This Content Downloaded From 13.234.156.0 On Sun, 02 Aug 2020 10:39:47 UTC

Artificial Intelligence and Law: Stepping Stones to a Model of Legal Reasoning
Author(s): Edwina L. Rissland

Source: The Yale Law Journal , Jun., 1990, Vol. 99, No. 8 (Jun., 1990), pp. 1957-1981
Published by: The Yale Law Journal Company, Inc.
Stable URL: http://www.jstor.com/stable/796679
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms
is collaborating with JSTOR to digitize, preserve and extend access to The Yale Law Journal
This content downloaded from

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
All use subject to https://about.jstor.org/terms
Artificial Intelligence and Law: Stepping
Stones to a Model of Legal Reasoning
Edwina L. Risslandt
This Comment discusses developments in the twenty-year-old interdis-

ciplinary field of Artificial Intelligence (AI) and law. This field is impor-
tant for both AI and law because it is directed at improving our under-
standing and modeling of legal reasoning.
The AI and law projects discussed here are landmarks in this field., A
unifying theme of the projects is the goal to understand and model legal
argument, a keystone of an overarching goal to understand and model
legal reasoning. These goals require that we know first how to represent
several types of knowledge, such as cases, rules, and arguments; second,
how to reason with them, such as to manipulate precedents, to apply and
make inferences with rules, and to tailor arguments to facts; and third,
how to use them ultimately in a computer program that can perform tasks
in legal reasoning and argumentation, such as analogizing favorable cases
and distinguishing contrary ones, anticipating parries in adversarial argu-
ment, and creating artful hypothetical.
The projects constitute a coherent set of studies about key topics in AI
and law: (1) reasoning with rules; (2) handling open-textured legal con-
cepts; (3) reasoning with cases and hypotheticals; (4) integrating reasoning
with rules and reasoning with cases; and (5) representing legal knowl-
edge. The limitations of techniques for handling the first topic are ad-
dressed by the second, and the third topic addresses critical issues not cov-
ered at all by work on the first. The fourth integrates work on the first
three, and the fifth provides underpinnings needed for all of them. The
projects are some of the major accomplishments in AI and law, especially
t Associate Professor, Department of Computer and Information Science, University of Massachu-

setts at Amherst and Lecturer on Law, Harvard Law School. I am grateful to my colleagues and
students, in particular, Kevin Ashley, David Blank, Martha Minow, Oliver Selfridge, and David
Skalak for numerous discussions in the preparation of this manuscript, and to the National Science
Foundation, the Defense Advanced Research Projects Agency of the Department of Defense, and
GTE Laboratories for their support.
1. For one of the earliest discussions of AI and law, see Buchanan & Headrick, Some Speculation
About Artificial Intelligence and Legal Reasoning, 23 STAN. L. REV. 40 (1970). See also Gardner,
Law Applications, in THE ENCYCLOPEDIA OF ARTIFICIAL INTELLIGENCE 456 (S. Shapiro ed. 1989)
(providing general overview of AI and law). For a representative collection of current work, see THE
SECOND INTERNATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE AND LAW: PROCEEDINGS OF
THE CONFERENCE (1989) [hereinafter ICAIL-89].
1957

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
1958 The Yale Law Journal [Vol. 99: 1957
those seeking to understand and model legal argument, and they exem-
plify and illustrate the progress, concerns, and style of its research.
I. BACKGROUND ON ARTIFICIAL INTELLIGENCE
A. What is Artificial Intelligence?
AI is the study of cognitive processes using the conceptual frameworks

and tools of computer science.2 As a distinct subfield of computer science,
AI had its beginnings in the mid-fifties.' In 1968 Marvin Minsky, one of
the founders of AI, said it well: Al is "the science of making machines do
things that would require intelligence if done by man."4 Thus, all manner
of intelligent behavior is in the realm of AI, including playing chess,' solv-
ing calculus problems,6 making mathematical discoveries,7 understanding
short stories,8 learning new concepts,9 interpreting visual scenes,10 diag-
nosing diseases,"1 and reasoning by analogy."2 Any discussion of AI must
2. For additional background, see E. CHARNIAK & D. MCDERMOTT, INTRODUCTION To ARTI-

FICIAL INTELLIGENCE (1985); 1-4 HANDBOOK OF ARTIFICIAL INTELLIGENCE (A. Barr, P. Cohen &
E. Feigenbaum eds 1981-89); G. LUGER & W. STUBBLEFIELD, ARTIFICIAL INTELLIGENCE AND THE
DESIGN OF EXPERT SYSTEMS (1989); THE ENCYCLOPEDIA OF ARTIFICIAL INTELLIGENCE, supra
note 1; P. WINSTON, ARTIFICIAL INTELLIGENCE (2d ed. 1984).
3. A summer conference held at Dartmouth College in 1956 is perhaps the most convenient
landmark to denote its beginning. See P. MCCORDUCK, MACHINES WHO THINK 93-114 (1979)
(describing history surrounding the Dartmouth Conference and its lasting effects in Al community).
4. SEMANTIC INFORMATION PROCESSING v (M. Minsky ed. 1968).
5. See, e.g., Newborn & Kopec, Results of The Nineteenth ACM North American Computer
Chess Championship, 32 COMM. ACM 1225 (1989) (describing several games).
6. Research on symbolic mathematics began with James Slagle's SAINT, a program which solved
52 out of 54 problems selected from MIT freshman calculus final examinations. See Slagle, A Heuris-
tic Program That Solaves Symbolic Integration Problems in Freshman Calculus, in COMPUTERS AND
THOUGHT 191 (E. Feigenbaum & J. Feldman eds. 1963). The next project was Joel Moses's pro-
gram SIN. See Moses, Symbolic Integration: The Stormy Decade, 14 COMM. ACM 548 (1971). Fur-
ther research on symbolic mathematics has led to the development of systems such as MIT's MAC-
SYMA, "a large, interactive computer system designed to assist mathematicians, scientists, and
engineers in solving mathematical problems." 2 HANDBOOK OF ARTIFICIAL INTELLIGENCE, supra
note 2, at 143. Commercial versions of such programs are now available for personal computers.
Wayner, Symbolic Math on the Mac, BYTE, Jan. 1989, at 239.
7. A program that modeled the process of discovery in mathematics was Doug Lenat's AM. Using
a knowledge base containing fundamental concepts and rules of thumb, AM was able to "discover"
such mathematical concepts as "divisor" and "prime." R. DAVIS & D. LENAT, KNOWLEDGE-BASED
SYSTEMS IN ARTIFICIAL INTELLIGENCE 1-225 (1982); AM, in 3 HANDBOOK OF ARTIFICIAL INTEL-
LIGENCE, supra note 2, at 438.
8. See, e.g., R. SCHANK & R. ABELSON, SCRIPTS, PLANS, GOALS, AND UNDERSTANDING: AN
INQUIRY INTO HUMAN KNOWLEDGE STRUCTURES (1977) (discussing theory for story understand-
ing); INSIDE COMPUTER UNDERSTANDING: FIVE PROGRAMS PLUS MINIATURES (R. Schank & C.
Riesbeck eds. 1981). For general collections of work on natural language processing, see READINGS IN
NATURAL LANGUAGE PROCESSING (B. Grosz, K. Jones & B. Webber eds. 1986); STRATEGIES FOR
NATURAL LANGUAGE PROCESSING (W. Lehnert & M. Ringle eds. 1982).
9. See, e.g., 1-2 MACHINE LEARNING: AN ARTIFICIAL INTELLIGENCE APPROACH (R. Michalski,
J. Carbonell & T. Mitchell eds. 1983, 1986).
10. See, e.g., READINGS IN COMPUTER VISION: ISSUES; PROBLEMS, PRINCIPLES, AND PARA-
DIGMS (M. Fischler & 0. Firschein eds. 1987).
11. See, e.g., RULE-BASED EXPERT SYSTEMS, THE MYCIN EXPERIMENTS OF THE STANFORD
HEURISTIC PROGRAMMING PROJECT (B. Buchanan & E. Shortliffe eds. 1984) [hereinafter Buchanan
& Shortliffe] (thorough discussion of MYCIN diagnostic system for bacterial infections).
12. See, e.g., Gentner, Structure-Mapping: A Theoretical Framework for Analogy, 7 COGNITIVE

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
1990] Artificial Intelligence and Law 1959
note that tasks involving "common sense" reasoning or perception, such as

