Content uploaded by Michael Anderson
Author content
All content in this area was uploaded by Michael Anderson
Content may be subject to copyright.
MedEthEx:
A Prototype Medical Ethics Advisor
Michael Anderson1, Susan Leigh Anderson2, Chris Armen3
1University of Hartford, 2University of Connecticut, 3Amherst College
1Dept. of Computer Science, 200 Bloomfield Avenue, West Hartford, CT 06776
2Dept. of Philosophy, 1 University Place, Stamford, CT 06901
3Dept. of Mathematics and Computer Science, Amherst, MA 01002
anderson@hartford.edu, susan.anderson@uconn.edu, carmen@amherst.edu
Abstract
As part of a larger Machine Ethics Project, we are
developing an ethical advisor that provides guidance to
health care workers faced with ethical dilemmas.
MedEthEx is an implementation of Beauchamp’s and
Childress’ Principles of Biomedical Ethics that harnesses
machine learning techniques to abstract decision principles
from cases in a particular type of dilemma with conflicting
prima facie duties and uses these principles to determine
the correct course of action in similar and new cases. We
believe that accomplishing this will be a useful first step
towards creating machines that can interact with those in
need of health care in a way that is sensitive to ethical
issues that may arise.
Introduction
Past research concerning the relationship between
technology and ethics has largely focused on responsible
and irresponsible use of technology by human beings, with
a few people being interested in how human beings ought
to treat machines. In all cases, only human beings have
engaged in ethical reasoning. We believe that the time
has come for adding an ethical dimension to at least some
machines. Recognition of the ethical ramifications of
behavior involving machines, recent and potential
developments in machine autonomy, as well as the
possibility of harnessing machine intelligence to aid
humans in ethical decision making, all support this
position. We explore adding an ethical dimension to
machines through what has been called machine ethics
(Anderson et al. 2004). In contrast to software property
issues, privacy issues and other topics normally ascribed to
computer ethics, machine ethics is concerned with the
behavior of machines towards human users and other
machines.
In order to create ethically sensitive machines, we need
a computable ethical theory. A long-term objective of our
work is to further research in both applied and theoretical
Ethics via application of techniques from research in
Artificial Intelligence. Ethics, by its very nature, is a
branch of Philosophy that must have practical application,
so we believe that we can advance the study of Ethical
Theory by
attempting to work out the details needed to apply a
proposed ethical theory to particular ethical dilemmas. In
this way, we can best determine whether the theory can be
made consistent, complete, practical and agree with
intuition, essential criteria that any good (action-based)
ethical theory must satisfy (Anderson 1999).
Currently, we are investigating the feasibility of systems
that can act as ethical advisors, providing guidance to
users faced with ethical dilemmas. To this end, we are
developing a system that provides such guidance in the
domain of health care. Healthcare workers and researchers
using human subjects face many ethical dilemmas in their
practices, yet it is not clear that all are equipped to think
through the ethically relevant dimensions of these
dilemmas to the extent that they feel confident about the
decisions that they make and act upon. In the absence of
having an ethicist at hand, a system that provides
guidance in such dilemmas might prove useful.
MedEthEx, our current effort in this vein, is a system that
extracts and analyzes ethically relevant information about
a biomedical ethical dilemma from the health care worker
or researcher to help decide the best course of action. This
project allows us to explore the computability of ethics in a
limited domain. We believe that creating an ethical
advisor, such as MedEthEx, will be a useful first step
towards creating machines that can interact with those in
need of health care in a way that is sensitive to ethical
issues that may arise. It can also function as a model for
creating machines that can follow more general ethical
principles, ones that can function in any domain.
Philosophical Foundations
MedEthEx is based upon a well-known multiple duty
ethical theory that is tailored to problems in biomedical
ethics: Tom L. Beauchamp’s and James F. Childress’
Principles of Biomedical Ethics (1979). There are four
duties or principles in this theory – the Principle of
__________
Copyright © 2006, American Association for Artificial Intelligence
(www.aaai.org). All rights reserved.
