harm to be avoided or good
to be achieved by taking a particular
medication, while not unnecessarily
challenging a patient’s autonomy.
EthEl minimally satisfies the require-
ments of an explicit ethical agent (in a
constrained domain), according to Jim
Moor’s definition of the term: A
machine that is able to calculate the
best action in ethical dilemmas using
an ethical principle, as opposed to
having been programmed to behave
ethically, where the programmer is fol-
lowing an ethical principle.
Step Six
As a possible means of assessing the
morality of a system’s behavior, Colin
Allen, G. Varner, and J. Zinser (2000)
describe a variant of the test Alan Tur-
ing (1950) suggested as a means to
determine the intelligence of a
machine that bypassed disagreements
about the definition of intelligence.
Their proposed “comparative moral
Turing test” (cMTT) bypasses disagree-
ment concerning definitions of ethical
behavior as well as the requirement
that a machine have the ability to
articulate its decisions: an evaluator
assesses the comparative morality of
pairs of descriptions of morally signif-
icant behavior where one describes the
actions of a human being in an ethical
dilemma and the other the actions of
a machine faced with the same dilem-
ma. If the machine is not identified as
the less moral member of the pair sig-
nificantly more often than the
human, then it has passed the test.
WINTER 2007 25
They point out, though, that human
behavior is typically far from being
morally ideal and a machine that
passed the cMTT might still fall far
below the high ethical standards to
which we would probably desire a
machine to be held. This legitimate
concern suggests to us that, instead of
comparing the machine’s behavior in
a particular dilemma against typical
human behavior, the comparison
ought to be made with behavior rec-
ommended by a trained ethicist faced
with the same dilemma. We also
believe that the principles used to jus-
tify the decisions that are reached by
both the machine and ethicist should
be made transparent and compared.
We plan to devise and carry out a
moral Turing test of this type in future
work, but we have had some assess-
ment of the work that we have done
to date. The decision principle that
was discovered in MedEthEx, and used
by EthEl, is supported by W. D. Ross’s
claim that it is worse to harm than not
to help someone. Also, the fact that
the principle provided answers to
nontraining cases that are consistent
with Buchanan and Brock’s judgments
offers preliminary support for our
hypothesis that decision principles
discovered from some cases, using our
method, enable a machine to deter-
mine the ethically acceptable action in
other cases as well.
We have argued that machine ethics is
an important new field of artificial
intelligence and that its goal should be
to create machines that are explicit
ethical agents. We have done prelimi-
nary work to show—through our
proof of concept applications in con-
strained domains—that it may be pos-
sible to incorporate an explicit ethical
component into a machine. Ensuring
that a machine with an ethical com-
ponent can function autonomously in
the world remains a challenge to
researchers in artificial intelligence
who must further investigate the rep-
resentation and determination of eth-
ical principles, the incorporation of
these ethical principles into a system’s
decision procedure, ethical decision
making with incomplete and uncer-
tain knowledge, the explanation for
decisions made using ethical princi-
ples, and the evaluation of systems
that act based upon ethical principles.
Of the many challenges facing
those who choose to work in the area
of machine ethics, foremost is the
need for a dialogue between ethicists
and researchers in artificial intelli-
gence. Each has much to gain from
working together on this project. For
ethicists, there is the opportunity of
clarifying—perhaps even discover-
ing—the fundamental principles of
ethics. For AI researchers, convincing
the general public that ethical
machines can be created may permit
continued support for work leading to
the development of autonomous
intelligent machines—machines that
might serve to improve the lives of
human beings.
This material is based upon work sup-
ported in part by the National Science
Foundation grant number IIS-
1. See
2. A demonstration of MedEthEx is avail-
able online at
3. For example, see
Allen, C.; Varner, G.; and Zinser, J. 2000.
Prolegomena to Any Future Artificial Moral
Agent. Journal of Experimental and Theoreti-
cal Artificial Intelligence 12(2000): 251–61.
Anderson, M., and Anderson, S., eds. 2006.
Special Issue on Machine Ethics. IEEE Intel-
ligent Systems 21(4) (July/August).
Anderson, M.; Anderson, S.; and Armen,
C., eds. 2005a. Machine Ethics: Papers from
the AAAI Fall Symposium. Technical Report
FS-05-06, Association for the Advancement
of Artificial Intelligence, Menlo Park, CA.
Anderson, M.; Anderson, S.; and Armen, C.
2005b. Toward Machine Ethics: Imple-
menting Two Action-Based Ethical Theo-
ries. In Machine Ethics: Papers from the
AAAI Fall Symposium. Technical Report FS-
05-06, Association for the Advancement of
Artificial Intelligence, Menlo Park, CA.
Anderson, M.; Anderson, S.; and Armen, C.