language understanding, are by far the most difficult for AI.'3 More tech-
nical tasks, like solving calculus problems or playing chess, are usually
much easier. That is because the latter can be framed in well-defined
terms"4 and come from totally black-and-white domains, while the former
cannot and do not. What distinguishes the AI approach from other studies
of cognition and knowledge, such as psychology or philosophy, is its insis-
tence on grounding the analysis in computational terms-preferably in a
successfully running computer program that embodies the analysis."5
AI is pursued for at least two reasons: to understand the workings of
human intelligence and to create useful computer programs and com-
puters that can perform intelligently. Most workers in the field of AI pur-
sue these goals simultaneously. For instance, in the course of designing a
computer program for commercial purposes like making credit card ap-
proval decisions, the program designer needs to examine how experienced
people make such decisions, since they are usually the best, and often the
only, source of information about how the job is done.17 Likewise, in en-
gaging in Al for the sake of understanding or modeling cognition, it is far
Sci. 155 (1983). Reasoning by analogy was the subject of one of the earliest programs in AI. Evans, A
Program for the Solution of a Class of Geometric-Analogy Intelligence-Test Questions, in SEMANTIC
INFORMATION PROCESSING supra note 4, at 271.
13. See, e.g., E. CHARNIAK, TOWARD A MODEL OF CHILDREN'S STORY COMPREHENSION
271-74 (MIT Al Lab Technical Report No. 266, 1972) (even understanding "simple" children's
stories can be quite messy).
14. For instance, the rules of chess completely define the game. Furthermore, in the case of
games, solutions can be found using high-powered, specialized "search" techniques. This is how
highly successful chess programs, such as Deep Thought, work. Deep Thought can examine 720,000
chess positions per second. See Newborn, & Kopec, supra note 5, at 1225. However, it is quite a
different matter whether such programs work in the same way human chess experts do and whether
they can shed light upon human thought processes. See Chase & Simon, The Mind's Eye in Chess, in
VISUAL INFORMATION PROCESSING 215, 278 (W. Chase ed. 1973).
15. See D. DENNETT, BRAINSTORMS: PHILOSOPHICAL ESSAYS ON MIND AND PSYCHOLOGY 112
(1978); MIND DESIGN (J. Haugeland ed. 1981).
16. For two interesting theories of human cognition, see M. MINSKY, THE SOCIETY OF MIND
(1986); A. NEWELL, UNIFIED THEORY OF COGNITION (forthcoming 1990).
17. Even if psychological validity is not usually paramount, it is often helpful. The MYCIN
Project illustrates this point. The goal of the MYCIN Project was to build a system that could diag-
nose bacterial blood infections at an expert level. Although the goal was not to model closely the
diagnostic behavior of expert physicians, observations of medical experts were critical during the early
phases of the project stage, when the Al researchers (known as "knowledge engineers") gathered,
structured, and encoded the experts' medical knowledge for use by the program. Later, having the
program operate in a comprehensible manner was critical for debugging and refining it. See generally
Buchanan & Shortliffe, supra note 11. It is usually the case that if there is no point of contact
between the program's processing style and the human's, the program behavior appears inscrutable,
impeding its development. Some similarity between the program's and the experts' processing also
enhances one's belief in the correctness of the output of the program; sometimes this is so because it is
easier for the program to explain its own reasoning in the user's terms. With respect to the issue of
capturing the style of expert reasoning, a chess playing program like Deep Thought is an extreme
case of a high performance program where there is no claim to cognitive validity. See Newborn &
Kopec, supra note 5. There was no attempt to make Deep Thought think like a grand master. See
Leithauser, Kasparov Beats Deep Thought, N.Y. Times, Jan. 14, 1990, ? 6 (Magazine), at 33, 74
(discussion of grand master Kasparov's thoughts on ramifications of some of these issues).

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
more satisfying to exhibit a running program: to some degree success is a

working computational model.18
In the context of law, these twin rationales translate into the twin goals
of understanding certain key aspects of legal reasoning and building com-
putational tools useful for legal practice, teaching, or research. An exam-
ple of the former is the development of an AI model for reasoning based
upon the doctrine of precedent. The process of developing an AI model
causes one to learn about legal reasoning. Modeling involves elucidating
key ingredients of precedent-based reasoning, such as making assessments
of the relevance of precedents to new situations, distinguishing contrary
cases, and drawing connections between relevant cases; then describing
them in detail and building a program to execute them.
An example of the second, more applications-oriented goal, is construc-
tion of a set of computational tools (a lawyer's workbench, so to speak) to
assist in the preparation of a brief. These may include functions to assist
in gathering relevant precedents from data bases, sorting them according
to their doctrinal approaches, and "Shepardizing" them. Building a prac-
tical system, like one to assist with writing a brief, requires developing
analytical models. Typically, satisfaction with an analytical model in-
creases if it offers insights leading to the practical advances.
The desire to develop a model of legal reasoning is not new. Certainly
key aspects of legal reasoning, such as the analysis of precedent, have been
the subject of many discussions.19 However, for the most part, previous
studies have not provided the level of detail required of an AI model; that
is, they have not provided enough detail to indicate how they could be
implemented as a computer program.20 In AI, one is forced to be detailed.
For instance, law has been described as "reasoning by example."2 This
may be an appropriate level of description for some purposes, but for AI,
it leaves too many questions unanswered. To take advantage of invaluable
18. Actually building a program is quite different from speculating about it. Programming makes
abundantly clear the weaknesses or difficulties of the model. Cf. McDermott, Artificial Intelligence
Meets Natural Stupidity, in MIND DESIGN, supra note 15, at 143, 156-59 (highlighting risk of
theorizing about implementation without actually implementing).
19. See, e.g., B. CARDOZO, THE NATURE OF THE JUDICIAL PROCESS (1921); E. LEVI, AN IN-
TRODUCTION TO LEGAL REASONING (1949); K. LLEWELLYN, THE BRAMBLE BUSH (1930); K.
LLEWELLYN, THE CASE LAW SYSTEM IN AMERICA (1989); Radin, Case Law and Stare Decisis:
Concerning Praijudizienrecht in Amerika, 33 COLUM. L. REV. 199 (1933).
20. Although this was obviously not their purpose, some discussions have come remarkably close.
For instance, some of the analyses done by Llewellyn and Radin capture the spirit of the sort of
description desired in Al. See, e.g., K. LLEWELLYN, supra note 19, at 50 n.I (describing doctrine of
precedent, particularly concerning broad and narrow reading of rule of case); Radin, supra note 19,
at 206-09 (describing concept evolution in law). Radin's description is uncannily similar to an al-
gorithm, called the "candidate elimination algorithm," used in machine learning. See 3 HANDBOOK
OF ARTIFICIAL INTELLIGENCE, supra note 2, at 385-91. For a comparison of this algorithm and
Radin's analysis, see Rissland & Collins, The Law as Learning System, in PROCEEDINGS OF THE
FOURTH ANNUAL CONFERENCE OF THE COGNITIVE SCIENCE SOCIETY 500, 501 (1986) (examining
evolution of concept of "inherently dangerous" from Radin).
21. E. LEVI, supra note 19, at 1.

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
insights offered by legal scholars about legal reasoning, an AI researcher

needs to specify both how the reasoning is to happen and what informa-
tion and methods are required.
The Al approach forces one to be relentlessly analytic and specific. It
advocates that one use ideas and methods of computer science to develop
conceptual and computational frameworks.
B. AI and Law: A Fruitful Synergy
The law offers abundant opportunities for developing analytic and com-
putational AI models. Law also has unique characteristics that make it a
particularly challenging field for AI:
1. Legal reasoning is multi-modal, rich and varied: it includes rea-

soning with cases, rules, statutes and principles;
2. Case law has an explicit style and standard of reasoning and justi-
fication: stare decisis.
3. Specialized legal knowledge, such as cases and statutory rules, is
well-documented and available from many sources, including case
reporters, treatises, restatements, statutes, commercial summaries,
and scholarly commentaries.
4. The law is self-aware and self-critical, and has an established tra-
dition of examining its processes and assumptions. There is lively
debate between proponents of competing jurisprudential schools.
5. The character of answers in the law is different from those in
many other disciplines: answers are much more a matter of degree
than clear-cut yes-or-no and they can change over time.
6. The knowledge used in legal reasoning is diverse, ranging from
common sense to specialized legal knowledge, and it varies greatly in
structure, character, and use.
These observations suggest the possibility of fruitful synergy between

law and Al, and have implications for Al approaches. That the law is
multi-modal means that an Al program will need to know about several
modes of reasoning and how to use them in concert. That the law has an
explicit, accepted style of reasoning gives the researcher a definite reason-
ing style to incorporate into a model and to use to analyze legal reasoning.
However, some of the modes of reasoning, such as reasoning with cases,
are very different from those used most widely in Al, and Al researchers
have become interested in these only recently.
That the law is well-documented, self-aware, and self-critical makes it
more accessible to Al researchers. If there were no repositories of legal
knowledge and no tradition of trying to describe and criticize the goals
and methods of legal reasoning, an Al researcher would have to start from
scratch in trying to understand legal reasoning and to elucidate the knowl-
edge involved. By no means is this to say that all the spadework has been
done-the philosophical descriptions of legal reasoning are vague for Al

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
purposes, and there is a lot more to legal knowledge than the "book
knowledge" of traditional legal materials. Still, an Al study of the law has
been provided with a good beginning, an epistemological leg up.
That what counts as an "answer" in the law is not clearcut is also
different from other disciplines. In law there is usually no unique right
answer; rather there are reasonable alternative answers, more a matter of
degree than of extremes. The answers are highly contextual, depend on
goals and points of view, and change as the law evolves. Even the rule-
based aspects of legal reasoning cannot be modeled with purely deductive
methods. This also means that, unfortunately, there is never the comfort
of a quod erat demonstrandum at the end of a reasoning episode to sanc-
tion it as sound and beyond reproach, as there is in mathematics. From
the legal point of view, this is no real deficit-it's a feature and not a bug,
to use the computer scientists' phrase-since it allows law to accommodate
to changes in circumstance and purpose. But, computationally, the nature
of legal answers adds complexity and difficulty as well as richness and
flexibility.
These observations all suggest that the law is an exceedingly challeng-
ing domain for Al. Research in Al and law will impel Al in new direc-
tions and thus benefit Al. In turn, law will also benefit from Al, both
analytically and practically. As an analytical medium, Al forces meticu-
lous attention to details and precise testing of theoretical ideas. This, in
turn, facilitates the unmasking of flaws, furthers the understanding of as-
sumptions, and leads to proposals for refinements. Al focuses a spotlight
on issues of knowledge and process to a degree not found in non-
computational approaches, which often assume that some key task, like
analogical reasoning, will be accomplished magically without a hint as to
how,22 or with too many critical details left.underspecified.23 The practi-
22. The philosopher Dan Dennett calls this the problem of a "loan of intelligence" or the hidden
homunculus. See D. DENNETT, supra note 15, at 12. In an AI model, some process, somewhere, must
actually do the work or else as Dennett puts it, the theory is "in the red." Id. A great danger is in
"designing a system whose stipulated capacities are miraculous." Id. at 112; see also McDermott,
supra note 18, at 143.
23. Some of Dworkin's models are intriguing in this regard, such as his model of hard cases. See,
e.g., R. DWORKIN, TAKING RIGHTS SERIOUSLY (1977). He argues that there may not really be any
"hard" cases since one can use a set of relatively weighted principles to resolve certain "hard" ques-
tions, which arise because the on point cases are in conflict or there is a tie in the principles. The
principles and weights are generated from a collection of relevant precedents. Dworkin omits the
details of how to decide which precedents to use and how to induce principles from them. He is not
necessarily wrong, but it would be instructive to extract a more detailed description of how his model
works. By declining to instruct further on how to develop the weighting system, Dworkin has simply
moved the problem of analysis back one step. Regarding the assignment of relative weights, he has
walked headlong into the "credit assignment" quagmire, well known to workers in machine learning,
where the problem is to assign credit or blame for the overall success or failure of a problem solution
to an individual step or aspect of it. For instance, is the credit for a win or the blame for a loss in a
game of checkers to be given to the penultimate move of a game, the first move, or some intermediate
move or moves? How can one tell which features or principles "caused" a case to be resolved for one
party or the other? See Minsky, Steps Toward Artificial Intelligence, in COMPUTERS AND THOUGHT,
supra note 6, at 432 (1963); Samuel, Some Studies in Machine Learning Using the Game of Check-

