Computer-Aided Law? by Philip Greenspun Abstract 1 Anatomy of a lawsuit 1 Three important systems 2 Meldman 2 Gardner 3 Ashley 6 Where do these systems fit? 7 Do these systems work? 7 What is the point of computer-aided law? 8 Answer 1: There is none. 8 Answer 2: We can replace traditional jurisprudence. 9 Answer 3: We can advance AI research. 10 Answer 4: We can help people work with today's jurisprudential system. 10 Which answer is right? 11 Meldman 13 Gardner 14 Ashley 16 We need a better substrate 18 Greenspun v. Smyly in our new substrate 18 What can be done now with such a substrate? 19 What can be done in the future with such a substrate? 20 Conclusion 20 Acknowledgments 20 References 21 Appendix A (Greenspun v. Smyly Complaint) 24 Abstract This paper sets forth the basic anatomy of a lawsuit, summarizes three of the most significant works in computer- aided law of the past two decades, determines where in a typical lawsuit those works would prove useful, evaluates the success of these three systems, and proposes a new substrate for future programs. Anatomy of a lawsuit Consider a typical lawsuit, Greenspun v. Smyly Autos, Malden District Court Civil Action 94-0872 [Complaint attached as Appendix A]. The suit begins with the plaintiff Greenspun telling his story to an attorney. The story goes something like this [] I took my car to Smyly for service. [] They stole my Alpine stereo. [] I sent them a letter demanding payment for the stereo. [] Smyly refused to pay and tried to get me to file a claim with my own insurance company The attorney's job is to figure out what claims Greenspun may have against Smyly based upon the facts he has alleged. The attorney comes up with (1) breach of contract, (2) negligence, (3) fraud, and (4) violation of the Consumer Protection Act (93A). As the case progresses, lawyers from both sides will argue the law, e.g., "what are the elements of breach of contract" and "what kind of duty of car does a car dealer have to a customer whose car they are servicing?", and the facts, e.g., "was the stereo stolen by a Smyly employee, by someone who broke into Smyly's parking lot, or was it never in the car at all?" A jury will decide whether to accept the Greenspun's version of the facts or Smyly's and a judge will decide what the law is. Three important systems Meldman Meldman designed a system that matches user-formalized facts to legal claims pre-formalized in the system. Meldman limited his work to the claims of assault and battery and never implemented it. The matching described in his thesis [Meldman 1975] could be performed by a simple unifier [Abelson and Sussman 1985] if not for the fact that matching is allowed for two objects that are at different levels in a kind-of hierarchy. The system operates as follows: [] user types in facts of case [] computer tries to match those facts to a legal claim (e.g., assault or battery) [] computer asks user questions to try to transmogrify facts that don't match anything into facts that match somewhere in the computer's model of legal claims [] computer coughs up a legal conclusion and a justification for that conclusion Meldman represents cases and legal theories in PSL (Preliminary Study Language), which appears to be a subset of first order logic. The syntax of the basic PSL statement is "( )". The legal claim of battery is represented by (element cjm-battery, battery comprises (instance person p), a person, p, and (instance person d), a person, d, and situation:contact, a situation called contact situation:intent) and, a situation called intent (element cjm-battery/contact contact comprises (s-object contact-event:c p) :f, contact to the plaintiff (cause c event:e), as a result of an act (agent e d)) of the defendant ... The facts of a specific case under investigation are represented like this (element hit:h1 The hit h1 (agent hit:h1 person:Joe), by Joe (instrument hit:h1 fist:f1), the instrument was a fist f1 (dest hit:h1 shoulder:s1), the destination of h1 was s1 (s-object person:Fred)) the object hit was Fred (part person:Fred shoulder:s1) shoulder s1 is part of Fred (part person:Joe fist:f1) fist f1 is part of Joe Note that the fact that "hit" is a kind-of "contact-event" should allow the hit in the specific case to match the contact event in the cjm-battery claim element. Gardner Gardner takes a deep cut at the law of offer and acceptance. The legal world of contract is a four-node augmented transition network (ATN) with 10 interesting arcs (page 124). The following example is used throughout the book [Gardner 1987]: [] On July 1 Buyer sent the following telegram to Seller: "Have customers for salt and need carload immediately. Will you supply carload at $2.40 per cwt?" Seller received the telegram the same day. [] On July 12 Seller sent Buyer the following telegram, which Buyer received the same day: "Accept your offer carload of salt, immediate shipment, terms cash on delivery." [] On July 13 Buyer sent by Air Mail its standard form "Purchase Order" to Seller. On the face of the form Buyer had written that it accepted "Seller's offer of July 12" and had written "One carload" and "$2.40 per cwt." in the appropriate spaces for quantity and price. Among numerous printed provisions on the reverse of the form was the following: "Unless otherwise stated on the face hereof, payment on all purchase orders shall not be due until 30 days following delivery." There was no statement on the face of the form regarding time of payment. [] Later on July 13 another party offered to sell Buyer a carload of salt for $2.30 per cwt. Buyer immediately wired Seller: "Ignore purchase order mailed earlier today; your offer of July 12 rejected." This telegram was received by Seller on the same day (July 13). Seller received Buyer's purchase order in the mail the following day (July 14). Through an agonizing hand process, Gardner transforms English phrases such as "Accept your offer carload of salt" into these MRS rules [Genesereth, et al. 1980] on page 114 ... (sentence S21) (text S21 "Accept your offer carload of salt") (presupposition S21 Prop2) ; S21 presupposes an offer (= Prop2 (prop exist Off2 (and (offer Off2) (ben Off2 Seller) (obj Off2 (prop exist Exch2 (exchange Exch2))) (terms Off2 Prop2a)))) (= Prop2a (prop exist Exch2 Trans21 Salt2 Vol2 (and (exchange Exch2) (event1 Exch2 Trans21) (transfer Trans21) (obj Trans21 Salt2) (salt Salt2) (quantity Salt2 Vol2) (carloads Vol2) (number Vol2 1)))) (prop-content S21 Prop21) ;;; The propositional content of S21 is that Seller accepts ;;; the presupposed offer. (= Prop21 (prop exist Acc21 (and (acceptance Acc21) (agent Acc21 Seller) (obj Acc21 Off2)))) (literal-force S21 Dec21) (declaration Dec21) Each arc in the ATN has one or more