2006a. An Approach to Computing Ethics.
IEEE Intelligent Systems 21(4): 56–63.
Anderson, M.; Anderson, S.; and Armen, C.
2006b. MedEthEx: A Prototype Medical
Ethics Advisor. In Proceedings of the Eigh-
teenth Conference on Innovative Applications
of Artificial Intelligence. Menlo Park, CA:
AAAI Press.
Anderson, S. L. 1995. Being Morally
Responsible for an Action Versus Acting
Responsibly or Irresponsibly. Journal of
Philosophical Research 20: 453–62.
Asimov, I. 1976. The Bicentennial Man. In
Stellar Science Fiction 2, ed. J.-L. del Rey. New
York: Ballatine Books.
Baral, C. 2003. Knowledge Representation,
Reasoning, and Declarative Problem Solving.
Cambridge, UK: Cambridge University
Beauchamp, T. L., and Childress, J. F. 1979.
Principles of Biomedical Ethics. Oxford, UK:
Oxford University Press.
Bentham, J. 1907. An Introduction to the
Principles and Morals of Legislation. Oxford:
Clarendon Press.
Bringsjord, S.; Arkoudas, K.; and Bello, P.
2006. Toward a General Logicist Methodol-
ogy for Engineering Ethically Correct
Robots. IEEE Intelligent Systems 21(4): 38–
Brody, B. 1988. Life and Death Decision Mak-
ing. New York: Oxford University Press.
Buchanan, A. E., and Brock, D. W. 1989.
Deciding for Others: The Ethics of Surrogate
Decision Making, 48–57. Cambridge, UK:
Cambridge University Press.
Capek, K. 1921. R.U.R. In Philosophy and
Science Fiction, ed. M. Phillips. Amherst, NY:
Prometheus Books.
Clarke, A. C. 1968. 2001: A Space Odyssey.
New York: Putnam.
Damasio, A.R. 1994. Descartes’ Error: Emo-
tion, Reason, and the Human Brain. New
York: G. P. Putnam.
Dennett, D. 2006. Computers as Prostheses
for the Imagination. Invited talk presented
at the International Computers and Philos-
ophy Conference, Laval, France, May 3.
Dietrich, E. 2006. After the Humans Are
Gone. Keynote address presented at the 2006
North American Computing and Philosophy
Conference, RPI, Troy, NY, August 12.
Ganascia, J. G. 2007. Using Non-Monoton-
ic Logics to Model Machine Ethics. Paper
presented at the Seventh International
Computer Ethics Conference, San Diego,
CA, July 12–14.
Gazzaniga, M. 2006. The Ethical Brain: The
Science of Our Moral Dilemmas. New York:
Harper Perennial.
Guarini, M. 2006. Particularism and the
Classification and Reclassification of Moral
Cases. IEEE Intelligent Systems 21(4): 22–28.
Horty, J. 2001. Agency and Deontic Logic.
New York: Oxford University Press.
Joy, B. 2000. Why the Future Doesn’t Need
Us. Wired Magazine 8(04) (April).
Kant, I. 1785. Groundwork of the Metaphysic
of Morals, trans. by H. J. Paton (1964). New
York: Harper & Row.
Lavrec, N., and Dzeroski, S. 1997. Inductive
Logic Programming: Techniques and Applica-
tions. Chichester, UK: Ellis Horwood.
Mappes, T. A., and DeGrazia, D. 2001. Bio-
medical Ethics, 5th ed., 39–42. New York:
McLaren, B. M. 2003. Extensionally Defin-
ing Principles and Cases in Ethics: An AI
Model. Artificial Intelligence Journal 150(1–
2): 145–1813.
Moor, J. H. 2006. The Nature, Importance,
and Difficulty of Machine Ethics. IEEE Intel-
ligent Systems 21(4): 18–21.
Onishi, N. 2006. In a Wired South Korea,
Robots Will Feel Right at Home. New York
Times, April 2, 2006.
Picard, R. W. 1997. Affective Computing.
Cambridge, MA: The MIT Press.
Pojman, L. J. 1996. The Case for Moral
Objectivism. In Do the Right Thing: A Philo-
sophical Dialogue on the Moral and Social
Issues of Our Time, ed. F. J. Beckwith. New
York: Jones and Bartlett.
Powers, T. 2006. Prospects for a Kantian
Machine. IEEE Intelligent Systems 21(4): 46–
Rawls, J. 1951. Outline for a Decision Pro-
cedure for Ethics. The Philosophical Review
60(2): 177–197.
Ross, W. D. 1930. The Right and the Good.
Oxford: Clarendon Press.
Rzepka, R., and Araki, K. 2005. What Could
Statistics Do for Ethics? The Idea of a Com-
mon Sense Processing-Based Safety Valve.