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
cal benefits to law from Al are intelligent computational tools. The rela-
tionship between Al and law is truly synergistic.
C. Some Desiderata for AI and Law Programs
Given the special characteristics of law described above, we can enu-

merate several goals that we would like an ideal Al and law program to
achieve. It should be able, among other things, to:
1. Reason with cases (both real and hypothetical) and analogies;

2. Reason with rules;
3. Combine several modes of reasoning;
4. Handle ill-defined and open-textured concepts;
5. Formulate arguments and explanations;
6. Handle exceptions to and conflicts among items of knowledge,
like rules;
7. Accommodate changes in the base of legal knowledge, particularly
legal concepts, and handle non-monotonicity, that is, changes in
which previous truths no longer hold as more becomes known;24
8. Model common sense knowledge;
9. Model knowledge of intent and belief;
10. Perform some aspects of natural language understanding.
We are a long way from such an ideal. There is, however, activity on all
of these important fronts, and impressive progress on a few.
In fact, one can today speculate about which of these desiderata can be
expected now, in the near future, someday, or probably never. Presently,
Al is actively pushing back the boundaries on case-based reasoning (goal
1), has well-understood methodologies in rule-based reasoning (goal 2),
and is exploring multi-paradigm reasoning (goal 3). Reasoning with
open-textured predicates (goal 4) has had an admirable first cut, but it
will require further contributions from other Al specialities like case-
based reasoning (CBR)25 and machine learning.26 Major contributions
ers, in 3 IBM J. RES. & DEV. 211 (1959).

24. For instance, the overturning of established doctrine with a new precedent.
25. "Case-based reasoning" is the process by which a program uses previous cases (e.g., prece-
dents, experiences, problem-solving episodes, detailed solutions, plans) to solve a new case. There are
two kinds of CBR: (1) "problem-solving CBR," in which the emphasis is on generating a solution to
a new problem (case) by reusing and modifying old solutions; and (2) "interpretive CBR," in which
the emphasis is on generating an interpretation of a new case by comparing and analogizing it with
interpretations made in past cases. For both varieties, the key issues are the structure and content of
case memory, indexing mechanisms for retrieving cases from memory, and metrics for assessing simi-
larity or relevancy of retrieved cases. The sort of precedent-based reasoning done in the law is a
paradigm for the second type of CBR. See generally PROCEEDINGS: WORKSHOP ON CASE-BASED
REASONING (J. Kolodner ed. 1989) (good sampling of current research); Kolodner, Rissland &
Waltz, Case-Based Reasoning from DARPA: Machine Learning Program Plan, in PROCEEDINGS:
CASE-BASED REASONING WORKSHOP 1 (1989) (sponsored by Defense Advanced Research Projects
Agency; sets forth major classes of CBR, ingredients of CBR, and outstanding research issues).
26. "Machine learning" is the ability of a program to change itself, particularly its knowledge, so
that it can perform better.

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
have been made to precedent-based argumentation,27 and progress is being

made on explanation (goal 5). Thus one can be optimistic that someday
Al will be able to deal with many of the key goals for an Al and law
program in at least a rudimentary manner. I would go so far as to say
that the first five desiderata are attainable now or in the near future.
The next two desiderata-handling exceptions and conflict (goal 6) and
accommodating to change in the base of knowledge (goal 7)-presently
are being addressed broadly in Al, particularly in machine learning. It is
reasonable to expect Al programs to be able to handle gradual change
soon, but not abrupt or rapid change.28 Non-monotonicity is being vigor-
ously explored by an active cohort of researchers, primarily using ap-
proaches grounded in sophisticated logics. Thus, these two desiderata
might be met someday.
As for the last three desiderata, I am much less optimistic. Modeling
common sense reasoning, knowledge of intent and belief, and natural lan-
guage capabilities are by far the hardest tasks. With regard to language,
some capabilities like interactive dialogue and understanding of short
summaries can be achieved in narrowly circumscribed domains; fuller ca-
pabilities are far off.29 Widely applicable general "solutions" are very dis-
tant. Some, like the understanding of written appellate opinions-the ulti-
mate fantasy-I expect never to see.30
II. SIGNIFICANT STEPS TOWARD A MODEL OF LEGAL ARGUMENT
This Section discusses several landmark projects in Al and law that

provide significant steps toward the long range goal of understanding and
modeling legal reasoning, particularly legal argument. They ad-
dress: (1) reasoning with rules: (2) reasoning with open-textured con-
cepts; (3) reasoning with cases, hypotheticals, and precedent-citing argu-
mentation; (4) mixed paradigm reasoning with rules and cases; and (5)
deep models of legal knowledge.
27. See infra Section II.C.

28. Changes of the type involved in "normal science," to use Kuhn's term, are amenable to tech-
niques from machine learning. In contrast, the types of change involved in a Kuhnian "paradigm
shift" are probably not, at least in the sense of the system on its own recognizing the need for a shift
and carrying it out through the creation of novel concepts. See T. KUHN, THE STRUCTURE OF SCIEN-
TIFIC REVOLUTIONS 10, 66 (2d ed. 1970); see also E. LEVI, supra note 19, at 8, 9 (describing change
in law).
29. See generally, Lehnert, Knowledge-Based Natural Language Understanding, in EXPLORING
ARTIFICIAL INTELLIGENCE: SURVEY TALKS FROM THE NATIONAL CONFERENCES ON ARTIFICIAL
INTELLIGENCE 83 (1988).
30. This is not so much because of any special difficulties with opinions, but because understand-
ing text is just plain hard. See supra notes 13, 29. Legal opinions use specialized vocabulary and
conventions, so they might be harder because these considerations require extra processing; on the
other hand, the extra constraints these considerations impose might make legal opinions more amena-
ble to computational analysis.

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
A. Reasoning with Rules
One of the earliest steps toward a model of legal reasoning was the use
of expert systems31 to model certain rule-based aspects of law.32 This step
reflects the development of Al: Rule-based expert systems were the first
type of Al system to become widely available and employed beyond the Al
research community.33 Furthermore, their underlying computational
mechanisms are conceptually clear and they have many computational
strengths. While from the legal standpoint there is a variety of opinions as
to the validity, usefulness, and status of rules, and there are acknowledged
difficulties in representing them,34 it is still quite natural to take some
body of legal rules and embed them in a standard rule-based computa-
tional framework.
In the rule-based approach, a rule is encoded in a simple, stylized if-
then format: If certain conditions are known to hold, then take the stated
action or draw the stated conclusion.35 Rule-based systems work by chain-
ing these rules together.36
31. An "expert system" is a special-purpose computer program, which can be said to be expert in
a narrow problem area. Typically, such a program uses rules to represent its knowledge and to rea-
son. See, e.g., P. HARMON & D. KING, EXPERT SYSTEMS (1985); D. WATERMAN, A GUIDE TO
EXPERT SYSTEMS (1986); B. BUCHANAN & R. SMITH, Fundamentals of Expert Systems, in 4 HAND-
BOOK OF ARTIFICIAL INTELLIGENCE, supra note 2, at 149.
32. An even earlier effort related to rule-based aspects of law was Layman Allen's work on "nor-
malization." His emphasis was on eliminating syntactic ambiguity in statutes and legal documents
rather than on using computational programs to reason with them. See Allen, Symbolic Logic: A
Razor-Edged Tool for Drafting and Interpreting Legal Documents, 66 YALE L.J. 833 (1957). Of
course, if certain ambiguities, for instance those about the scope of logical connectives and the meaning
of words like "if," "except," and "unless," were eliminated from legal sources, encoding legal rules for
use by an expert system would- be easier and less open to debate. Allen & Saxon, Some Problems in
Designing Expert Systems to Aid Legal Reasoning, in THE SECOND INTERNATIONAL CONFERENCE
ON ARTIFICIAL INTELLIGENCE AND LAW: PROCEEDINGS OF THE CONFERENCE 94 (1987) [hereinaf-
ter ICAIL-871 (discussing 48 alternative interpretations of structure of proposed limitations on exclu-
sionary rule).
33. For example, the expert system DENDRAL has been widely used by organic chemists, see R.
LINDSAY, B. BUCHANAN, E. FEIGENBAUM & J. LEDERBERG, APPLICATIONS OF ARTIFICIAL INTEL-
LIGENCE FOR ORGANIC CHEMISTRY: THE DENDRAL PROJECT (1980); sources cited supra note 31.
34. Besides "syntactic" difficulties, see Allen, supra note 32, there are "semantic" difficulties such
as the presence of conflicting rules, imprecise terms, and incompleteness. See Berman & Hafner,
Obstacles to the Development of Logic-Based Models of Legal Reasoning, in COMPUTER POWER AND
LEGAL LANGUAGE 183 (C. Walter ed. 1988) (discussing difficulties with logic-based approaches);
Berman, Cutting Legal Loops, in ICAIL-89, supra note 1, at 251 (discussing definitional circularity,
recursion, and self-referencing in statutes).
35. For instance, Rule 1, concerning "responsibility for use of the product," from the case settle-
ment system of Waterman and Peterson states, "IF the use of [the product] at the time of the plain-
tiff's loss is foreseeable and (that use is reasonable-and-proper or that use is an emergency or (there is
a description by the defendant of that use and that description is improper) or there is not a descrip-
tion by the defendant of that use) THEN assert the defendant is responsible for the use of the prod-
uct." D. WATERMAN & M. PETERSON, MODELS OF LEGAL DECISION MAKING 37 (Institute for
Civil Justice of The Rand Corporation Memo R-2717-ICJ 1981) (parentheses used to denote scope
and distribution of logical connectives; thus, here there are two principal antecedents, second of which
can be satisfied in alternative ways).
36. The systems can work either "forward" by reasoning from facts to a desired conclusion sup-
ported by them, or "backward" from a desired conclusion to find facts supporting it. Forward chain-
ing simply is the repeated application of the logical inference rule modus ponens: If one has a rule "If
A then B" and the fact A, conclude the fact B. Alternatively, in backward chaining, to establish the