rules associated with it. If the assertions in the facts database so far satisfy the antecedents of a rule according to MRS, then the arc is traversed. For example, a transition from "no legal relations" to "offer pending" is possible when nine antecedents are satisfied (page 142). Every node in the ATN is potentially terminal, i.e., after being fed all the facts, Gardner's system could say "there is a contract and a proposal to modify pending" or "there is an offer pending but no contract." The output of the program, however, is not an answer, but an analysis. Gardner's program identifies the "hard questions" and produces a tree of possibilities (page 176) where branch points indicate different assumptions about what legal conclusions may be drawn from a given set of facts. One of the most interesting aspects of the program is how hard and easy questions are distinguished. Section 7.1.l walks through the program's consideration of whether "Have customers for salt and need carload immediately. Will you supply carload at $2.40 per cwt?" is an offer. There are eight terminal nodes in the decision tree (don't confuse this with the ATN), one resulting in traversing the arc to the "offer pending" node and the others resulting in staying put. A trivial "stay put" node is the result of binding the act for the offer to the speech act of asserting that Buyer has customers for salt. The first nontrivial stay put node occurs when the program assumes that the terms of the offer are not reasonably certain. The second occurs when it is assumed that the terms are reasonably certain, but that Buyer was not apparently ready to be bound. These two have nearly all the same variable bindings as the decision tree node that leads to an ATN arc traversal and hence are picked out of the crowd as worthy of future consideration. Because these lead to the same ATN node, they are condensed together, thus reducing the branches at the top level from 8 to 2. Reduction in the degree of branching at each ATN node is critical. This is not because searching a large space would exhaust a modern computer, but because the complete decision tree has to be presented at the end to a human user. If there are thousands of terminal nodes, most of which could only be reached via absurd bindings, the program isn't going to be of much use. Meldman doesn't really address the issue of whether his system's output would be 100% reliable, but Gardner explicitly recognizes that even the smartest computer will be limited to spitting out a list of hard questions that have to be answered by a jury or judge. Two contexts result from the analysis of "Have customers for salt....". One context is in the ATN node of "offer pending" and comes with a page of MRS conclusions. The other is in the initial ATN node and its only conclusions are that either Buyer wasn't apparently ready to be bound or that the terms weren't reasonably certain. This process continues until there are nine contexts at the end of processing all the facts from the "salt shortage" example. Gardner's principal contribution seems to be dividing up legal reasoning into rule following and reasoning from examples. She's right in that it is pointless to reason by example to show that an acceptance leads from "offer pending" to "contract exists." Nobody is going to dispute that rule. Yet for deciding whether a telegraphed rejection supersedes an airmailed acceptance letter, it would be best to rely on precedent. Ashley Ashley built a Lisp program that can take a new fact situation, compare it to precedent, and construct arguments with those precedents that are recognized as reasonable by attorneys. His program [Ashley 1990] does this without attempting to represent the specific facts of any case or depending on deep knowledge of any laws. HYPO's ability rests fundamentally upon Ashley's manual determination that 13 dimensions are important in trade secret disputes and the formalization of 30 precedents so that each can be fixed in all 13 dimensions. For each dimension, Ashley figures out a way of comparing two magnitudes and which direction is pro-plaintiff. Ashley comes up with a formal description frame for each case and procedures for turning that into magnitudes along the 13 dimensions. This isn't a big step. For example, the change of employment object in the case description has a slot "records, devices, code brought by employee." If that slot is filled, the case's rating along dimension "Common- Employee-Transferred-Product-Tools" will be "something" (as opposed to the alternative) "nothing." The dimension says that the pro-plaintiff dimension is "something" so that a company suing a former employee for stealing trade secrets will be in a better position if this slot is filled. Here are the steps HYPO takes when presented with a new case: [] analyze case with respect to the 13 dimensions [] construct a partial order of all 30 precedents according to how many dimensions are shared with the case under analysis [] select the most on-point precedent for each side (shares the most dimensions), the most on-point citable precedent for each side (decided in side's favor and shares at least one dimension that favors the winning side), and the most on-point near-miss precedents [] generate 3-Ply arguments for each side, where each argument is of the form [] Ply1: "My side should win because we share Dimension P with Precedent Y that was decided in favor of my side." [] Ply2: "My learned opponent has overlooked the fact that Precedent Y was much stronger for their side than the current case [distinguishing precedent from current case]. In Precedent Z, our side won even though it also shares Dimension P [citing counterexamples]." [] Ply3: "Precedent Z shares Dimension P, but it also was very strong for your side along Dimension Q, unlike the present case." [] perturb the facts of the current dispute until new dimensions apply, old dimensions no longer apply, or the position along a dimension changes enough so that the old 3-Ply arguments look different. Offer the user analyses of these hypothetical disputes HYPO generates some remarkably interesting behavior with simple techniques. In evaluating HYPO's performance, Ashley feeds in actual cases and compares HYPO's 3-Ply arguments with those made by the litigants and the court. He notes with pride that HYPO cited many of the same cases and made many of the same arguments as the humans involved. Where do these systems fit? Given the task of deploying Meldman, Gardner, and Ashley's systems in the anatomy of a lawsuit, the most logical division of labor is the following: [] Meldman's program writes the Complaint [] Ashley's program argues before the Court [] an improved version of Gardner's program