1759
Respect for Autonomy, the Principle of Nonmaleficence,
the Principle of Beneficence and the Principle of Justice –
that are each considered to be prima facie duties. A prima
facie duty is not absolute, but rather is thought of as an
obligation to which we should adhere unless it is
overridden by a stronger obligation (i.e. one of the other
duties). To elaborate upon each of the four duties that
form the Principles of Biomedical Ethics: The Principle of
Autonomy (A) states that the health care professional
should not interfere with the effective exercise of patient
autonomy. For a decision by a patient concerning his/her
care to be considered fully autonomous, it must be
intentional, based on sufficient understanding of his/her
medical situation and the likely consequences of foregoing
treatment, sufficiently free of external constraints (e.g.
pressure by others or external circumstances, such as a
lack of funds) and sufficiently free of internal constraints
(e.g. pain/discomfort, the effects of medication, irrational
fears or values that are likely to change over time). The
Principle of Nonmaleficence (N) requires that the health
care professional not harm the patient, while the Principle
of Beneficence (B) states that the health care professional
should promote patient welfare. Finally, the Principle of
Justice (J) states that health care services and burdens
should be distributed in a just fashion. (Mappes and
DeGrazia 2001)
What makes ethical decision-making difficult with a
theory involving multiple prima facie duties is
determining which duty (duties) should prevail in a case
where the duties give conflicting advice. This requires
ethical sensitivity and expert judgment. We contend that
this sensitivity can be acquired systematically through
generalization of information learned about particular
cases where biomedical ethicists have a clear intuition
about the correct course of action. There will still,
undoubtedly, be borderline cases where experts, and so
also an ethical advisor system, will not be able to give a
definite answer; but even in these cases the advisor will be
able to elicit from the user the ethically relevant features
of the case, which can be quite helpful in and of itself.
John Rawls’ “reflective equilibrium” approach (Rawls
1951) to creating and refining ethical principles can be
used to help solve the problem of determining the correct
action when duties conflict. This approach involves
generalizing from intuitions about particular cases, testing
those generalizations on further cases, and then repeating
this process towards the end of developing a decision
procedure that agrees with intuition. This approach, that
would very quickly overwhelm a human being, lends itself
to machine implementation. For this reason, we believe
that machines can play an important role in advancing
ethical theory.
MedEthEx
MedEthEx (Medical Ethics Expert) is an implementation
of Beauchamp’s and Childress’ Principles of Biomedical
Ethics that, as suggested by Rawls’ reflective equilibrium
approach, hypothesizes an ethical principle concerning
relationships between these duties based upon intuitions
about particular cases and refines this hypothesis as
necessary to reflect our intuitions concerning other
particular cases. As this hypothesis is refined over many
cases, the principle it represents should become more
aligned with intuition and begin to serve as the decision
procedure lacking in Beauchamp’s and Childress’ theory.
MedEthEx is comprised of three components (Fig. 1): a
training module that abstracts the guiding principles from
particular cases supplied by a biomedical ethicist acting as
a trainer, a knowledge-based interface that provides
guidance in selecting duty intensities for a particular case,
and an advisor module that makes a determination of the
correct action for a particular case by consulting learned
knowledge. The first module is used to train the system
using cases in which biomedical ethicists have a clear
intuition about the correct course of action; the last two
Figure 1. MedEthEx Architecture.
1760
modules are used in concert to provide advice for an
ethical dilemma
The training module (used to refine the current
hypothesis) prompts the trainer for the name of an action
and an estimate of the intensity of each of the prima facie
duties satisfied or violated by this action (very violated,
somewhat violated, not involved, somewhat satisfied, very
satisfied). The trainer continues to enter this data for each
action under consideration. When data entry is complete,
the system seeks the intuitively correct action from the
trainer. This information is combined with the input case
to form a new training example which is stored and used
to refine the current hypothesis. After such training, the
new hypothesis will provide the correct action for this
case, should it arise in the future, as well as those for all
previous cases encountered. Further, since the hypothesis
learned is the least specific one required to satisfy these
cases, it may be general enough to satisfy previously
unseen cases as well.