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
The rule-based approach is particularly useful because in many do-

mains much of an expert's knowledge is amenable to expression in if-then
rules, many of which are "heuristic."37 Heuristics are typically an ex-
pert's individual synthesis of past problem solving, and they capture meth-
ods for making educated hunches. Expert systems provide a straightfor-
ward way to harness heuristic expertise, expressed as rules.
The first uses of the expert systems approach in law were those of
Donald Waterman and Mark Peterson at the RAND Corporation's
Center for Civil Justice.38 In their systems, rules were used to encode
statements of doctrine as well as legal expertise, cast as rules of thumb.39
Their work, done while expert systems approaches were beginning to be
widely applied, is representative of both what can be accomplished in law
with the expert systems methodology and also of the difficulties inherent
in such an approach. Waterman and Peterson investigated how the expert
systems approach fared as a practical tool in legal applications, and
demonstrated it as a methodology for modeling legal expertise.40
The problem area of one of their systems, called LDS (for Legal Deci-
sion-making System), was assessing the "worth" of a products liability
case for settlement purposes. The program, using rule-based models for
strict liability, comparative negligence, and calculation of damages, was
able to compute the worth of a case, to capture chains of inferences sup-
porting a conclusion like negligence, and to model certain aspects of settle-
ment negotiation. The project demonstrated the applicability of rule-based
techniques for legal applications.
Perhaps their system's greatest weaknesses were that it glossed over dif-
ficulties inherent in reasoning with imprecise terms and it underplayed
the adversarial nature of legal reasoning. For instance, LDS asked the
user whether the use of the product was "foreseeable."'" This is a notori-
ously subtle question, whose answer is open to interpretation and is en-
tirely contextual. One can argue that coming to a conclusion about the
foreseeability of use is at the heart of reasoning in the negligence area,
fact B, one looks for such a rule and verifies that one has information to satisfy its precondition, A; if
A were not satisfied, one then would look for rules establishing A, and so on until the necessary
factual basis were reached and the desired conclusion logically supported. See W. SALMON, LOGIc (2d
ed. 1973); sources cited supra note 31.
37. A "heuristic" is a rule of thumb, a bit of encapsulated wisdom. Much problem-solving behav-
ior is guided by heuristic rules of thumb, such as "If it is fourth down and long yardage is required,
then punt," or "If you need to get your citations in correct law journal form, then ask a law journal
editor." Heuristics are methods that past experience has shown to be "good" or "wise" things to do;
they do not necessarily guarantee a solution, as might algorithms or mathematical theorems, and
occasionally they might even produce wrong answers or lead in counter-productive directions. The
word "heuristic" stems from the Greek for invention or discovery. The mathematician George Polya
discussed heuristic reasoning and the use of heuristics in mathematical reasoning and problem solving.
See G. POLYA, How To SOLVE IT (1973).
38. See generally D. WATERMAN & M. PETERSON, supra note 35.
39. See, e.g., supra note 35 (example of one of their rules).
40. See D. WATERMAN & M. PETERSON, supra note 35, at 13-15.
41. Id. at 45..

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
and that by asking the user to make such an in

has given away the whole game. LDS also illustrates another general
problem with the rule-based approach: it does not give enough promi-
nence to the adversarial nature of legal reasoning, where the opposing
sides seek to establish different, and often contradictory, conclusions. Per-
haps more fundamentally, the rule-based approach assumes that the set of
rules has no inherent difficulties, like ambiguities, gaps, and conflicts. To
make a rule-based system work, the programmer must usually eliminate
these problems and make the rules appear more consistent and complete
than they are.
Although oversimplified, the system built by Waterman and Peterson
was a landmark in unknown territory. It is unfair to condemn it for fail-
ing to solve all the hard problems of modeling legal reasoning. To their
credit, Waterman and Peterson confront these criticisms head-on in their
own critiques. For instance, they address both the difficulty in handling
imprecise legal predicates and the need for subtler reasoning with such
42
concepts.
Today, expert systems are one of the most important approaches used
and there are many additional examples of current applications. Two re-
cent efforts in Britain are particularly noteworthy. The first is the Latent
Damage System by Richard Susskind and Phillip Capper, which ad-
dresses a problem area concerning the 1986 British Latent Damage Act.43
Their system addresses the complex legal issues relating to time periods
within which claimants may start proceedings when the damage or suf-
fered loss was "latent."" Besides providing a useful application in an in-
tricate body of law, this work is also interesting in the way in which Suss-
kind explains the project in terms of analytic jurisprudence.45
The second effort is a group of projects of the Logic Programming
Group at Imperial College, University of London, one of which modeled
the recent 1981 British Nationality Act." A major goal of this project was
to test the suitability of rule-based approaches for representing large, com-
plicated statutes. The process of modeling the statute with rules helped
the programmers uncover problems specific to the Act, such as undefined
legal predicates and loopholes in the Act, as well as a variety of general
problems in the rule-based approach concerning the formalization of ne-
gation, the use of defaults and counterfactuals, the representation of com-
42. See id. at 26. Waterman, a pioneer in machine learning, understood how learning issues were
also critical to expert systems. He was well ahead of his time and his early death in 1987 was a great
loss.
43. Latent Damage Act, 1986.
44. P. CAPPER & R. SUSSKIND, LATENT DAMAGE LAW-THE EXPERT SYSTEM (1988). The
book contains diskettes for installing and running the system on a machine like the IBM PC or X
45. See R. SUSSKIND, EXPERT SYSTEMS IN LAW (1987).
46. British Nationality Act, 1981. See Sergot, Sadri, Kowalski, Kriwaczck, Hammond & Cory,
The British Nationality Act as a Logic Program, 29 COMM. ACM 370 (1986) [hereinafter Sergot].

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
mon sense knowledge, and the use of administrative discretion in statutes,

each of which poses technically interesting problems for AI.47
Although rule-based techniques are but one component in a full model
of legal argument, they can be satisfyingly powerful from the perspective
of building useful applications. There are abundant examples of applica-
tions, many motivated by a need for practical, labor-saving tools in the
trenches, where expert resources are terribly strained.48 Given the availa-
bility of inexpensive expert systems development tools and rule-based pro-
gramming languages, such application systems will continue to proliferate;
anyone having access to such tools can write one.49 In summary, rule-
based techniques have demonstrated utility for frequently performed anal-
yses on stereotypical cases in stable, well-developed bodies of law.50
B. Reasoning with Open-Textured Concepts
Whether there are run-of-the-mill cases amenable to straightforward

rule-based analysis is a question with a long jurisprudential history. This
brings us to understanding the limitations of rule-based reasoning and the
special problems inherent in legal concepts. Anne Gardner's work both
elucidates the limitations of rule-based approaches and also points the
way toward the use of case-based techniques.5' Gardner investigated the
adequacy of rule-based approaches through the framework of "hard/easy
questions."52 In the hard/easy paradigm, a question is considered "easy"
if the legal experts agree as to its analysis or outcome; otherwise, it is
"hard." Hard questions typically arise in conjunction with "open-
textured" legal concepts.53 To implement this distinction in a program,
47. See, e.g., Sergot, supra note 46, at 379 (defining "negation as failure" as the conclusion that
something is false if all known ways of showing it true fail); id. at 382 ("counterfactual conditionals"
as in the statutory phrase "became a British citizen by descent or would have done so but
for ...."); id. at 382 (discretion as in the phrase "If ... the Secretary of State sees fit"); Berman,
supra note 34; Berman & Hafner, supra note 34.
48. See, e.g., Grady & Patil, An Expert System for Screening Employee Pension Plans for the
Internal Re7venue Service, in ICAIL-87, supra note 32, at 137 (describing Internal Revenue Service
project to process employee pension plans); Pethe, Rippey & Kale, A Specialized Expert System for
Judicial Decision Support, in ICAIL-89, supra note 1, at 190-94 (system for processing claims under
The Federal Black Lung Benefits Act); Weiner, CACE: Computer-Assisted Case Evaluation in the
Brooklyn District Attorney's Office, in ICAIL-89, supra note 1, at 215-23 (describing system for
post-arrest/pre-trial processing of drug busts by police).
49. See, e.g., Wiehl, Computers Assuming New Roles at Law Firms, N.Y. Times, Jan. 20, 1989,
at B4, col. 3; Blodgett, Artificial Intelligence Comes of Age, 73 A.B.A. J. 68 (1987).
50. For example, Paul Brest has commented that constitutional law would be "wildly unsuited for
an expert system because the principles are vague . . . . Expert systems are best for areas of law that
are rule-bound." Blodgett, supra note 49, at 70 (quoting Brest).
51. See generally A. GARDNER, AN ARTIFICIAL INTELLIGENCE APPROACH TO LEGAL REASON-
ING (1987). Gardner holds a J.D. from Stanford and practiced law before returning to Stanford to
obtain a Ph.D. in computer science.
52. For a discussion of this framework, see Hart, Positivism and the Separation of Lauw and
Morals, 71 HARV. L. REV. 593 (1958), reprinted in H.L.A. HART, ESSAYS IN JURISPRUDENCE. AND
PHILOSOPHY (1983); R. DWORKIN, supra note 23, at 81.
53. For a discussion of "open textured" legal concepts, see H.L.A. HART, THE CONCEPT OF
LAW 121-32 (1961).