decides who is right If all three systems worked perfectly, we could automate the entire legal system. Do these systems work? Before one can ask how well these systems work, one needs to determine the purpose of computer-aided law. What is the point of computer-aided law? Answer 1: There is none. Law is English. Computers do not understand English. Game Over. The life of the law has not been logic: it has been experience - Oliver Wendell Holmes 1881 Consider a typical U.S. Supreme Court opinion, Alexander v. United States, No. 91-1526, handed down in the summer of 1993. The lengthy writings of the justices are preceded by a syllabus summarizing the facts of the case, its history in lower courts, and the holding of the Supreme Court. The syllabus explains that Alexander was a bookstore owner found to possess seven obscene items. "In addition to imposing a prison term and fine, the District Court ordered petitioner, as punishment for the RICO violations, to forfeit his businesses and almost $9 million acquired through racketeering activity." Alexander complained that RICO forfeiture constituted a prior restraint on speech and that the forfeiture violated the Eight Amendment ("Excessive bail shall not be required, nor excessive fines imposed, nor cruel and unusual punishment inflicted."). The syllabus summarizes the Supreme Court's holding in the case, i.e., what part of the opinion is binding on lower courts in the future: "1. RICO's forfeiture provisions, as applied here, did not violate the First Amendment.... 2. The case is remanded for the Court of Appeals to consider petitioner's claim that the forfeiture, considered atop his prison term and fine, is 'excessive' within the meaning of the Excessive Fines clause of the Eight Amendment." It sounds like law, but is it? No. The law is what the majority wrote about the this particular dispute. The words of the Reporter of Decisions have no value except in helping legal researchers quickly determine whether Alexander is worth examining more carefully. The syllabus reads much more clearly than the opinion in many cases, so why can't it be used as law? Because we don't trust the very brightest humans, with the most education and experience, to rewrite laws in English, the most flexible representation language. The nation's best legal scholars spend ten years sifting through cases and come up with something like Restatement of Torts. Murky specific cases turn into crystal clear general rules: One who discloses or uses another's trade secret, without a privilege to do so, is liable to the other, if (a) he discovered the secret by improper means, or ... (d) he learned the secret with notice of the facts that it was a secret and that its disclosure was made to him by mistake. Every lawyer has read it, every law library has a copy. Is this magnum opus law? Not unless a legislature has adopted it as a statute. It can be cited in a legal brief, but no judge is bound to follow this rule. Why not? Because we don't trust the very brightest humans, with the most education and experience, to rewrite laws in English, the most flexible representation language. If a brilliant human cannot rewrite the law in English, how can we expect a computer, for whom the law must be rewritten in an impoverished formal language, to draw any useful legal conclusions? Although the preceding ideas are my own, the idea that AI and law don't mix is an old one [Leith 1986]. Even without throwing machines into the debate, nobody has been able to agree on how the legal system works. Legal positivists [Hart 1961] argue that law is essentially logical when cases fall within the core of rules, legal realists respond that judges decide based on personal bias then find a legal rule to justify that decision, and feminists from the Critical Legal Studies school see law as part of the "liberal state coercively and authoritatively [constituting] the social order in the interest of men as a gender... It achieves this through embodying and ensuring male control over women's sexuality at every level." [MacKinnon 1989] Answer 2: We can replace traditional jurisprudence. Traditional jurisprudence bears the same relation to a modern science of jurimetrics as astrology does to astronomy, alchemy to chemistry, or phrenology to psychology. [Loevinger 1949] Litigation is too expensive for individuals, corporations, and society. Legislators are hucksters in the pay of special interests. Judges and juries decide cases according to whim, politics, or prejudice. Let's tear down the whole system and replace it with something better, such as Pound's mechanical jurisprudence [Pound 1908], Loevinger's jurimetrics, Hohfield's analytical jurisprudence [Hohfield 1919], symbolic logic and propositional calculus [Allen 1957], or deontic logic [von Wright 1963]. Once the sloppy vagaries of the current system are replaced by a formal system, the computer will be able to dispense mechanical justice [Kafka 1914]. That glorious day can be hastened by developing programs to sift through precedent and demonstrate that cases are not decided according to fair rules, but rather according to the each party's race, sex, or economic position [Smith and Deedman 1987]. Mathematics and natural science have made substantial progress since the Enlightenment, why not law? This answer is certainly in the best tradition of AI. What could be more familiar than the sight of graduate students streaming out of 545 Technology Square and into a domain where thousands of experts have labored for hundreds of years? One can almost hear them speak... "Everything you are doing is pinheaded. We've studied this problem for three months and come up with The Right Thing." Answer 3: We can advance AI research. AI is all about rules. The Law is the largest body of well- understood rules and meta-rules extant. Legal reasoning is structured, yet flexible enough to deal with virtually every human situation, and produces justified decisions for every dispute. Answer 4: We can help people work with today's jurisprudential system. By exploring legal issues with the help of an expert system that can justify its conclusions, a proto-lawyer can learn about a new field of the law [Ashley and Aleven 1991] or general argument skills [Aleven and Ashley 1993]. For working lawyers, computer systems can help draft and assemble documents [Lauritsen 1993]; even judges can be assisted with writing opinions in certain kinds of routine cases [Branting 1993], an important consideration now that number of cases filed in U.S. state courts is crossing the 100 million/year mark [Rottman and Ostrom 1991]. Should ambiguities in a computer-drafted document lead to litigation, the same computer system stands ready to render