Volume 1, Spring Issue, 1988 BOOKREVIEW AN ARTIFICIAL INTELLIGENCE APPROACH TO LEGAL REASONING By Anne vonder Lieth Gardner. Cambridge, Massachusetts: The MIT Press, 1987, pp. 193. Reviewed by Edwina L. Rissland* Anne Gardner is both a lawyer and a coruputer scientist. She obtained her J.D. from Stanford in 1958, ar:d her book is a revision of her 1984 dissertation submitted to Stanford's Department of Computer Science. She plays, in part, the role of pioneer; artifi- cial intelligence ("AI") techniques have not yet been widely ap- plied to perform legal tasks. Therefore Gardner, and this review, first describe and define the field, then demonstrate a working model in the domain of contract offer and acceptance. I, THE FIELD: AI AND THE LAW A. Artificial Intelligence Malwin Minsky, in his preface to the collection Semantic In- formation Processing, defined artificial intelligence as "the :science of making machines do things that would require intel- ligence if done by [people]. 'u Better known successful applications have included playing chess, identifying bacterial strains, recog- nizing and manipulating structures built with childrens' blocks and formulating concepts in set theory. 2 Typical programming languages include LISP and PROLOG, while typical specialties in AI research include case-based reasoning, expert (rule-based) systems, natural language processing, nonmonotonic reasoning, and learning from examples. Above all, the field is marked by its current diversity, rapid growth, and apparent potential2 * Associate Professor of Computer and Information Science, University of Massachusetts at Amherst; Lecturer on Law, Harvard LaW School. Professor Rissland will serve as Program Chair for the second International Conference of Artificial Intelligence and the Law, current- ly scheduled for late Spring, 1989 in London. Asecond, more technical version of this review will appear in the Fall, 1988 issue ofA/Magazine. I. SEMANTIC INFORMATION PROCESSING (M. Minsky ed. 1968). 2. See generally Waltz, Artificial Intelligence, SCI. AM., Oct. 1982. For a recent, readable overview, see Linden, Putti,~g Knowledge to Work, TIME, March 28, 1988. 3. Optimism is founded in part upon parallel developments in ha~'dware and operating systems, such as Intel's 386 and MicmsoR's OS/2. See, e.g., Gralla, Chips Offthe Old Block: A History, PC WEEK, Jan. 19, 1988 at S/13 (noting the AI capabilities of new microcomputers).224 Harvard Journal of Law and Technology [Vol. 1 B. Attraction The law is an attractive domain for AI research for several reasons. First, the law has a tradition of examining its own reasoning process. Second, legal reasoning is stylized: one reasons according to stare decisis, with cases and by analogy. Third, much legal knowledge is readily accessible and relatively well struc- tured, codified and indexed. Nevertheless it will not surprise, and may even please, lawyers to learn that the Restatement, the Uniform Commercial Code, and case law, like the theories of legal reasoning proposed by Karl Llewellyn, Ronald Dworkin, Edward Levi, and H. L. A. Hart, are of limited immediate use to AI programmers. There is want amidst plenty because the central questions-what we know and how we know it-are answered only partially by such material, and even these partial answers prove difficult to harness computationally. Lawyers' interest in AI techniques and software is nascent but concisely expressed in marketing terms. Lawyers use tools that gather, sift, and/or structure legal arguments in a cost-effective manner. LEXIS and WESTLAW, which are essentially keyword- based programs siftinglarge full-text databases, operate at a level of sophistication far below that contemplated by AI programmers such as Alane Gardner, Marek Sergot, 4 my colleague Kevin Ash- ley, 5 or many others. 6 Yet products such as LEXIS and WESTLAW do indicate the potential ubiquity of legal analysis software and the symbiosis between computer science and legal research. C. The Hurdles Several specific, familiar aspects of legal analysis that chal- lenge AI may be noted. First, logical deduction alone cannot resolve all legal issues. In Gardner's terms, legal reasoning is a "rule-guided" rather than rule-governed activity (p. 3). Legal rules have the status more of heuristics than of theorems, in that all rules have exceptions, and rules may contradict. Second, the terms employed in legal discourse are "open-tex- tured." That is, the meanings or definitions of many legal terms 4. Sergot & Sadri, The British Nationality Act as a Logic Program, 29 COMMUNICATlONS OF THE ACM 370 ( 1987l. 5. Rissland & Ashley, A Case.Based System for 7}'ade Secrets Law, PROC. OF THE ANN. CONF. OF THEAM. A. FOR ARTIFICIAL INTELLIGENCE 60 (1984); Ashley, ModellingL~galAr- gument: Reasoning with Cases and Hypotheticals (unpublished manuscript). 6. See, e.g,, PROC. OF THE FIRST INT'L CONF. ON ARTIFICIAL ][N~rELLIGENCE AND LAW (ACM 1987}. Gardner herself offers an excellent bibliography (pp. 201-14), and presents in Chapter 4 a survey of the field complete through the spring of 1986.Spring, 1988] Artificial Intelligence 225 and predicates are inherently indeterminate in the philosophical sense of natural kind classes. Third, legal questions commonly invite more than one answer, yet legal argument is both time constrained and resource limited. For while conflict, disagreement, and argument are part and par- cel of the law, 7 the law must provide a timely answer for one side or the other, reached at a socially acceptable cost. Finally, in law, the answers change. Whether the change is in- cremental, as in the mode of Kuhn's normal science, or abrupt, as in a Kuhnian paradigm shift, a hard learning issues lurk close to the surface of the legal issues that AI programs attempt to resolve. To accomodate even gradual change, the algorithms must adapt to a dynamic knowledge base: a shifting foundation of cases, statutes, and the indices, rules, and norms which manipulate them. Eventually one must confront the change in predicates and in the representation itself, either through the emergence of new legal concepts or through the substantial modification of old ones. 9 In summary, legal reasoning requires certain minimum capabilities: 1. the ability to reason with cases and examples, particular- ly through analogy; 2. the ability to handle