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
several specifics must be provided, such as who are the experts to be used,
what counts as a disagreement among them, and how to tell if a disagree-
ment exists. From a practical standpoint, if one could sift the easy from
the hard, one could employ rule-based techniques, for instance, to solve
the easy questions, and attack the hard questions with other methods.64
Gardner built a computational model for the hard/easy paradigm and
for reasoning with open-textured legal concepts. The problem area of her
program was classic offer-and-acceptance law, and its task was to analyze
issue-spotter questions, taken from sources such as bar exams and Gil-
bert's Law Summaries. The program analyzed what the legal questions
were, spotted which were hard, and answered the easy ones. Gardner re-
quired that her program be computationally reasonable; it should not take
an unduly hard computation to decide which questions are easy and hard,
or else nothing much would have been gained, at least computationally, if
not theoretically.
In Gardner's model, hard questions arise from problems with rules in
two ways: either the rules relevant to the domain, such as those from the
Restatement of Contracts,55 are incomplete, circular, or contradictory; or
the legal predicates used in the rules are indeterminate and their interpre-
tation cannot be resolved.6 When there is a mixture in the disposition of
cases used to resolve these problems, Gardner's program categorizes the
problem as hard. Typically it is not possible to handle indeterminate con-
cepts because, as Gardner puts it, the rules "run out;" that is, a needed
rule's premise uses an undefined term.57
Gardner's approach to these problems was to give her program a rich
body of knowledge and several powerful heuristics. Her program's knowl-
edge included (1) Restatement-like rules for the doctrine of offer and ac-
ceptance,58 (2) a "network" to represent various states in which parties
54. This is of course not the only way to model the hard/easy problem; for instance, case-based
techniques could be used for both types of questions. However, given that rule-based methods are
well-understood, Gardner's approach is quite natural.
55. RESTATEMENT (SECOND) OF CONTRACTS (1981).
56. Note that a typical rule-based approach to handling difficulties with the rule set is to resolve
conflicts by hand before encoding the rules in the program. Of course, one cannot always eliminate
conflicts in the rule set because even the experts disagree. For example, Gardner's program has pairs
of conflicting rules about a rejection that revokes a previous acceptance of an offer, and about simulta-
neous acceptance and proposal to modify the contract. See A. GARDNER, supra note 51, at 134. Gard-
ner deliberately leaves conflicting rules in her system because she wants her model to be able to
handle the fact that legal experts, and thus their rules, can be in conflict. See id. at 3-4. A typical
rule-based way to finesse problems with open-textured legal predicates is either simply to ask the user
to make the judgment or engage in what might be called "definitional backchaining:" write rules
whose pre-conditions specify whether a legal predicate obtains, and then write rules defining what
those pre-conditions mean, and so on. See Sergot, supra note 46, at 378 (implementing this approach);
D. WATERMAN & M. PETERSON, supra note 35, at 45 (suggesting this approach).
57. A. GARDNER, supra note 51, at 33-34 (discussing problem of inadequacy of rules). This
brings the traditional rule-based technique (of backward chaining) to a complete standstill before an
adequate definitional basis is reached and application of the ambiguous term can be resolved.
58. For example, her program uses about twenty rules covering contract doctrines such as offer
and acceptance. See id. at 133.

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
could be in offer and acceptance situations, and transitions between these

states,69 (3) relevant common sense knowledge,80 and (4) prototypical fact
patterns (exemplars) for certain key concepts."' The exemplars serve as
simplified precedents in that their classifications have already been deter-
mined and are used to interpret new cases. For convenience in this discus-
sion, the rules, network, and common sense knowledge will be called as a
group the "non-exemplar" knowledge.
Gardner's program tries to answer a question by first using the non-
exemplar knowledge, and then, if that fails, by using the exemplars,
which function as prototypical clear cases. If an answer can be derived
with non-exemplar knowledge, the validity of the answer is checked
against relevant exemplars.2 Thus, her program uses exemplars both in a
primary way to derive an answer, and in a secondary way to check the
validity of an answer. In any case, if there is a mix in the disposition of
relevant exemplars, the question is considered hard.3
Gardner took the skeleton of an idea from legal philosophy and pro-
vided it with computational flesh. Obviously, hers is not the only model
that one can propose for the hard/easy paradigm, but it is a computation-
ally disciplined one. By examining the model's computational underpin-
nings and assumptions, one can learn something about the hard/easy dis-
tinction.64 Her work raises interesting questions about the use and content
59. For example, there can be an offer pending, and an acceptance will enable one to make a
transition to the state of being in a contractual relationship. See id. at 123-25; id. at ? 6.2.
60. For instance, that a telegram is a kind of document, which in turn is a kind of inanimate
object, which in turn is a kind of physical object. See id. at 90-91. See generally id. at 85-117 (very
detailed presentation of representation issues).
61. For example, her program has two prototypical fact patterns for the legal predicate "produces
a manifestation with content:" making an utterance with some content, and sending a document with
some content. See id. at 156-57.
62. The determination of hard/easy cases is made as follows. If a problem such as the application
of a legal predicate to the current facts can be (tentatively) answered using the program's non-
exemplar domain knowledge, primarily rule-like in character, and if there are no opposing case exem-
plars, then the question is deemed easy and its answer is the one derived. If the case exemplars used
to check the tentative answer all point the opposite way, then the question is also considered easy but
the answer is that supported by the case exemplars. If, on the other hand, there is a mixture in the
disposition of the case exemplars, then the question is flagged as hard and the program does not
attempt to provide an answer. If a (tentative) answer cannot be derived with the domain rules be-
cause, for example, a predicate cannot be resolved, but all the relevant cases point the same way and
thus can be used to posit and support an answer, then the question is considered easy and the answer
is that indicated by the case exemplars. On the other hand, if an answer cannot be derived using the
domain rules and there is a mixture in the cases, then the question is also deemed hard. See id. at
54-55, 160-61 (abbreviated descriptions of her program's algorithm).
63. One might ask, why not go to the cases right off? My interpretation is that the reasoning
involving the other knowledge, the rules, the transition network, and the common sense knowledge,
provides a means by which to index relevant cases. Furthermore, even in the easy cases, one would
need to derive the answer by some means, and in her model this is attempted with the rule-guided
reasoning, which would need to be done at some point anyway. Alternatively, one could omit looking
at the rules altogether and just consider cases. Of course, one would need to say which cases to
consider and what exactly to do with them. See infra Section II.C (discussing work of Ashley).
64. For instance, too many questions might be categorized as "hard" in this model since there is
very often a mixture in the disposition of the cases. Perhaps one needs to be more discriminating in
the use of cases.

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
of cases, and about which cases to include in a decision maker's case

base.65
While Gardner's work certainly addresses some of the shortcomings of
purely rule-based models, her program still falls short with regard to rea-
soning with cases. Some of the core features of precedent-based argumen-
tation, such as determining relevancy and drawing analogies, are beyond
the scope of her program; these issues are the focus of the next set of
projects to be explored here. Nonetheless, Gardner's research provides an
exceptionally important step toward the goal of understanding and model-
ing legal argumentation.
C. Precedent-Based Reasoning with Cases and Hypotheticals
The next landmark Al project was Kevin Ashley's, which was done in
this author's research group at the University of Massachusetts.66 Ashley
developed a program called HYPO to model certain aspects of case-based
reasoning as exemplified by appellate-style argumentation with prece-
dents. The problem area of this project was trade secret law. The task
was to produce elements of precedent-based argument. The HYPO pro-
ject focused solely on reasoning with cases, and was the first project-not
only in Al and law, but also in Al in general-to attack squarely the
problem of reasoning with cases and hypotheticals in a precedent-based
manner.
HYPO performs as follows: Given a fact situation, HYPO analyzes it
according to its model of trade secret law and then retrieves relevant cases
from its knowledge base of cases. It then determines which relevant cases
are most on point, or potentially so, for whose point of view, and from
which analytic approach. HYPO then generates the skeleton of an argu-
ment. In such an argument snippet, HYPO first argues for side one
(plaintiff or defendant) by making a legal point and citing its best, mos
on point cases; then it argues for side two by responding with a counter-
point, citing a most on point case supporting side two's point of view or
also by distinguishing the current facts from side one's cases; and finally,
HYPO argues again for side one with a rebuttal of side two's position,
which may include distinguishing side two's cases and strengthening the
65. For instance, in Gardner's program cases are prototypical fact patterns rather than actual
cases. What about using real fact patterns instead, or in addition? Which ones? How should the
program's memory of cases be organized? Are certain cases to be preferred to others?
66. See generally K. ASHLEY, MODELING LEGAL ARGUMENT: REASONING WITH CASES AND
HYPOTHETICALS (forthcoming 1990) (originally completed as Ph.D. dissertation, Department of
Computer and Information Science of the University of Massachusetts, Amherst, February, 1988,
under this author's direction); Ashley & Rissland, A Case-Based Approach To Modeling Legal Ex-
pertise, IEEE EXPERT, Fall 1988, at 70-77 (short overview and example). Ashley holds a J.D. from
Harvard Law School and practiced as a litigator with White & Case in New York before returning to
study computer science at the University of Massachusetts.