assistance. Meldman's system will help us write pleadings. Current text retrieval systems for precedent and statutes are in a fairly sorry state [Bing 1986]. However, based on a fact summary, HYPO will cough up not only the most on- point cases for each side and the best cases for each side to cite, but also 3-ply arguments that can serve as the basis for a legal memorandum [Ashley 1990]. We can figure out which memoranda we are currently able to file by formalizing a court's civil procedure [Meldman 1978]. If one's memoranda are not immediately persuasive, various commercial [Wilmer et al. 1989] and research [Greenleaf et al. 1991] litigation support systems are available. Legislatures could reduce the amount of litigation by reducing the ambiguity inherent in laws drafted purely in English. If a legislature could agree on a formal representation of a law, either in a textual or graphical language, then that formalism would be easy to interpret by potential litigants. Which answer is right? It depends on whom you ask and when you ask. Imagine telling Loevinger in 1949 that "Forty-five years from now, a typical attorney will have a computer on his desk capable of executing 30 million instructions per second with instant access to the full text of every statute and precedent in the United States. This computer will be used primarily for word processing. Sometimes it will be used to retrieve relevant precedent, but if asked about cases where dogs bite children, it will miss those where the judge referred to a German Shepherd specifically. Furthermore, if a C programmer at Microsoft has a bad hair day, the entire system will periodically crash and destroy an hour's worth of the attorney's work." Loevinger probably would have had you committed for psychiatric evaluation because of your extreme pessimism. In 1975, when Meldman handed in his dissertation, optimism about computers was still sufficiently high that he felt safe in assuming that "computer-aided legal analysis" would be seen by readers as inherently useful. A careful reading of Meldman's journal article and technical report reveals virtually no clue to his private thoughts on the subject, except that he presumed the user of his system to be a lawyer. By 1987, AI had lost enough credibility that Gardner, in her Chapter 8, explicitly claims all of the last three answers are right. "If a legal reasoning program can be made to behave appropriately in the face of these complications [referring mostly to open texture], there should be implications for knowledge representation and natural- language processing in general. [Answer 3] ... AI work on legal reasoning may be important for the law as well as for AI. Three levels can be distinguished. First is the level of legal philosophy. [Answer 2] ... The second level is the development and criticism of the substantive law. ... Some rules might be formulated rather differently if an AI program had tested them on a stored set of problems. [Answer 4]." Ashley, in 1990, surveys an AI landscape but slightly changed from Gardner's time and comes up with the same answers as Gardner. Ashley is a little more specific, however. In his introduction, he notes that his model "addresses a central problem in AI: controlling the complexity of inference required for a system to locate relevant information. In the short run, AI's principal contribution to society may be to provide intelligent access to our vast databases of information, in particular, helping us to select and organize information that is relevant" [Answer 3]. His program "demonstrates how to build a program for assisting attorneys in an important part of legal practice" [Answer 4]. In his conclusion, Ashley makes a weak claim that his system contributes to legal philosophy [Answer 2]. Greenspun, in 1994, draws the following conclusions: [] Answer 1 ("there is no point to computer-aided law") is untenable in light of the success achieved by Ashley's one-man one-program virtually know-nothing effort. [] Answer 2 ("computer-aided law can replace traditional jurisprudence") is untenable. Computers that cannot process words without crashing are not going to contribute anything fundamental to jurisprudence in the foreseeable future. [] Answer 3 ("computer-aided law can advance AI research") is only true in the trivial sense that law is AI-complete. If we wrote a program that could do law, we could pull out the common sense reasoning and natural language understanding modules and solve the rest of AI. However, it is also true that if we obtained these modules elsewhere, we could send Samantha (Greenspun's PowerBook) to Harvard Law School. The formalisms and meta-rules of law are not nearly as powerful as those of mathematicians; we could advance AI research just as much by working on theorem proving and restaurant ordering. [] Answer 4 ("help people work with today's jurisprudential system") remains and I'm going to assume it is correct and hope that the reader is convinced by the end of this paper. After accepting that the purpose of computer-aided is to assist traditional practitioners, let us re-examine the three systems. Meldman It is difficult to say exactly where Meldman's approach breaks down because he did not implement his system. Nor did he provide a complete hand simulation of an analysis of a fact description. Meldman expected an implemented system to suffer primarily from efficiency problems. Computers in his day were thought to be smart but weak, groaning under the load of 20 users editing: "if our Corpus Juris Mechanicum contained representations for all of the doctrines in the general law encyclopedia Corpus Juris Secondum, we would need a system ten thousand times larger than the prototype." A modern anthropomorphism of a computer is more like Lenny in Of Mice and Men: powerful but incredibly stupid. In the light of 100 MIPS on a desktop and the linear time Patterson- Wegman unification algorithm, Meldman's concern for efficiency seems absurd. If we could adequately represent 10,000 laws with rules and patterns, it would be a very happy day for computer-aided law. Unfortunately, it is far from clear that Meldman's language is adequate for modeling 10,000 laws. Even if it were, what about the commonplace occurrence of lawyers arguing about whether summaries in C.J.S. are correct? What about open texture? Consider using Meldman's program to help Greenspun file his Complaint against Smyly. Greenspun has to show that the facts alleged support one or more claims for legal liability of the defendant. It shouldn't be too hard to formally represent the change of state in the car (radio->no radio). Deontic logic will serve to represent the obligation of a bailee to return an object left in his care in