The interface uses knowledge derived from ethicists
concerning the dimensions and duties of particular ethical
dilemmas. This knowledge is represented as finite state
automata (FSA) for each duty entailed. Questions
pertinent to the dilemma serve as start and intermediate
states, and intensities of duties as final states (as well as a
request for more information state). The input to the
interface is the user’s responses to the questions posed; its
output is a case with duty intensities corresponding to
these responses. This interface provides the experienced
guidance necessary to navigate the subtleties of
determining duty intensities in particular cases.
The advisor module consults the current version of the
hypothesis (as well as background knowledge) and , using
a resolution refutation system, determines if there is an
action that supersedes all others in the current case. If
such an action is discovered, it is output as the correct
action (in relation to the system’s training, a qualification
throughout this paper) in this dilemma. It further uses the
hypothesis, as well as stored cases, to provide an
explanation for its output.
As an example of how the ethical advisor MedEthEx
works, let us consider a common type of ethical dilemma
that a health care worker may face: A health care worker
has recommended a particular treatment for her competent
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 worker may not want to risk diminishing the patient’s
autonomy by challenging his decision; on the other hand,
the health care worker may have concerns about why the
patient is refusing the treatment. Three of the four
Principles/Duties of Biomedical Ethics are likely to be
satisfied or violated in dilemmas of this type: the duty of
Respect for Autonomy, the duty of Nonmaleficence and
the duty of Beneficence.
The system accepts a range of integers for each of the
duties from –2 to +2, where -2 represents a serious
violation of the duty, -1 a less serious violation, 0 indicates
that the duty is neither satisfied nor violated, +1 indicates
a minimal satisfaction of the duty and +2 a maximal
satisfaction of the duty.
MedEthEx uses inductive logic programming (ILP)
(Lavrac and Dzeroski 1997) as the basis for its learning
module. ILP is concerned with inductively learning
relations represented as first-order Horn clauses (i.e.
universally quantified conjunctions of positive literals Li
implying a positive literal H: H←(L1∧…∧Ln).
MedEthEx uses ILP to learn the relation
supersedes(A1,A2) which states that action A1 is preferred
over action A2 in an ethical dilemma involving these
choices (Anderson et al. 2005).
This particular machine learning technique was chosen
to learn this relation for a number of reasons. First, the
properties of the set of duties postulated by Beauchamp’s
and Childress are not clear. For instance, do they form a
partial order? Are they transitive? Is it the case that
subsets of duties have different properties than other
subsets? The potentially non-classical relationships that
might exist between duties are more likely to be
expressible in the rich representation language provided
by ILP. Further, a requirement of any ethical theory is
consistency. The consistency of a hypothesis regarding
the relationships between Beauchamp’s and Childress’
duties can be automatically confirmed across all cases
when represented as Horn clauses. Finally, commonsense
background knowledge regarding the superseding
relationship is more readily expressed and consulted in
ILP’s declarative representation language.
The object of training is to learn a new hypothesis that
is, in relation to all input cases, complete and consistent.
Defining a positive example as a case in which the first
action supersedes the remaining actions 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 negative cases. In
MedEthEx, 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.
MedEthEx starts with the most general hypothesis
(where A1 and A2 are variables): supersedes(A1,A2). This
states that all actions supersede each other and, thus,
covers all positive and negative cases. The system is then
provided with a positive case (and its negative) and
modifies its hypothesis such that it covers the given
positive case and does not cover the given negative case.
The following will help to illustrate this process. It
details MedEthEx training using a number of particular
cases within the type of dilemma we are considering, as
well as its use as an advisor in this dilemma.
Training Case 1. The patient refuses to take an antibiotic
that is almost certain to cure an infection that would
otherwise likely lead to his death. The decision is the
result of an irrational fear the patient has of taking
medications. (For instance, perhaps a relative happened
1761
to die shortly after taking medication and this patient now
believes that taking any medication will lead to death.)