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
relationship of side one's analysis to the current facts.7 At various points

in the argument, HYPO may generate and employ hypotheticals-for in-
stance, to refute a claim when no actual counter-example case exists in
HYPO's knowledge base of cases.8
HYPO is able to: assess relevancy of cases; decide for each side which
relevant cases are most on point and which of these are best to cite in
argument; analogize and distinguish cases; generate and reason with hy-
potheticals; cite various types of counterexamples to a point; and construct
the skeleton of a case-citing argument. HYPO does not in any way at-
tempt to bring in policy-level concerns or argumentation; rather, it sticks
to arguing with cases, on their facts, in a technical way. HYPO also does
not include other aspects of legal reasoning, such as reasoning with rules.
A key feature of HYPO is a kind of case index, called a "dimension,"
which it uses to retrieve and analyze cases. Dimensions represent impor-
tant legal factors. They encode the knowledge that the presence of certain
facts enables a case to be addressed from a certain point of view.9 A
dimension enables HYPO to retrieve a set of cases that support the same
analytic approach and to compare and assess the relative strength of the
cases within this group.70
HYPO uses dimensions to define concepts like "relevant," "most on
point" and "best" cases.7' In HYPO, a case is considered "relevant" if it
shares at least one dimension with the fact situation. HYPO ranks cases
according to how on point they are by examining the overlap between the
set of dimensions present in the fact situation and the sets of dimensions
present in the cases. If the overlapping set of dimensions of one case, B, is
contained within that of another case, A, then A is defined as more on
point than B, the rationale being that A has more lines of analysis in
common with the fact situation than does B.72
67. For several examples, see K. ASHLEY, supra note 66; Ashley & Rissland, supra note 66, at
74-76.
68. K. ASHLEY, supra note 66.
69. For instance, in HYPO's domain of trade secret misappropriation law, knowing facts about
the relative costs incurred by the plaintiff and defendant in developing their putative secrets into a
product enables one to make arguments about the gain of competitive advantage, and knowing facts
about the disclosures made by the plaintiff enables one to make arguments about how well the plain-
tiff kept his knowledge secret. See, e.g., Gilburne & Johnston, Trade Secret Protection for Software
Generally and in the Mass Market, 3 COMPUTER L.J. 211 (1982). For further explanation and
examples, see K. ASHLEY, supra note 66; Rissland, Valcarce & Ashley, Explaining and Arguing
with Examples, in PROCEEDINGS OF THE NATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE
288, 291-93 (1984).
70. In HYPO, for instance, the greater the disparity between the development costs of plaintiff
and defendant, the easier it is to argue that, all other things being equal, the defendant gained an
unfair competitive advantage through misappropriation of plaintiff's trade secret. Thus, for example,
since there is a case in which the plaintiff prevails with a ratio of two to one for plaintiff's to defend-
ant's costs, see Telex Corp. v. IBM Corp., 367 F. Supp. 258 (N.D. Okla. 1973), then a new case with
a ratio of four to one would represent an even stronger position for the plaintiff in HYPO.
71. See K. ASHLEY, supra note 66.
72. So, for instance, suppose the dimensions that apply to the current fact situation (CFS) are W.
X, Y, and Z. Suppose Case A shares X, Y, and Z with CFS, and Case B shares just X and Y. Then

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
By then taking into account which cases support or cut against a party,
their relative strength along shared dimensions, and the dimensions not
shared, HYPO is able to analogize, distinguish, and otherwise manipulate
cases to construct snippets of precedent-citing argument. For instance, to
distinguish a case that the opposing party, say the plaintiff, has cited as
controlling the outcome of the current fact situation, HYPO looks at the
cited case for the existence of dimensions unshared with the current situa-
tion. HYPO then argues, for the defendant, that the presence or absence
in the current factual setting of these dimensions diminishes the applica-
bility of the cited case; in other words, these are differences that really
make a difference."
To summarize, the HYPO project models some of the key ingredients
of reasoning with precedents. The model provides computational defini-
tions of relevant cases, on point cases, and best cases to cite.74 Because of
the precision of HYPO's model, a researcher can examine specific details
and assumptions.75 By providing an analysis and computational model for
reasoning with cases, perhaps the most vital aspect of legal reasoning,
Ashley's work is a giant step toward the goal of understanding legal
argumentation.
An earlier effort addressing some of the same concerns as HYPO was
the work of Thorne McCarty on legal argument.76 One of McCarty's
longstanding goals has been the construction of knowledge representation
mechanisms with which to model certain aspects of legal argumentation.
For instance, McCarty has proposed a mechanism called "prototypes and
deformations" to model certain aspects of appellate argument. He has
used it to examine the sort of situation in which one incrementally steps
from a desirable precedent to the current case, through a sequence of in-
termediate real and hypothetical cases, in order either to construct an ar-
gument by analogy for the result one desires, or else to show the inconsis-
tency of the result argued for by one's opponent.77 His model requires a
Case A is more on point than Case B because the set that B shares with CFS is a subset of the set
that A shares. Suppose there is a third relevant case, Case C, sharing dimensions W and X with CFS;
Case C is neither more nor less on point than A or B. Both Case A and Case C are most on point
cases since each shares maximally with respect to a subset of dimensions.
73. For instance, if the case cited by the plaintiff is very strong on a pro-plaintiff dimension not
present in the current case, the defendant can argue that that particular dimension is responsible for
the pro-plaintiff outcome. If, in addition, the defendant can point to a different case, the same in all
respects to the plaintiff's cited case except for that one dimension, in which the defendant prevailed,
then the assignment of credit for the outcome in the plaintiffs case to the missing dimension is all the
more convincing.
74. For instance, the best cases for a side to cite are defined as those most on point cases for that
side that share at least one applicable dimension favoring that side.
75. For instance, an alternative way to measure relevancy is to minimize the sets of dimensions
not shared. One could then swap this alternative definition for relevancy in the HYPO model and
observe the repercussions.
76. See generally McCarty, Reflections on TAXMAN: An Experiment in Artificial Intelligence
and Legal Reasoning, 90 HARV. L. REV. 837 (1977).
77. See McCarty & Sridharan, The Representation of an Evolving System of Legal Concepts: II

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
great deal of attention to details concerning the representation of cases and

their constituent parts.78 Thus, although some of McCarty's research is
clearly tied to that on cases, hypothetical, and argument, the bulk of his
work is more aptly grouped with efforts on developing "deep models" of
legal knowledge, which concentrate more on developing highly detailed
models of knowledge than the other projects discussed thus far, and will
be discussed briefly later.79
One of the shared concerns of McCarty and this author is the use of
hypotheticals in legal arguments and development of computational meth-
ods to generate them.80 For example, HYPO uses hypotheticals to fulfill
the need for counter-examples to refute another's position when real cases
cannot be found or to show the sensitivity of a given fact to the interpreta-
tion of a case.8' It uses its internal mechanisms, like dimensions, to gener-
ate hypothetical, for instance by adding, deleting, or changing a fact af-
fecting a dimension.82
The conceptual framework provided by such Al models can also be
used to analyze expert use of hypotheticals in Socratic law school dia-
logues, advocacy, judical opinion, and appellate oral argument.83 Improve-
ments of advocacy and pedagogical skills are two potential benefits from
these analyses. Another is increased understanding of the development of
legal concepts and theories. For instance, a good hypothetical provides a
way to test the validity of a theory: it provides a gedanken experiment.
This is especially valuable when there are very few real cases. Hypotheti-
cals play a special role in case-based reasoning, and increased understand-
ing of them would contribute to our goal of modeling legal reasoning.
Another contribution to understanding the case-based aspects of legal
reasoning would be to understand how new arguments could be generated
by modification of old arguments, without starting "from scratch" from a
statement and analysis of the facts, proceeding through an analysis of rele-
Prototypes and Deformations, PROCEEDINGS OF THE SEVENTH INTERNATIONAL JOINT CONFER-

ENCE ON ARTIFICIAL INTELLIGENCE 246 (1981) (discussion of prototypes and deformations); L. MC-
CARTY & N. SRIDHARAN, A COMPUTATIONAL THEORY OF LEGAL ARGUMENT (Laboratory for
Computer Science Research, Rutgers University, New Brunswick, N.J. Technical Report No. LRP-
TR-13, 1981) (presenting example from Justice Pitney's opinion and Justice Brandeis' dissent in
landmark tax case Eisner v. Macomber, 252 U.S. 189 (1920)).
78. McCarty & Sridharan, supra note 77.
79. See infra Section II.E.
80. See, e.g., McCarty & Sridharan, supra note 77; L. MCCARTY & N. SRIDHARAN, supra note
77; Rissland, Examples in the Legal Domain: Hypotheticals in Contract Law, in PROCEEDINGS OF
THE FOURTH ANNUAL CONFERENCE OF THE COGNITIVE SCIENCE SOCIETY 96 (1982).
81. See K. ASHLEY, supra note 66; Rissland & Ashley, Hypotheticals as Heuristic Dev
CEEDINGS OF THE FIFTH NATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE 289 (1986).
82. For instance, one can make a trade secret misappropriation case incrementally weaker or
stronger by variation of a key fact, like the relative competitive advantage gained by the defendant
the number of disclosures made by the plaintiff. See, e.g., supra notes 69-70.
83. See, e.g., Rissland, Argument Moves and Hypotheticals, in COMPUTING POWER AND LEGAL
REASONING 129 (C. Walter ed. 1985) (examples of hypotheticals in Socratic teaching); Rissland &
Ashley, supra note 81, at 290-91 (examples in context of advocacy); McCarty & Sridharan, supra
note 77 (examples in context of judicial opinion).