good condition. However, even if Meldman's program had enough common sense reasoning ability to connect the missing radio with a reduction in value of the vehicle, it might miss the liability of Smyly to Greenspun altogether if the facts were characterized as Greenspun contracting with Smyly to fix certain problems with the vehicle. Unless the user explicitly stated that the vehicle was left in their care, the bailor/bailee relationship might never be uncovered and the most important claims never asserted. Gardner Questions immediately raised by Gardner's thesis include the following: [] Does Gardner have branching under control? [] Given how painful it is to represent natural language in MRS assertions, can this system ever be remotely useful? If not, then won't CS and AI be so advanced by the time natural language understanding is solved that all these ideas will appear laughably simple? [] Could we implement this more cleanly using standard procedural techniques, saving contexts with continuations? Gardner only works through a handful of problems. She knew exactly how her program worked so that when the system started branching like crazy, she could add in a few more MRS rules or assertions. If a system like this were ever unleashed upon a user with a fresh case, why wouldn't half the questions strike the program as hard and result in exponential branching? Even a cursory reading of Gardner's book reveals that it might be less trouble to litigate a contracts case right through the appeals court than to recast the facts of a dispute in MRS assertions. Gardner addresses this in the conclusion: "On the input side, manual encoding of the facts of cases is not convenient. ... Automatic encoding appears to require abilities not found in current natural- language programs. ... These extensions to the program are worth pursuing." As noted previously, a computer that can understand natural language won't have to read Gardner's thesis in order to dispense legal advice, it can just hook itself up to Lexis and have an FTP party. Logic programming and rule-based systems don't enjoy the same vogue they did in the 1980's. People built hundreds of toy systems that somehow never scaled the way procedural systems did. In that light, we could look at implementing the bulk of Gardner's system in a procedural language such as Lisp. The most suitable Lisp dialect for this application appears to be Scheme [Clinger and Rees 1991] because of its ability to save its state as a continuation (potentially more efficient is a nondeterministic extension of Common Lisp [Siskind and McAllester 1993]). For example, whenever a program came to a hard question, it would save a continuation just before taking one path. After a terminal context was reached, it would be pushed onto a list and the continuation invoked to restart computation at the hard question but with a different answer. This might result in a clearer more extensible implementation and could piggyback on efficiently implemented backtracking mechanisms. Having addressed these theoretical questions, let us turn out attention to the practical world. Gardner's program would be of little help in preparing Greenspun's Complaint against Smyly. It is obvious that Smyly and Greenspun had a contract and the real question is whether or not Smyly had deeper obligations under contract or common law. Before we consider whether a Gardner-type system would be useful in the trenches of litigation, we have to remember that we'd need four Gardner-type systems, each one with a deep knowledge of a different legal claim: breach of contract, negligence in bailor/bailee situations, fraud, and 93A. Many of the techniques in Gardner's contracts program would be inapplicable in, for example, the bailor/bailee area. We'd effectively be starting from scratch in any area where events along a time line didn't map to a path through an ATN. In Greenspun v. Smyly, there aren't enough facts under dispute to make a "path managing" program like Gardner's appear useful. Does trying to convince a consumer that his car insurance covers something it doesn't constitute fraud or not? Even if we were able to use Gardner's contracts program, we'd want a more practical system combining Gardner's formalisms and an interactive interface that exploits the strengths of humans. Humans have two hands and are good at "on the one hand but then on the other hand" arguments. They aren't good at maintaining a list of 15 paths through a finite-state machine or augmented transition network. Machines aren't good at natural language understanding or common sense reasoning yet are excellent at maintaining lists of paths through a finite-state machine. Let's drop into the middle of a user's interaction with such a system: [] program asks "the first telegram was an offer, the second telegram (carload of salt, immediate shipment, COD) was a counteroffer. Is the Air Mail letter a NOP, notification of death, acceptance, rejection, counteroffer, or acceptance plus proposal to modify? [] users clicks both the "counteroffer" and "acceptance plus proposal to modify" boxes. [] program asks "we're assuming now that the Air Mail letter was a counteroffer. What was the third person's offer (salt at $2.30 per cwt)? [] user clicks "NOP" [] program asks "what was the last telegram (ignore letter)?" [] user clicks "revocation of counteroffer" [] program says "out of events. This series of assumptions terminates in No K. We'll backtrack to the point where we assumed the Air Mail letter was a counteroffer. Now the first telegram was an offer, the second telegram was a counteroffer, the Air Mail letter was an acceptance plus proposal to modify. What was the third person's offer?" [] user clicks "NOP" [] program asks "What was the last telegram?" [] user clicks "Not sure." [] program prints out rules relevant to the "acceptance plus proposal to modify" state and then offers to help find caselaw. ... Ashley Ashley makes a very compelling point at the beginning of his book: "a good ethics tutor knows how to pose a series of hypotheticals that make a clear moral judgment seem more and more dubious. Why would we expect less from expert systems?" Ashley's thesis is also notable for its determinedly know- nothing attitude. We can't represent the specific facts in the case. We can't say that the "Agreement-Supported-By- Consideration" dimension is more or less important than the "Common-Employee-Sole-Developer" dimension. Ashley's HYPO system is unambitious and pessimistic about the possibilities for machine reasoning, yet for me it produces the most striking results of the three systems considered here. How would HYPO have done if its database included 70,000 decisions of the Massachusetts Supreme