The correct answer is that the health care worker should
try again to change the patient’s mind because if she
accepts his decision as final, the harm done to the patient
is likely to be severe (his death) and his decision can be
considered as being less than fully autonomous. This case
is represented using the values previously described as:1
Training Case 1 Autonomy Nonmaleficence Beneficence
√ Try Again -1 +2 +2
Accept +1 -2 -2
As the system’s starting hypothesis not only covers this
positive example (where try again serves as the correct
action over accept) but also the negative example
generated from it (where accept serves as the erroneously
correct action over try again), learning must be initiated.
No clauses are present in the starting hypothesis, so the
empty clause (which covers the only negative case) must
have all least specific specializations (LSS) generated
from it.
A specialization of clause C0 is a new clause C that
covers no more positive examples than C0 while covering
fewer negative cases. Such a specialization C is
considered least specific if there is no other specialization
of C0 that covers more positive examples (Bratko 1999).
MedEthEx specializes clauses by adding or modifying
conjuncts of the form favors (A,DA1,DA2,R) where A is a 1
or 2 signifying in which action’s favor the given duties lie,
Di is action i’s value (-2 to 2) for a particular duty D, and
R is a value (1 to 4) specifying how far apart the values of
these duties can be. favors is satisfied when the given
duty values are within the range specified. More formally:
favors(1,DA1 ,D A2 ,R) ← DA1 - D A2 >= R
favors(2,DA1 ,DA2,R) ← DA2 -DA1 >= 0 ∧ DA2 -D A1 =< R
The intuition motivating the use of favors as
MedEthEx’s specifying operation is that actions supersede
other actions based on the intensity differentials between
corresponding duties. The value of range R moderates the
specificity of the predicate. In the case where Action 1 is
favored in the pair of duties, a smaller R is less specific in
that it covers more cases. For instance,
favors(1,NA1,NA2,1) is satisfied when the difference
between Action 1’s and Action 2’s value for non-
maleficence is 1 through 4, whereas favors(1,NA1,NA2,2) is
only satisfied when the difference between Action 1’s and
Action 2’s value for non-maleficence is 2 through 4. In
the case where Action 2 is favored in the pair of duties, a
larger R is less specific in that it covers more cases. For
instance, favors(2,NA1,NA2,4) is satisfied when the
difference between Action 1’s value for non-maleficence is
1 through 4 where favors(2,NA1,NA2,3) is only satisfied
when the difference between Action 1’s value for non-
maleficence is 1 through 3. The intuition behind the
favors predicate is that, since Action 1 is the correct action
1 In analyzing this and the cases that follow, we are extrapolating from
material in Buchanan and Brock (1989).
in all training examples, if a duty differential favors it
then it follows that a larger differential will favor it as
well. Further, if a duty differential favors Action 2 (the
incorrect action in a training example of only two actions)
while still permitting Action 1 to be the chosen correct
action, it follows that a smaller differential will still
permit Action 1 to be chosen as well.
Refinement in MedEthEx favors duties whose
differentials are in favor of Action 1 as this is a more
likely relationship given that Action 1 is the correct action
in a training example and is clearly the only relationship
that, on its own, will support the claim that Action 1 is
favored. (Differentials that are in favor of Action 2 clearly
do not.) The range of these clauses is then incremented as
more specificity is required from them. When additions
and modifications of duty differentials in favor of Action 1
are not sufficient, clauses concerning duties whose
differentials are in favor of Action 2 are added and
decremented as necessary.