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
vant precedents, and culminating in an argument, as HYPO describes. Of

course, a model to describe this alternative approach would require repre-
sentation and manipulation of a knowledge base of arguments. This is the
subject of the on-going research.84 Another step would be to explore more
fully the relationship between relevance and analogy.86
However, just as Waterman and Peterson artificially confined them-
selves to reasoning with rules, others, like Ashley, considered only reason-
ing with cases. From the long range perspective of developing an all-
encompassing model for legal reasoning, both approaches leave something
out, since many areas of the law are neither purely case-based nor purely
rule-based. Both types of reasoning are clearly needed in statutory areas,
where, despite the drafters' best efforts, there are always legal terms and
rules whose meaning and scope are not totally fixed and for which one
must look to cases for interpretation. The next major step toward a model
of legal reasoning is development of approaches to combine reasoning with
cases and reasoning with rules.
D. Reasoning with Rules and Cases
Reasoning in a statutory area like tax law is typical of mixed paradigm

reasoning. For instance, consider the subsection of tax law governing the
deductibility of expenses for a so-called home office. The tax code includes
requirements that the office be used regularly and exclusively, and that it
is the principal place of business or a place where one meets or deals with
clients.86 However, nowhere in the tax code are the requirements for what
constitutes "regular" or "exclusive" use, etc., set forth. All such terms
have been the subject of numerous cases,87 and to determine whether they
apply to a new situation, one must typically examine precedents and cre-
ate precedent-based arguments or subarguments to justify the interpreta-
tion. Of course, the argumentation must be tied back to the requirements
and structure of the statute.
Methods for accomplishing hybrid rule-based and case-based reasoning
go beyond those needed for pure case-based and or rule-based reasoning
in isolation. For instance, suppose that one had a rule with three pre-
conditions for the awarding of a certain benefit: "In order to receive bene-
84. See, e.g., Branting, Representing and Reusing Explanations of Legal Precedents, in ICAIL-
89, supra note 1, at 103 (developing model of how past arguments can be modified for re-use in new
case).
85. See K. BELLAIRS, CONTEXTUAL RELEVANCE IN ANALOGICAL REASONING: A MODEL OF
LEGAL ARGUMENT (submitted to the University of Minnesota, 1989) (Ph.D. thesis) (how relevance
can be defined in terms of analogy). This work continues the examination begun by Ashley of what it
means for one case to be relevant to another. It draws on work such as Gentner's, supra note 12,
which views analogy as a mapping between structured objects like legal concepts and cases.
86. I.R.C. ? 280A(c)(1) (1988).
87. For example, cases examining "exclusive use" include Frankel v. Commissioner, 82 T.C. 318
(1984), Baie v. Commissioner, 74 T.C. 105 (1980), and Chauls v. Commissioner, 41 T.C.M. (CCH)
234 (1980).

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
fit B, one must satisfy requirements R1, R2, and R3." Suppose that R1
and R2 are clearly satisfied. To argue for receiving the benefit, one has
several strategies to consider. For instance, one can try to find on point
precedents addressing the interpretation of antecedent R3, and then argue
that, in fact, the facts actually do satisfy R3 and thus the rule. Alterna-
tively, one can try to find on point precedents in which only the first two
requirements were met and the benefit was awarded nonetheless, and then
argue that the first two are sufficient by themselves. These are examples
of "near miss" strategies. To carry them out requires not only the use of
both the rule and the cases, but also that both types of reasoning be done
in concert. It is not good enough to confine one's attention to the cases; the
case-based aspects must speak to the rule. Similarly, one gets nowhere
using the rules alone, since the rules "run out." Figuring out how to coor-
dinate the reasoning activities of the rule-based and case-based aspects is a
question of "control."88
Gardner's research suggests one model for integrating rule-based and
case-based reasoning: use cases when reasoning with rules comes to an
impasse, and also use cases to validate the conclusions reached using
rules.89 In this approach, case-based reasoning is subjugated to rule-based
reasoning. Of course, selecting the dominant mode of reasoning should
really depend on the context or circumstances: sometimes the rules should
drive the reasoning, and sometimes the cases.
A different approach for an Al program would be to have independent
processes-one for case-based tasks, one for rule-based tasks-where each
works from its own point of view independently. Later, an executive pro-
cess would integrate the results. Metaphorically, this is like sending off
associates to do specific tasks, such as library work on case law or statu-
tory law, and having the senior associate integrate the information and the
reasoning.
For a computational model, one must spell out when and how the
processes interact. In particular, how should the associate processes com-
municate their results with one another? Should they wait until they are
done with their individual tasks before communicating? Should they share
intermediate results and insights?
One Al model for integrating the work of the associate processes would
be the following: each process could have access to a common blackboard,
on which it could write anything interesting or useful it learns, and from
which it could read any piece of information it found to be interesting or
useful. In this model, the processes reason largely independently of each
other, and yet they can capitalize opportunistically on each other's results.
88. "Control" refers to issues concerning how a program is organized, how its parts interact, and,
in general, how it decides what to do next.
89. See supra Section II.B.

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
This is in fact the approach taken in so called

course, many more details must be specified before an actual AI and law
program could be built using this framework.91
In the same spirit, but slightly simpler, is the so-called "agenda-based"
model in which, at any given moment, an individual process is in control
as it tries to accomplish its assigned task.92 As the process performs its
assignment, it makes note of other tasks it would find necessary, interest-
ing, or useful to be accomplished by another process or itself. As a task is
being worked on, other tasks are proposed for consideration.9' Specifically,
the tasks spawned are added to an overall "wish list" of things to do.
Since computational resources are limited, this list (the agenda) is ordered
according to some standard, such as interest, novelty, or importance. The
task with the highest ranking is selected from the agenda and is worked
on next. Heuristic rules, like strategies for dealing with near misses, are
used to specify what sort of tasks should be proposed and how they should
be ordered.94
Blackboard and agenda-based systems behave in a dynamic, opportunis-
tic way, sometimes working from one approach, sometimes from another,
and deciding which task is best to do as it proceeds. The methods that a
system uses to manage the posting and execution of proposed tasks largely
define its overall style of reasoning.96
Research on hybrid reasoners is currently underway at several sites.
Two using an agenda-based approach are the PROLEXS Project at the
Vrije Universiteit in Amsterdam in the area of Dutch landlord-tenant
law,96 and the CABARET (for CAse-BAsed REasoning Tool) project of
this author's group at the University of Massachusetts in statutory areas,
90. See Nii, Blackboard Systems, in 4 HANDBOOK OF Al, supra note 2, at 3. The blackboard
model was first used in the 1970's in the HEARSAY II project for understanding continuous speech
(this has nothing to do with the hearsay rule in evidence). See, e.g., Erman, Hayes-Roth, Lesser &
Reddy, The HEARSAY-II Speech-Understanding System: Integrating Knowledge To Resolve Uncer-
taintT, 12 COMPUTING SURVEYS 213 (1980).
91. For example, one would need to determine which "knowledge sources," as the rule-based and
case-based associates would be called, have access to what information on the blackboard. For in-
stances, can some access more information than others? What is the relative import of their results?
92. The program AM used an agenda. See generally R. DAVIS & D. LENAT, supra note 7.
93. For instance, in our benefits rule example, the first task of applying the benefit rule would
spawn several near-miss tasks concerning the third antecedent, R3, such as finding cases to show R3
is satisfied or not needed.
94. The overall behavior of agenda-based systems is to work in a "best first" manner. Since the
decisions as to what counts as "best" are often based on evaluative heuristics, and at any given mo-
ment the tasks on the agenda represent alternatives, agenda-based systems perform "heuristic best first
search."
95. The management scheme can be used to bias the system to behave in a certain way. For
instance, one can bias the system always to look at rule-based tasks in preference to case-based ones.
96. See, e.g., Oskamp, Knowledge Representation and Legal Expert Systems, in ADVANCED Top-
ICS OF LAW AND INFORMATION TECHNOLOGY 195 (G. Vandenberghe ed. 1989); Oskamp, Walker,
Schrickx & van den Berg, PROLEXS Divide and Rule: A Legal Application, in ICAIL-89, supra
note 1, at 54.

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
such as the home office deduction.7 There is a rising tide of interest in

such systems throughout Al.
CABARET contains full-fledged rule-based and case-based reasoners,
each with its own dedicated "monitoring" process, which harvests obser-
vations and describes them in terms understandable by the agenda-based
controller. The controller, in turn, applies heuristics to these observations
to propose and order tasks.98 For instance, the rule-based monitor would
post the observation that a near-miss situation had obtained with regard
to antecedent R3 of our benefits example, and this would trigger near-
miss heuristics. CABARET currently uses approximately thirty heuris-
tics, grouped into ten or so categories, such as ways to begin reasoning, to
check on reasoning, to respond to failures in reasoning, to respond to near
misses, to broaden rules and concepts, and to narrow rules and concepts.99
These were developed through observations of experts as well as from
legal sources.100
E. Deep Models of Legal Knowledge
The goal of research on deep models is to allow an AI program to

reason about ingredients of its own legal and general knowledge by repre-
senting this knowledge in great detail.101 Types of knowledge used in deep
models include knowledge about temporal and spatial relations, quantity,
common sense taxonomies, relations among individuals and beliefs. This
topic is often addressed under the rubric of "deep models" and is very
much related to fundamental problems concerning common sense reason-
ing. The motivating assumption is that the richer the representation, both
in scope and detail, the better the reasoning. The need to use deeper mod-
els of knowledge will ultimately forge a link between the other research
97. See supra text accompanying notes 85-87; Rissland & Skalak, Interpreting Statutory Predi-
cates, in ICAIL-89, supra note 1, at 46; Rissland & Skalak, Combining Case-Based and Rule-Based
Reasoning: A Heuristic Approach, in 1 PROCEEDINGS OF THE ELEVENTH INTERNATIONAL JOINT
CONFERENCE ON ARTIFICIAL INTELLIGENCE 524 (1989) [hereinafter Rissland & Skalak, A Heuristic
Approach].
98. For some of the computational details, see Rissland & Skalak, A Heuristic Approach, supra
note 97.
99. Id. at 526.
100. See, e.g., W. TWINING & D. MIERS, How To Do THINGS WITH RULES (2d ed. 1982)
(discussing various problems, examples, and approaches for reasoning with statutes and rules).
101. By contrast, in most of the systems discussed in this Comment, perhaps with the exception of
Gardner's system, there is no method by which the program may reason about such knowledge ex-
plicitly. While questions of time and number come up, for instance, in HYPO's reasoning about
disclosure events, see supra notes 69, 82, HYPO cannot reason explicitly about time or numeracy as
topics in their own right. However, note that since HYPO knows that 10,000 disclosures is a far
worse number of disclosures from the plaintiff's point of view than two would be, in a sense HYPO
implicitly knows the difference between big and small numbers. One might even say that since, with
respect to the dimensions about disclosures, fact situations with the same number of disclosures are
treated as being the same (all other things aside), HYPO could be said to know what 2 or 10,000
means. However, numbers and their absolute magnitudes are not topics HYPO can reason about in
an explicit way. If HYPO were redesigned so that it used a deep model containing information about
numbers it would be able to do so.