Judicial court? Did HYPO do well because Ashley enshrined in its database precisely those cases that reference each other in a little self-contained area of the law? Aren't we kidding ourselves in thinking a program that keeps itself in ignorance of the actual facts of a case can sort through the thousands of precedents that are out there? If HYPO pulls too many cases out of the database, the nodes in the partial order lattice will each have contains dozens of cases. HYPO will have to either pick cases at random to cite or generate so many 3- Ply arguments that a brief-writing attorney would be boggled by choice. HYPO works well in its chosen corner of the law. However, a case that fits neatly into trade secret can probably be analyzed by a decent intellectual property lawyer faster than the facts of a case could be formalized for HYPO (in fairness to Ashley it must be noted here that HYPO is much less demanding of the user than Gardner's system). Ashley is mute on the subject of finding alternative claims from the same facts, which is what one would expect a creative lawyer to do ("if we can't get him on trade secrets, we'll go after him for common law fraud or breach of fiduciary duty"). This problem comes to the fore when we consider using Ashley's system to help Greenspun construct his Complaint against Smyly. How do we categorize this case in advance? That is precisely the hard part of pleading. Is it contract or tort? Consumer protection or fraud? What leads to the greatest recovery? Imagine that a large case base of consumer versus car dealer precedents had been assembled. Most of these would involve breach of contract or warranty claims. Would it ever occur to any of the "knowledge engineers" looking at these precedents to include "part-of- car-stolen" and "theft-by-dealer-employee" as dimensions? If not, it is hard to see how genuinely on-point cases would be uncovered. HYPO might help Smyly add a Counterclaim to its Answer. After all, it doesn't really matter why Smyly is being maliciously and frivolously sued, it only matters what are the liabilities created by the filing of a malicious and frivolous suit. Discovery fights also have similarities that transcend the facts to some extent and HYPO might help an inexperienced lawyer to quickly find leading cases. Imagine now that HYPO is invoked to assist in the filing or defending of a summary judgment motion on the fraud count. We supposed earlier that HYPO had a special car dealer/consumer precedent collection, nicely dimensionally analyzed. Unfortunately, relevant precedent for the inducement to insurance fraud claim here is likely to come from cases with all kinds of underlying factual disputes. If those cases are not analyzed along the same dimensions, then HYPO will have no way of constructing its partial order lattice. Although through Greenspun HYPO's case bases are likely to be either too broad, too narrow, or too fragmented, Ashley's methodology has a lot of promise. He doesn't expect more from computers than they are currently capable of delivering. The effort that would be required to formalize precedent for his system is not out of line with the effort that publishers such as West already go to when digesting or indexing cases. It would be very interesting to see what kinds of practical tools could be built by hooking up HYPO to a large database of precedent and the case indexing skills of a big publisher. We need a better substrate Interesting computer-aided law programs are not going to be built upon a substrate of English text, e.g. Lexis. We need to start with something that [] does not depend on natural language understanding, common sense reasoning, or laborious formalization, [] can represent at least three views of a dispute simultaneously (plaintiff, defendant, judge), [] justifies its existence by reducing the administrative burdens of a lawsuit, saving time for both parties and the judge, and [] incrementally builds a substantial database of formalized legal argument. It is upon such a substrate that machine learning programs will learn the law, probabilistic reasoning systems will make reasonable predictions, and new ideas for computer- aided law will grow. Greenspun v. Smyly in our new substrate Figure 1 shows a graphical representation of Greenspun's 93A claim against Smyly. The box on the right labeled "93A violation" indicates that five elements must be present for a 93A claim: (1) plaintiff is a consumer, (2) defendant is a business, (3) defendant committed an unfair or deceptive act in trade or commerce, (4) plaintiff gave business 30 day notice (which itself must have 4 sub-elements, one of which may be satisfied by mentioning one of six items in the demand letter), and (5) business did not make a reasonable settlement offer in the 30 days following its receipt of plaintiff's demand letter. Note that the claim box points back to a statute (Massachusetts General Laws, chapter 93A) for authority. In other situations, this might be a case or set of cases. Each blob on the left is one of the facts asserted by the plaintiff. The one labeled "Greenspun is a consumer" has an arrow pointing directly to the element "plaintiff is a consumer." This arrow corresponds to an assertion by the plaintiff that the element is satisfied by the fact. The element "unfair act" is satisfied by either the fact "Smyly stole stereo" OR "demand sent" AND "Smyly stonewalled" (the AND and OR symbols are standard ones from digital logic design). Specific paragraphs in the demand letter fact are asserted to match up to requires of the "notice" 93A element. Figure 2 shows the same situation as Figure 1, but it includes some of Greenspun's private litigation strategy. It shows the fact "Smyly stole stereo" supported by four pieces of evidence: (1) a receipt from Tweeter, which installed the original Alpine, (2) a receipt from Rich's Car Tunes, which installed the replacement Alpine, (3) testimony from Neil Mayle, a passenger in the car shortly before the theft, and (4) Greenspun's own testimony. Until required to disclose this information by discovery, a motion fight, or the trial, these links will remain hidden from anyone but Greenspun. Figure 3 shows a hypothetical answer by Smyly. Items in red are assertions by Smyly or consequences of those assertions. Smyly admits that Greenspun is a consumer and that Smyly is a business. Smyly denies that paragraph 2 in Greenspun's demand letter adequate describes his injury suffered; the computer can infer that therefore the notice element of 93A is under dispute. Smyly denies that they stonewalled and asserts a defendant fact that "Smyly offered to pay." Smyly denies that they stole the stereo and asserts "stereo never in car". Figure 4 shows