Given the current example case, the list of least specific
specializations is (favors(1,AA1,AA2,1), favors(1,NA1,NA2,1),
favors(1,BA1,BA2,1)) and it is found that two of these
clauses covers a case: (favors(1,NA1,NA2,1),
favors(1,BA1,BA2,1)). The first clause is removed from the
list and found to cover no negative examples, so further
refinement is not necessary and it becomes a clause in the
new rule. As all positive cases are covered, the process
stops and a new hypothesis, complete and consistent
through Training Case 1, has been generated:
supersedes(A1,A2) ← favors(1, NA1, NA2, 1)
That is, action A1 supersedes action A2 if the A1’s value
for the duty of nonmaleficence is at least 1greater than the
value for the duty of nonmaleficence for A2. To further
refine this hypothesis, another case is presented to the
training module.
Training Case 2. Once again, the patient refuses to take
an antibiotic that is almost certain to cure an infection that
would otherwise likely lead to his death, but this time the
decision is made on the grounds of long-standing religious
beliefs that don’t allow him to take medications.
The correct answer in this case is that the health care
worker should accept the patient’s decision as final
because, although the harm that will likely result is severe
(his death), his decision can be seen as being fully
autonomous. The health care worker must respect a fully
autonomous decision made by a competent adult patient,
even if she disagrees with it, since the decision concerns
his body and a patient has the right to decide what shall be
done to his or her body. This case is represented as:
Training Case 2 Autonomy Nonmaleficence Beneficence
Try Again -1 +2 +2
√ Accept +2 -2 -2
The current hypothesis does not cover Training Case 2
(i.e. is not complete) and covers the negative generated
from Training Case 2 (i.e. is not consistent) as well, so
1762
learning is initiated once again. To reinstate the current
rule’s consistency, a list of least specific specializations
(LSS) is generated from the only clause of the current
hypothesis, favors(1,NA1,NA2,1). These include the next
range increment (2) for this clause, as well as conjuncts of
this clause with other duties favoring both action 1 and
action 2:
favors(1, NA1, NA2, 2),
favors(1, NA1, NA2, 1) ∧ favors(1, AA1, AA2, 1),
favors(1, NA1, NA2, 1) ∧ favors(1, BA1, BA2, 1),
favors(1, NA1, NA2, 1) ∧ favors(2, AA1, AA2, 4),
favors(1, NA1, NA2, 1) ∧ favors(2, BA1, BA2, 4)
Note that, since the current clause does not cover Case
2, no amount of specialization will ever cause it to do so,
so we are only interested in specializations that continue
to cover Case 1. The only clauses from the list of LSS
found to do so are:
favors(1,NA1, NA2, 2),
favors(1,NA1, NA2, 1) ∧ favors(1,BA1, BA2, 1),
favors(1,NA1, NA2, 1) ∧ favors(2,AA1, AA2, 4)
As the search for a clause that does not cover the
negative case generated from Training Case 2 (i.e. is
consistent) continues, it is found that no single clause
favoring action 1 in nonmaleficence in any range will be
consistent, so this branch terminates. The same is true of
any clause that is a conjunct of nonmaleficence and
beneficence in favor of action 1, terminating this branch.
It is found, however, that a clause consisting of a conjunct
favoring action 1 in nonmaleficence with a range of 1 or
more and a conjunct favoring action 2 in autonomy with a
range of 2 or less does not cover the negative generated
from Training Case 2 while still covering Case 1. Case 1
is removed from consideration and this conjunct becomes
the first disjunct of the new hypothesis:
favors(1, NA1, NA2, 1) ∧ favors(2, AA1, AA2, 2)
As this hypothesis still needs to cover Training Case 2,
the process continues with the search for a new clause that
does so without covering the negative cases generated
from Training Cases 1 and 2. This search starts with an
empty clause which, being the most general, covers all
positive and negative examples. All LSS are generated
from it, garnering the same clauses generated originally
for Training Case 1. It is found that only one of these
clauses covers Training Case 2 (the only uncovered case
left):
favors(1, AA1, AA2, 1)
Since this clause covers the negative case generated
from Training Case 1 (i.e. is not consistent), all LSS are
generated from it which includes the next increment (2)
favoring action 1 in autonomy (among other clauses). It is
found, through further search, that the next increment (3)
of this clause covers Training Case 2 without covering any
negative cases so it becomes the second clause of the new
hypothesis and Training Case 2 is removed from further
consideration. As there are no uncovered cases, the new
hypothesis, complete and consistent through Training
Case 2, is then generated:
supersedes(A1,A2) ←
(favors(1, NA1, NA2, 1) ∧ favors(2, AA1, AA2, 2)) ∨
favors(1, AA1, AA2, 3)
This rule states that if action 1 favors nonmaleficence with
a value at least 1 greater than action 2 and action 2 favors
autonomy with a value no greater than 2 over action 1 or
action 1 favors autonomy 3 or greater over action 2, then it
is the preferred action. This rule begins to tease out the
subtle relationship between nonmaleficence and autonomy
in Beauchamp’s and Childress’ theory in a way that
proves useful in other circumstances. With just these two
cases, the ethical advisor has learned a rule that would
give correct advice in a third, entirely new case of within
the same type of dilemma. To provide an example use of
the trained system, the duty intensities of this test case will
be generated via the knowledge-based interface.