In Machine Ethics: Papers from the AAAI
Fall Symposiu. Technical Report FS-05-06,
Association for the Advancement of Artifi-
cial Intelligence, Menlo Park, CA.
Turing, A. M. 1950. Computing Machinery
and Intelligence. Mind LIX(236): 433–460.
Michael Anderson is an
associate professor of
computer science at the
University of Hartford,
West Hartford, Con-
necticut. He earned his
Ph.D. in computer sci-
ence and engineering at
the University of Con-
necticut. His interest in further enabling
machine autonomy brought him first to
diagrammatic reasoning where he co -
chaired Diagrams 2000, the first conference
on the topic. This interest has currently led
him, in conjuction with Susan Leigh
Anderson, to establish machine ethics as a
bona fide field of study. He has cochaired
the AAAI Fall 2005 Symposium on Machine
Ethics and coedited an IEEE Intelligent Sys-
tems special issue on machine ethics in
2006. His research in machine ethics was
selected for IAAI as an emerging applica-
tion in 2006. He maintains the machine
ethics website (
and can be reached at anderson@hart-
Susan Leigh Anderson,
a professor of philoso-
phy at the University of
Connecticut, received
her Ph.D. in philosophy
at UCLA. Her specialty is
applied ethics, most
recently focusing on bio-
medical ethics and
machine ethics. She has received funding
from NEH, NASA, and NSF. She is the
author of three books in the Wadsworth
Philosophers Series, as well as numerous
articles. With Michael Anderson, she has
presented work on machine ethics at
national and international conferences,
organized and cochaired the AAAI Fall
2005 Symposium on Machine Ethics, and
coedited a special issue of IEEE Intelligent
Systems on machine ethics (2006). She can
be contacted at Susan.Anderson@uconn.
Proceedings of the Twenty-Second AAAI Conference
on Artificial Intelligence
July, 2007 Vancouver, British Columbia, Canada
2 vols., references, index, illus., ISBN 978-1-57735-323-2
    • This paper is a result of a week long discussion among experts of different fields within the scope of the Dagstuhl Seminar 16222 2 , and incorporates various ideas that arose as a result of discussions among interdisciplinary experts. Position papers that focus on machine ethics as a whole have been produced and they offer interesting insights in the problem as a whole, see for example [36, 3, 8, 4], but to the best of our knowledge, this is the only document devoted specifically to the problem of engineering machine ethics.
    [Show abstract] [Hide abstract] ABSTRACT: Both the ethics of autonomous systems and the problems of their technical implementation have by now been studied in some detail. Less attention has been given to the areas in which these two separate concerns meet. This paper, written by both philosophers and engineers of autonomous systems, addresses a number of issues in machine ethics that are located at precisely the intersection between ethics and engineering. We first discuss different approaches towards the conceptual design of autonomous systems and their implications on the ethics implementation in such systems. Then we examine problematic areas regarding the specification and verification of ethical behavior in autonomous systems, particularly with a view towards the requirements of future legislation. We discuss transparency and accountability issues that will be crucial for any future wide deployment of autonomous systems in society. Finally we consider the, often overlooked, possibility of intentional misuse of AI systems and the possible dangers arising out of deliberately unethical design, implementation, and use of autonomous robots.
    Full-text · Article · Mar 2017 · Robotics and Autonomous Systems
    • For example, a highly autonomous robot could be said to have some causal responsibility because it is capable of making decisions that cause actions in a broad range of environments whereas the actions of a robot with low autonomy, typically caused by the human that controls it, would not. Computers may be superior to humans in terms of the accuracy and quality of their decisions [4] because they have a greater ability to calculate all the consequences of an action that may be performed in a certain situation. So, with these innately superior capacities they might perform their social tasks [27] both perfectly and accurately and therefore be able to be more responsible than humans — at least in the sense of role responsibility — in so far as they are better able to perform tasks effectively.
    [Show abstract] [Hide abstract] ABSTRACT: Autonomous robots such as self-driving cars are already able to make decisions that have ethical consequences. As such machines make increasingly complex and important decisions, we will need to know that their decisions are trustworthy and ethically justified. Hence we will need them to be able to explain the reasons for these decisions: ethical decision-making requires that decisions be explainable with reasons. We argue that for people to trust autonomous robots we need to know which ethical principles they are applying and that their application is deterministic and predictable. If a robot is a self-improving, self-learning type of robot whose choices and decisions are based on past experience, which decision it makes in any given situation may not be entirely predictable ahead of time or explainable after the fact. This combination of non-predictability and autonomy may confer a greater degree of responsibility to the machine but it also makes them harder to trust.