Smyly's "hidden evidence", i.e., whose testimony they expect to use to establish their facts at trial. What can be done now with such a substrate? Before Smyly files a motion for summary judgment with Judge Foobar, it can ask "how many summary judgment motions on 93A claims were brought by defendants before Judge Foobar? How many were granted?" If Smyly files the motion, the judge can ask to see the whole case assuming all of Greenspun's facts to be true, exposing only the disputed characterizations of the law. Anyone can ask "what happens if the finder of fact doesn't believe this one... how many claims do we lose?" With some relationships among the facts, a belief revision can answer the question "Now that the jury isn't going to buy Fact X, what set of consistent beliefs leads to the minimum liability for my client?" What can be done in the future with such a substrate? Before you think about this question, write down what you think are the most important new applications of computers developed since 1975. Do not count, for example, CAD, computer algebra, or word processing because those were already well-established among users of powerful computers in 1975; the fact that more people use these applications today is evidence of the prowess of hardware engineers not innovation in computer science. Look at your list and ask, Were these programs developed before or after the substrate that made each possible? My list is short: the spreadsheet and the World Wide Web. Most computer science researchers in the late 1960's expected that computers would one day be cheap enough to put on everyone's desk. Did anyone of them invent the spreadsheet? The spreadsheet was not invented until after computers were in fact cheap enough to put on everyone's desk. In the mid-1970's, thousands of people were using ARPAnet, which grew into Internet. Many of these people knew that powerful graphics workstations would be one everyone's desk eventually. Did any of them conceive the World Wide Web? It was not until powerful graphics workstations were in fact on everyone's desk that the Web was conceived. There is no reason to believe that we can envision the applications until we have built the substrate. Conclusion Efforts to make computers think like human lawyers are no more or less likely to succeed than efforts to make computers think like other kinds of humans. If we want to build anything of practical value, we should concentrate on approaches that exploit the comparative strengths of computers and humans. In particular, building a substrate capable of formally representing the state of a lawsuit will provide short run benefits to litigants, long run benefits to society, and will result in currently unimaginable computer-aided law applications. Acknowledgments I am grateful to David McAllester for noting that "AI is just an efficiency hack" and answering countless questions. Marc Lauritsen was kind enough to steer me through the unfamiliar territory of AI and law literature. The entire High Testosterone Floor (4AI) turned out to give me the shellacking I deserved with an early version of a talk based on this paper; James O'Toole and Franklyn Turbak stayed to sew together my tattered ideas and garments. References Abelson, H. and Sussman, G. 1985. Structure and Interpretation of Computer Programs. MIT Press, Cambridge, 1985 Aleven, V. and Ashley, K.D. 1993. "What Law Students Need to Know to WIN," in Proceedings of the Fourth International Conference on Artificial Intelligence and Law (New York: ACM Press), pp. 152-161 Allen, Layman 1957. "Symbolic Logic: A Razor Edge Tool for Drafting and Interpreting Legal Documents." 66 Yale Law Journal 833 Ashley, Kevin D. 1990. Modeling Legal Argument: Reasoning with Cases and Hypotheticals (Cambridge: MIT Press) Ashley, K.D. and Aleven, V. 1991. "Towards an Intelligent Tutoring System for Teaching Law Students to Argue with Cases," in Proceedings of the Third International Conference on Artificial Intelligence and Law (Oxford: ACM Press), pp. 42-52 Bing, J. 1986. "Legal Text Retrieval Systems: The Unsatisfactory State of the Art," Journal of Law and Information Science, 2(1) Branting, L. Karl 1993. "An Issue-Oriented Approach to Judicial Document Assembly," in Proceedings of the Fourth International Conference on Artificial Intelligence and Law (New York: ACM Press), pp. 228-235 Clinger, William and Rees, Jonathan (editors) 1991. "Revised^4 Report on the Algorithmic Language Scheme." Lisp Pointers 4(3):1-56 (July-September) Doyle, Jon 1992. "Reason Maintenance and Belief Revision," in Belief Revision (Peter GŠrdenfors, ed), Cambridge University Press Gardner, Anne von der Leith 1987. An Artificial Intelligence Approach to Legal Reasoning. MIT Press, Cambridge. Genesereth, M.R., Greiner, R. and Smith, D.E. 1980. MRS Manual. Memo HPP-80-24. Stanford Heuristic Programming Project, Stanford University Greenleaf, G., Mowbray, A. and Tyree, A. 1991. "The DataLex Legal Workstation-Integrating Tools for Lawyers," in Proceedings of the Third International Conference on Artificial Intelligence and Law (Oxford: ACM Press), pp. 215-224 Hart, H.L.A. 1961. The Concept of Law. (Oxford: The Clarendon Press) Hohfield, W. 1919. Fundamental Legal Conceptions as Applied in Judicial Reasoning. Kafka, Franz 1914. In der Strafkolonie (In the Penal Colony). Available in translation by J.A. Underwood in Franz Kafka Stories 1904-1924, Futura, 1983. Lauritsen, Marc 1993. "Knowing Documents," in Proceedings of the Fourth International Conference on Artificial Intelligence and Law (New York: ACM Press), pp. 184-191 Leith, P. "Legal Expert Systems: Misunderstanding the Legal Prodess in Computers and Law," in Computers and Law, 49 (September 1986) Loevinger, Lee 1949. "Jurimetrics: The Next Step Forward" 33 Minnesota Law Review 455 MacKinnon, Catherine. "Feminism, Marxism, Method and the State: Toward Feminist Jurisprudence," in Critical Legal Studies (ed. Hutchinson) (Totowa, New Jersey: Rowman & Littlefield) Meldman, Jeffery A. 1975. A Preliminary Study in Computer- Aided Legal Analysis. MIT Project MAC Technical Report 157 (Condensed version available as "A Structural Model for Computer-Aided Legal Analysis" in Vol. 6 of the Journal of Computers and Law (1977)). Meldman, Jeffrey A. 1978. "A Petri-Net Representation of Civil Procedure," IDEA 19(2) Pound, R. 1908. "Mechanical Jurisprudence," 8 Columbia Law Review 605-623 Rottman, David and Ostrom, Brian 1991. "Caseloads in the State Courts," State Court Journal 15(2) (Spring) Siskind, Jeffrey M. and McAllester, David A. 1993. "Nondeterministic Lisp as a Substrate for Constraint Logic Programming," in Proceedings of the Eleventh National Conference on Artificial Intelligence (Washington, DC) Smith, J.C. and Deedman, C. 1987. "The Application of Expert Systems Technology to Case-Based Law," Proceedings of the First International Conference on Artificial Intelligence and Law (New York: ACM Press), pp 84-93 Wilerm, Cutler and Pickering 1989. Manual on Litigation Support Databases. (2nd edition) (New York: Wiley Law Publications) von Wright, G. 1963. Norm and Action. Appendix A (Greenspun v. Smyly Complaint) COMMONWEALTH OF MASSACHUSETTS MIDDLESEX, ss. MALDEN DISTRICT COURT __________________________________ ) PHILIP GREENSPUN, ) Plaintiff, ) v. ) COMPLAINT ) SMYLY AUTOS, INC. ) Defendant. ) __________________________________) 1. The plaintiff, Philip Greenspun (hereinafter referred to as "Greenspun") is a consumer with a usual place of residence of 55 Russell Street, Melrose, MA 02176. 