Test Case. The patient refuses to take an antibiotic that is
likely to prevent complications from his illness,
complications that are not likely to be severe, because of
long-standing religious beliefs that don’t allow him to take
medications.
When the system is consulted, it first seeks information
to determine the satisfaction/violation level of the duty of
autonomy for each action. To do so, it presents questions
as required. The system first asks whether or not the
patient understands the consequences of his decision. If
the health care worker is not sure, she may need to seek
more information from the patient or, depending upon her
answers to later questions, the system may determine that
this is not a fully autonomous decision. If we assume that
the health care worker believes that the patient does
indeed know the consequences of his action, the system
then asks questions to determine if the patient is externally
constrained. The healthcare worker answers “no” because
the reason why the patient is refusing to take the antibiotic
has nothing to do with outside forces. Finally, it asks
questions to determine if the patient is internally
constrained. Since the patient is not constrained by
pain/discomfort, the effects of medication, irrational fears
or values that are likely to change over time, the answer is
“no.” This is because the belief that has led to his refusing
the antibiotic is a long-standing belief of his. The answers
provided to these questions have the system conclude that
the patient’s decision is fully autonomous, giving the
value +2 to the duty of autonomy for accepting the
patient’s decision. The value for challenging the patient’s
decision is -1 because questioning the patient’s decision,
which challenges his autonomy, is not as strong as acting
against the patient’s wishes which would have been a -2.
The system then seeks information to determine the
satisfaction/violation level of the duty of nonmaleficence
for each action. To do so, it presents questions concerning
the possibility and severity of harm that may come to the
patient given his decision. As harm will likely result from
the patient’s decision, but it will not be severe, the system
gives the value of -1 to the duty of nonmaleficence for
accepting the patient’s decision. Challenging the patient’s
1763
decision could avoid this moderate harm, so a +1 to the
duty of nonmaleficence is assigned to this action.
The system then seeks information to determine the
satisfaction/violation level of the duty of beneficence for
each action. To do so, it presents questions concerning
the possibility and level of improvement of quality of the
patient’s life that may result from accepting/challenging
his decision. As the quality of the patient’s life would
worsen somewhat if the patient’s decision were accepted
and improve somewhat if not, the system gives the value
of -1 to the duty of beneficence for accepting the patient’s
decision and a +1 for challenging it. The test case, then, is
generated as:
Test Case Autonomy Nonmaleficence Beneficence
Try Again -1 +1 +1
Accept +2 -1 -1
The system then consults the current hypothesis for
both supersedes(try again, accept) and supersedes(accept,
try again). It finds that the first is not covered by the
current hypothesis but the second is covered by the clause
favors(1,AA1,AA2,3), that is, autonomy is favored by at least
3 in action 1 (the correct action). As action 1 in this case
is accept, the system advises the user to accept the
patient’s decision. The correct answer is indeed that the
health care worker should accept his decision, since once
again the decision appears to be a fully autonomous one
and there is even less possible harm at stake than in
Training Case 2.