    Full-text · Conference Paper · Nov 2016 · Robotics and Autonomous Systems
    • Neither extreme is entirely satisfactory due to the complexity of oversight and the volatility of decision-making structures. Beyond the nature of moral agency in machines, work in machine ethics also investigates how best to design moral reasoning and behaviours into autonomous algorithms as artificial moral and ethical agents 18 (Anderson and Anderson, 2007; Crnkovic and C ¸ u¨ru¨klu¨,u¨ru¨u¨ru¨klu¨u¨ru¨klu¨, 2011; Sullins, 2006; Wiegel and Berg, 2009 ). Research into this question remains highly relevant because algorithms can be required to make realtime decisions involving ''difficult trade-offs . . .
    [Show abstract] [Hide abstract] ABSTRACT: In information societies, operations, decisions and choices previously left to humans are increasingly delegated to algorithms, which may advise, if not decide, about how data should be interpreted and what actions should be taken as a result. More and more often, algorithms mediate social processes, business transactions, governmental decisions, and how we perceive, understand, and interact among ourselves and with the environment. Gaps between the design and operation of algorithms and our understanding of their ethical implications can have severe consequences affecting individuals as well as groups and whole societies. This paper makes three contributions to clarify the ethical importance of algorithmic mediation. It provides a prescriptive map to organise the debate. It reviews the current discussion of ethical aspects of algorithms. And it assesses the available literature in order to identify areas requiring further work to develop the ethics of algorithms.
    Full-text · Article · Oct 2016
    • Let us evaluate whether these criteria can be satisfied for a particular set of ethical rules in the domain of robot assistance. Anderson and Anderson [1] proposed a set of ethical rules for a robot charged with reminding a patient to take her medication . In particular, they considered what the robot should do in case she refuses to take the medicine.
    [Show abstract] [Hide abstract] ABSTRACT: Robots are becoming ever more autonomous. This expanding ability to take unsupervised decisions renders it imperative that mechanisms are in place to guarantee the safety of behaviours executed by the robot. Moreover, smart autonomous robots should be more than safe; they should also be explicitly ethical -- able to both choose and justify actions that prevent harm. Indeed, as the cognitive, perceptual and motor capabilities of robots expand, they will be expected to have an improved capacity for making moral judgements. We present a control architecture that supplements existing robot controllers. This so-called Ethical Layer ensures robots behave according to a predetermined set of ethical rules by predicting the outcomes of possible actions and evaluating the predicted outcomes against those rules. To validate the proposed architecture, we implement it on a humanoid robot so that it behaves according to Asimov's laws of robotics. In a series of four experiments, using a second humanoid robot as a proxy for the human, we demonstrate that the proposed Ethical Layer enables the robot to prevent the human from coming to harm.
    Full-text · Article · Sep 2016 · Robotics and Autonomous Systems
    • The remaining three articles focus on machine ethics as a research field. Colin Allen, Wendell Wallach, and Ida Smith paper's (Allen et al. 2006) as well as Michael and Susan Anderson's papers (Anderson & Anderson, 2007) both offer seminal contributions in delineating the scope of the field and its relevance both within and outside academia. As Anderson and Anderson stress in their papers, " machine ethics is concerned with ensuring that the behavior of machines toward human users, and perhaps other machines as well, is ethically acceptable " .
    Full-text · Chapter · Feb 2016 · Robotics and Autonomous Systems
    • Such ethical requirements may prove essential for an autonomous system to be certified by a regulatory body, since ethical autonomy is obviously desirable. Machine ethics is an emerging discipline concerned with ensuring that the behaviour of machines towards humans and other machines they interact with is ethical [1]. It is an open question whether machines are, or will ever be, moral agents, i.e., in possession of an innate ability to distinguish between right or wrong.
    [Show abstract] [Hide abstract] ABSTRACT: Autonomous systems such as unmanned vehicles are beginning to operate within society. All participants in society are required to follow specific regulations and laws. An autonomous system cannot be an exception. Inevitably an autonomous system will find itself in a situation in which it needs to not only choose to obey a rule or not, but also make a complex ethical decision. However, there exists no obvious way to implement the human understanding of ethical behaviour in computers. Even if we enable autonomous systems to distinguish between more and less ethical alternatives, how can we be sure that they would choose right? We consider autonomous systems with a hybrid architecture in which the highest level of reasoning is executed by a rational (BDI) agent. For such a system, formal verification has been used successfully to prove that specific rules of behaviour are observed when making decisions. We propose a theoretical framework for ethical plan selection that can be formally verified. We implement a rational agent that incorporates a given ethical policy in its plan selection and show that we can formally verify that the agent chooses to execute, to the best of its beliefs, the most ethical available plan.
    Full-text · Article · Dec 2015
Show more