2. The defendant, Smyly Autos, Inc. (hereinafter referred to as "Smyly") is a corporation engaged in the business of selling and servicing automobiles. Smyly is based at 700 Broadway, Malden, MA 02148 in Middlesex County. 3. On April 13, 1994, Greenspun delivered his 1993 Dodge Grand Caravan (the "Caravan") to Smyly requesting that they repair a rattle, a hesitating engine, the brakes, a mirror motor, and a door lock. Greenspun also requested an oil change and state inspection. 4. Smyly contracted to perform repairs under Greenspun's new car warranty and undertake the oil change and state inspection at Greenspun's expense. 5. On April 15, 1994, Greenspun picked up the Caravan at Smyly. Greenspun paid a bill of $33.28 for maintenance and $31.50 for his rental car (from Enterprise, operating out of the same business as Smyly). Greenspun found that his car was locked but that the alarm had not been set. Greenspun's Alpine 7525 radio/cassette player (the "Alpine") had been neatly removed from the Caravan dashboard, apparently by someone skilled in the art of disassembling dashboards, in possession of the proper tools, in possession of a key to the vehicle, and with plenty of time. 6. Greenspun requested the return of the Alpine from Smyly personnel, all of whom were evasive and denied that the Alpine had been in the Caravan when Greenspun brought it in for service. Smyly personnel refused to return or replace the Alpine, to apologize for having stolen the Alpine or having allowed it to be stolen, or to assist Greenspun in any way. 7. On April 16, 1994, Greenspun went to the Malden central police station at 5:10 pm and reported the stolen Alpine 7525 radio/cassette player, receiving police report number 9403151. 8. On April 19, 1994, Alec DeSimone telephoned Greenspun, identifying himself as a vice president of Smyly. DeSimone repeated Smyly's allegation that there was no radio in the car when Greenspun brought it in. DeSimone asserted that Smyly's actions were not covered by G.L. c. 93A, that Smyly went to court all the time, and that Smyly always won the numerous lawsuits people filed against it. DeSimone further asserted that he didn't care if Greenspun filed suit because Smyly's insurance company would pay for all litigation expenses and awards of damages. DeSimone refused to compensate Greenspun for the Alpine or his time and inconvenience. 9. On April 19, 1994, Greenspun sent by facsimile and first class mail a letter to Smyly setting forth the facts in the case and demanding compensation under Massachusetts General Laws ch. 93A. Greenspun gave Smyly 30 days in which to make a settlement offer. 10. Smyly made no unconditional offer of settlement at any time since April 19, 1994. Smyly and its insurance company insisted that, despite all evidence to the contrary (assuming the Alpine had been stolen while in Smyly's possession) that the act was not committed by a Smyly employee and that therefore Greenspun should file a claim with his car insurance carrier. Smyly and its insurance company expressly denied that a standard Massachusetts automobile insurance policy did not provide coverage while a vehicle was in the possession of an automobile dealer. 11. Greenspun suffered a loss of $800 for the stereo and installation and $2400 for lost work time. 12. Greenspun has given due notice to Smyly of his claim for $3200 in compensation. 13. The plaintiff is entitled to receive the sum of no less than $3200 from Smyly. COUNT I BREACH OF CONTRACT OF DEFENDANT SMYLY 1. That the defendant Smyly breached said defendant's contract with plaintiff Greenspun. 2. Wherefore, as a consequence of said breach by defendant Smyly the indebtedness of the defendant Smyly to the plaintiff be established by the Court. 3. That Smyly be ordered forthwith to pay the damages as outlined in paragraph 13 above together with interest from April 15, 1994. COUNT II NEGLIGENCE OF DEFENDANT SMYLY 1. That the defendant was negligent in its care of plaintiff's vehicle. 2. Wherefore, as a consequence of said negligence the indebtedness of the defendant to the plaintiff be established by the Court. 3. That Smyly be ordered forthwith to pay the damages as outlined in paragraph 13 above together with interest from April 15, 1994. COUNT III FRAUD OF DEFENDANT SMYLY 1. That the defendant concealed its knowledge of the events surrounded the theft of the Alpine with the intent of inducing the plaintiff to file a false claim with his own insurance carrier. 2. That the defendant knowingly misrepresented the responsibilities of a car insurance provider in the Commonwealth of Massachusetts with the intent of inducing the plaintiff to abandon his attempts at recovery from defendant and its insurance and file a false claim with his own insurance carrier. 3. Wherefore, as a consequence of said fraud the indebtedness of the defendant to the plaintiff be established by the Court. 4. That Smyly be ordered forthwith to pay the damages as outlined in paragraph 13 above together with interest from April 15, 1994. COUNT IV CONSUMER PROTECTION STATUTE 1. The plaintiff herein incorporates all of the allegations contained in paragraphs 1-13 and each of the Counts above. 2. At all relevant times hereto the defendant was engaged in trade or commerce. 3. The acts of the defendant herein were performed willfully and knowingly. 4. The acts of the defendant herein constitute unfair or deceptive acts or practices within the meaning of G.L. c. 93A, section 9. Wherefore the plaintiff requests this Court to enter a judgment for plaintiff against the defendant, award treble damages to plaintiff, award interest from the dates that plaintiff incurred expenses, and award costs and attorneys' fees to the plaintiff. That Court order such further and other relief as seems meet and just to it. THE PLAINTIFF DEMANDS A TRIAL BY JURY AS TO ALL ISSUES SO TRIABLE. Philip G. Greenspun 55 Russell Street Melrose, MA 02176 (617) 662-8735 Dated: May 23, 1994