Three additional training cases are sufficient to learn a
rule that correctly covers all eighteen possible cases
(combinations of 2 sets of satisfaction/violation values
possible for the duty of respect for autonomy, 3 for the
duty of nonmaleficence, and 3 for the duty of beneficence)
of the type of dilemma under consideration.
Training Cases 3-5.
The cases are represented as:
Training Case 3 Autonomy Nonmaleficence Beneficence
Try Again -1 0 +1
√ Accept +1 0 -1
Training Case 4 Autonomy Nonmaleficence Beneficence
√ Try Again -1 +1 +1
Accept +1 -1 -1
Training Case 5 Autonomy Nonmaleficence Beneficence
√ Try Again -1 0 +2
Accept +1 0 -2
The final rule that results from these training cases is:
supersedes(A1,A2) ←
(favors(1, NA1, NA2, 1) ∧ favors(2, AA1, AA2, 2)) ∨
favors(1, AA1, AA2, 3) ∨
(favors(1, AA1, AA2, 1) ∧
favors(2, BA1, BA2, 3) ∧ favors(2, NA1, NA2, 1)) ∨
(favors(1, BA1, BA2, 3) ∧ favors(2, AA1, AA2, 2))
This rule states, in relation to the type of dilemma under
consideration, that a health care worker should challenge
a patient’s decision if it is not fully autonomous and either
there is any violation of the duty of nonmaleficence or
there is a severe violation of the duty of beneficence.
This philosophically interesting result lends credence to
Rawls’ Method of Reflective Equilibrium. We have,
through abstracting a principle from intuitions about
particular cases and then testing that principle on further
cases, come up with a plausible principle that tells us
which action is correct when specific duties pull in
different directions in a particular ethical dilemma.
Furthermore, the principle that has been so abstracted
supports an insight of Ross’ that violations of the duty of
nonmaleficence should carry more weight than violations
of the duty of beneficence.
We have described a proof-of-concept system that is
constrained to a single type of ethical dilemma in which
only three of Beauchamp’s and Childress’ four Principles
of Biomedical Ethics are involved. Future extensions of
this system include widening its scope to include other
ethical dilemmas, some involving the duty of justice as
well, and further enhancement of the user interface to
incorporate more detailed knowledge elicitation as well as
explanatory information. Decision principles gleaned and
past cases pertinent to a new case can be used both to
guide the user in the process of abstracting ethically
relevant information from a case and, further, to provide
support for conclusions reached by the system.
Beyond MedEthEx
As an example of how machine ethics can be used to
improve the performance of a system, consider an
artificially intelligent care provider or eldercare system.
One duty of an eldercare system is to provide reminders
for taking medications, eating meals, etc., which ensure
that the duties of beneficence and nonmaleficence will be
satisfied. As another important goal of an eldercare system
is the maintenance of a patient's autonomy, an ethical
tension arises when these conflict: constant reminding
and/or reporting to overseers can erode patient autonomy.
The decision principles developed by MedEthEx may
prove useful to such a system as a theoretically valid
foundation for comparing the ethical weight of the
system’s candidate actions, determining a partial order of
these actions along an ethical dimension.
Given candidate actions "don't remind", "remind", and
"report", each action’s satisfaction/violation values for
relevant ethical duties (respect for autonomy, beneficence,
and nonmaleficence) could be determined by tracking
pertinent variables over time, such as the risk of harm of a
refusing a particular medication. When the system is
presented with a set of candidate actions (along with the
satisfaction/violation values for each action’s relevant
duties), the supersedes predicate developed by MedEthEx
can be used to order these actions along an ethical
dimension. This information can then be combined with
1764
extra-ethical information to decide the system's next
action. Given the number of things for which reminders
may need to be given, this framework may provide a
verifiable abstraction better able to deal with the ensuing
complexity than an ad hoc approach. An eldercare
system, guided by the developed ethical principles, will be
better equipped to handle conflict in its duties with greater
sensitivity to the needs of the human with which it
interacts.