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
discussed in this Comment and that research seeking to develop rich rep-
resentation schemes for legal knowledge.
Work on representation of legal knowledge is exemplified by the re-
search projects of Thorne McCarty'02 and Tom Gordon.'03 Deep models
stand in contrast to situations where one simply uses representation items
to encode domain knowledge, but cannot reason further about their mean-
ing. If one were to employ deep knowledge in a system like HYPO, for
instance, one would represent in much more detail information about such
things as disclosure events, companies, time periods of employment and
product development, obligations of employers and employees, and prod-
ucts. For instance, in a deep model an Al researcher could reify a relation
such as employer-employee at a deeper level, perhaps in terms of relations
of permission or obligation, so that the program could reason about the
employer-employee relation as a topic in itself.'04
III. CONCLUSIONS
Obviously, I cannot do justice to the number and variety of projects in

Al and law in this Comment. However, I hope my discussion gives some
idea of the richness and variety of current projects.
I foresee the continuation of work on rule-based, case-based, mixed
paradigm, hypothetical reasoning, and argument, and predict it will be
extended by results concerning deep models of representation, sophisti-
cated control regimes, and machine learning. These lines of research even-
tually will merge and culminate in an extensive theory of legal argumen-
tation. In time, more and more of the desiderata for Al models of legal
reasoning will be incorporated into running Al systems. One can expect
some research systems, like HYPO, to have an impact on practical sys-
tems, just as Waterman and Peterson's did.105 What was first a research
idea will find its way into practical systems.'06
102. See, e.g., McCarty, Permissions and Obligations, 1 PROCEEDINGS OF THE EIGHTH INTER-
NATIONAL JOINT CONFERENCE ON ARTIFICIAL INTELLIGENCE 287 (1983) (using "permissions" and
"obligations" to represent certain relations between parties in case); cf. Hohfeld, Some Fundamental
Legal Conceptions as Applied in Legal Reasoning, 23 YALE L.J. 16 (1913) (developing similar
relationship-type structures). The need for representing such deontic relationships in his scheme re-
quired McCarty to develop a great deal of logical apparatus, which has been the focus of his work for
many years. Of late, McCarty has been coming back to some of his primary concerns: development of
a comprehensive representational theory for legal knowledge and ultimately, legal argument. See, e.g.,
McCarty, A Language for Legal Discourse I. Basic Features, in ICAIL-89, supra note 1, at 180.
103. See Gordon, Issue Spotting in a System for Searching Interpretation Spaces, in ICAIL-89,
supra note 1, at 157 (1989); Gordon, OBLOG-2: A Hybrid Knowledge Representation System for
Defeasible Reasoning, in ICAIL-87, supra note 32.
104. Hohfeld, supra note 102.
105. See supra Section II.A.
106. For instance, I predict that work on case-based reasoning will lead to practical tools for
creating and managing case data bases of individual practitioners and firms, which can then be used
in preparation of new cases. A beneficial side-effect of such CBR tools, and of course, of traditional
expert systems, will be the capturing and preservation of a firm's "institutional memory" and its use
to leverage new or inexperienced attorneys in the areas of the firm's expertise to higher levels of

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
I have tried to show how Al and law researchers are pursuing their
twin goals of analytic and practical advances, and how past and ongoing
research can be viewed as a coherent attempt to model legal reasoning,
particularly argumentation. Even though we may be a long way from
some vision of the ideal legal reasoning Al program, and in fact may
never be able to achieve certain aspects of human expertise, we can al-
ready accomplish very interesting and useful projects. We can use the fine
lens of Al to explicate the process of legal reasoning, for instance the crea-
tion of case-based arguments; to shed light on questions in legal philoso-
phy, such as the nature of open-textured predicates; and to provide practi-
cal, even socially beneficial, applications, such as expert systems in certain
administrative areas. The insistence on using computational methods in
the AT approach provides a useful discipline for considering longstanding
issues in legal reasoning.
Although I have not discussed it here, this body of research can also be
fruitfully applied to the law school curriculum. For instance, we can pro-
vide our law students with environments in which to examine their own
legal knowledge.107 The conceptual framework of AT can also provide a
way to describe our own expertise, such as posing of artful hypotheticals,
and show students how they may also possibly acquire it. AT provides a
set of tools, based on detailed models of representation and process.
By this discussion, I hope both to encourage those wishing to attempt
AT-style projects, and also to reassure those unsure of whether we should.
For instance, some might be concerned that the use of AT models will
somehow trivialize legal reasoning by making it seem too simple, under-
mine the importance of lawyers and judges by relegating them to the role
of mere users of systems which do all of the interesting reasoning, or de-
humanize us by describing intelligent behavior in well-defined terms. I
think AT research shows just the opposite: The more we understand
human reasoning, the more we marvel at its richness and flexibility, the
more questions we ask as we try to understand its workings, and the more
we require of a computer program exhibiting intelligence.
Legal reasoning is complex. Our current AT models, albeit too simple,
are but steps to more subtle and complete models, and at each step we
understand more. There will always be a need for human lawyers and
judges. The goal is to assist, not to replace. Demystification of some of the
performance, at the very least by keeping them from asking "obvious" questions and making "silly"
mistakes.
107. For example, one can use currently available commercial expert systems shells to allow stu-
dents to build their own small applications and then to experiment with them. The very exercise of
developing a rule base forces students to develop a theory of the application area and to think about
general issues regarding the validity and appropriateness of using rule-based approaches. For in-
stance, developing rule sets for consumer tort law or offer-and-acceptance law requires understanding
of the specifics of the law as well as general issues about rules and legal predicates and problematic
aspects of them. See supra Sections ILA-B.

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC
most precious qualities of human intelligence is not to be feared; under-

standing does not destroy or diminish that which is understood. We
should not be afraid to know how we know.
Of course, use of Al techniques, whether in basic research or straight-
forward applications, requires detailed knowledge of their underlying as-
sumptions, simplifications, strengths, and weaknesses. Successful projects
require not only thorough knowledge of both Al and the law, but also a
willingness to try new approaches without knowing exactly where they
will lead, or whether or not they will achieve exactly the desired result. In
short, work in AT and law is no different from work in other fields: you
cannot get anywhere without trying; and one invariably learns from
trying.
In seeking to understand and model legal reasoning, AT will be chal-
lenged and enriched. By engaging in AT endeavors, the law will be chal-
lenged and enriched too; it will better understand its own modes of rea-
sonings, including the knowledge and assumptions underlying them, and
it will benefit from practical computational tools and models. The rela-
tionship between AT and law is a true synergy, the shared specialty of AT
and law adding value to both.

13.234.156.0 on Sun, 02 Aug 2020 10:39:47 UTC

This Content Downloaded From 13.234.156.0 On Sun, 02 Aug 2020 10:39:47 UTC

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

This Content Downloaded From 13.234.156.0 On Sun, 02 Aug 2020 10:39:47 UTC

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

This Content Downloaded From 13.234.156.0 On Sun, 02 Aug 2020 10:39:47 UTC

Uploaded by

Copyright:

Available Formats

Artificial Intelligence and Law: Stepping Stones to a Model of Legal Reasoning

Author(s): Edwina L. Rissland

Stable URL: http://www.jstor.com/stable/796679

This content downloaded from

This Comment discusses developments in the twenty-year-old interdis-

t Associate Professor, Department of Computer and Information Science, University of Massachu-

This content downloaded from

I. BACKGROUND ON ARTIFICIAL INTELLIGENCE

A. What is Artificial Intelligence?

AI is the study of cognitive processes using the conceptual frameworks

2. For additional background, see E. CHARNIAK & D. MCDERMOTT, INTRODUCTION To ARTI-

This content downloaded from

note that tasks involving "common sense" reasoning or perception, such as

This content downloaded from

more satisfying to exhibit a running program: to some degree success is a

This content downloaded from

insights offered by legal scholars about legal reasoning, an AI researcher

B. AI and Law: A Fruitful Synergy

1. Legal reasoning is multi-modal, rich and varied: it includes rea-

These observations suggest the possibility of fruitful synergy between

This content downloaded from

This content downloaded from

C. Some Desiderata for AI and Law Programs

Given the special characteristics of law described above, we can enu-

1. Reason with cases (both real and hypothetical) and analogies;

ers, in 3 IBM J. RES. & DEV. 211 (1959).

This content downloaded from

have been made to precedent-based argumentation,27 and progress is being

II. SIGNIFICANT STEPS TOWARD A MODEL OF LEGAL ARGUMENT

This Section discusses several landmark projects in Al and law that

27. See infra Section II.C.

This content downloaded from

A. Reasoning with Rules

This content downloaded from

The rule-based approach is particularly useful because in many do-

This content downloaded from

and that by asking the user to make such an in

This content downloaded from

mon sense knowledge, and the use of administrative discretion in statutes,

B. Reasoning with Open-Textured Concepts

Whether there are run-of-the-mill cases amenable to straightforward

This content downloaded from

This content downloaded from

could be in offer and acceptance situations, and transitions between these

This content downloaded from

of cases, and about which cases to include in a decision maker's case

C. Precedent-Based Reasoning with Cases and Hypotheticals

This content downloaded from

relationship of side one's analysis to the current facts.7 At various points

This content downloaded from

This content downloaded from

great deal of attention to details concerning the representation of cases and

Prototypes and Deformations, PROCEEDINGS OF THE SEVENTH INTERNATIONAL JOINT CONFER-

This content downloaded from

vant precedents, and culminating in an argument, as HYPO describes. Of

D. Reasoning with Rules and Cases

Reasoning in a statutory area like tax law is typical of mixed paradigm

This content downloaded from