Related Work
Although there have been a few who have called for it,
there has been little to no serious scientific research
conducted in machine ethics. A few interesting exceptions
were presented in 1991 at the Second International
Workshop on Human & Machine Cognition: Android
Epistemology (Ford et al 1991). Unfortunately, none of
the work of this workshop seems to have been pursued any
further.
A more extended effort in computational ethics can be
found in SIROCCO (McLaren 2003), a system that
leverages information concerning a new problem to
predict which previously stored principles and cases are
relevant to it in the domain of professional engineering
ethics. This system is based upon case-based reasoning
techniques. Cases are exhaustively formalized and this
formalism is used to index similar cases in a database of
previously solved cases that include principles used in
their solution. Deductive techniques, as well as any
attempt at decision-making, are eschewed by McLaren due
to “the ill-defined nature of problem solving in ethics.”
We contend that an “ill-defined nature” does not make
problem solving in ethics completely indefinable and are
embarking on attempts of just such definition in
constrained domains. Furthermore, we maintain that
decisions offered by a system that are consistent with
decisions made previously by ethicists in clear cases have
merit and will be useful to those seeking ethical advice
(Anderson et al. 2004, 2005).
Conclusion
Our research advances from speculation to
implementation by building systems grounded in ethical
theory and, further, advances this theory through analysis
of these implemented systems. It is a domain-specific
extension of work of Anderson, Anderson, and Armen
(2005) where the use of cases and inductive logic
programming rule learning (based upon Ross’ Theory of
Prima Facie Duties) is first postulated.
We have developed MedEthEx, to our knowledge the
first system that helps determine the best course of action
in a biomedical ethical dilemma. This approach can be
used in the implementation of other such systems that may
be based upon different sets of ethical duties and
applicable to different domains. Further, the formally
represented ethical principles developed in this research,
as well as the formal methods adapted for their
consultation, will be useful in creating machines that can
interact with those in need of health care in a way that is
sensitive to ethical issues that may arise.
Acknowledgement
This material is based upon work supported in part by the
National Science Foundation grant number IIS-0500133.
References
Anderson, M., Anderson, S. and Armen, C. 2005. Toward
Machine Ethics: Implementing Two Action-Based Ethical
Theories. Proceedings of the AAAI 2005 Fall Symposium
on Machine Ethics, Crystal City, VA.
Anderson, M., Anderson, S. and Armen, C. 2004. Toward
Machine Ethics. Proceedings of AAAI 2004 Workshop on
Agent Organizations: Theory and Practice, San Jose, CA.
Anderson, S., 1999. “We Are Our Values”, in Questioning
Matters, Kolak, D. (ed.), Mayfield Publishing Company,
p. 599.
Buchanan, A.E. and Brock, D.W. 1989. Deciding for
Others: The Ethics of Surrogate Decision Making, pp48-
57, Cambridge University Press.
Bratko, I. 1999. Refining Complete Hypotheses in ILP.
Inductive Logic Programming, LNAI 1634, Springer.
Beauchamp, T.L. and Childress, J.F. 1979. Principles of
Biomedical Ethics, Oxford University Press.
Ford, K., Glymour, C. and Hayes, P.J. 1991. Android
Epistemology. MIT Press.
Lavrac, N. and Dzeroski, S. 1997. Inductive Logic
Programming: Techniques and Applications. Ellis
Harwood.
Mappes, T.A and DeGrazia, D. 2001. Biomedical Ethics,
5th Edtion, pp. 39-42, McGraw-Hill, New York.
McLaren, B. M. 2003. Extensionally Defining Principles
and Cases in Ethics: an AI Model, Artificial Intelligence,
Volume 150, November, pp. 145-181.
Rawls, J. 1951. Outline for a Decision Procedure for
Ethics. Philosophical Review, 60.
Ross, W.D. 1930. The Right and the Good, Clarendon
Press, Oxford.
1765
