Journal of Law, Information and Science
|
|
Home
| Databases
| WorldLII
| Search
| Feedback
Journal of Law, Information and Science |
|
by
Daniel Hunter[*], Alan Tyree[*] and John Zeleznikow[*]
In this article the authors continue the Artificial Intelligence and the Law Debate begun with Moles' 1991 article. In it the authors answer the latest criticisms made by Moles and others as they explain and argue the case for the practical benefits to be gained by AI systems involving the law.
___________________________
Over the last year a debate between the authors and Bob Moles, amongst others, has played itself out in the pages of this journal.[1] The debate has focussed on the narrow question of whether the current research into the use of computers in law is justified on a jurisprudential basis. The research of which we speak is the sub-speciality of Artificial Intelligence and Law. If the debate could be seen as litigation then Bob Moles as plaintiff submits that the use of AI in law is fundamentally flawed, because so few of the researchers have considered the jurisprudential background in the law they model. The Defendants, Messrs Tyree, Hunter and Zeleznikow, submit in most of the systems considered, jurisprudential questions are irrelevant to building practical systems, or alternatively that where appropriate researchers have given due regard to jurisprudence. A quick overview of the ‘pleadings’ may be instructive.
The ‘Statement of Claim’ consists of Moles’ first piece,[2] criticising the work done in logic programming at Imperial College of Science and Technology, London, and in particular the published work on the logic programming model of the British Nationality Act.[3] The ‘Defence’ was in two separate articles, one by Zeleznikow and Hunter[4] and the other by Tyree.[5] Both articles in essence argued that the concerns about logic programming and legal expert systems (LESs) are ill-founded for two main reasons. First, it is not necessary for the LES to model legal reasoning—all that is necessary is that the damn thing works. If it can provide a useful tool for lawyers in research and advice then it has served its function. Secondly, the field of AI and Law is far wider than the narrow area examined in Moles’ first article. He criticised logic programming, without discussing the current research in case-based reasoning, neural networks, fuzzy logic, to name but few.[6] This sin of omission gave the wrong impression of the work currently undertaken. A number of other points were raised in response to Moles’ criticisms, and the reader is directed to the articles.
The Plaintiff Moles, along with a secondnamed Plaintiff, Surendra Dayal, filed what may be either a ‘Reply to the Defendants’ Defence’ or perhaps an ‘Amended Statement of Claim,’[7] it matters not since the litigation analogy is far from perfect.[8] This article then constitutes the ‘Answer to the Plaintiffs’ Reply.’ It is necessary to analyse the issues raised by Moles and Dayal in some depth.
The Moles and Dayal piece (‘hereinafter ‘MD’) may be divided into three parts:
1. The question of intelligence
2. The question of rules
3. The question of Semiotics
The first part of MD takes our earlier work to task, mostly querying the definitions and use of terms such as ‘intelligent’, ‘expert systems’, and ‘artificial intelligence’. MD has enormous difficulties with the idea that any of these systems are or will be intelligent. MD says:
‘[T]hey [Zeleznikow and Hunter] are clearly implying that current work in expert systems will lead us to intelligent systems. We would disagree, and suggest that their discussion does not support this claim. Their linking of “expert” with “intelligent” systems in this way is misleading.’[9]
In using terms such as ‘expert system’, ‘artificial intelligence’ and ‘intelligent system’ we were using the standard nomenclature of that part of computer science called ‘Artificial Intelligence’. We assumed, perhaps incorrectly, that the readers and the authors of MD would understand these expressions are shorthand for various concepts. For example, there is no computer system in all of the artificial intelligence community which has anything approaching ‘intelligence’, as we understand it in the human world. The term ‘Artificial Intelligence’ (AI) is itself merely shorthand for the fields of endeavour in computer science which seeks to understand some aspects of the way human beings reason, and hopefully, simulate these aspects. We explained in our earlier piece that we have not come close to understanding human reason nor to modelling it. AI, as a discipline within computer science, has a number of discrete sub-areas all seeking to understand and simulate some feature of human achievement: these include computer vision, speech understanding[10] and reproduction, and expert systems. Researchers then use the expression ‘intelligent system’ as a shorthand for any computer system which uses some or all of the techniques to be found within any of the sub-areas of AI research. None of the researchers yet believe that these systems are intelligent as the layperson understands it; they simply know that the system uses AI techniques and is therefore an ‘intelligent system.’
MD disagree with the idea that expert systems will lead to ‘intelligent systems’. However, by definition, expert systems are part of AI research, and are therefore already intelligent systems. MD continues.
It says that Zeleznikow and Hunter’s categorisation of legal expert systems within the category of expert systems, which in turn are subsumed within intelligent systems, is ‘misleading.’ because the ‘connections are assumed, but not established’[11] A brief glance at any AI introductory text will show that expert systems are part of ‘intelligent systems.’[12] And legal expert systems must be part of expert systems, since they are merely expert systems within the legal domain. MD however says that Zeleznikow and Hunter fail to establish this link, but fails to explain why LESs are not part of the ES discipline. One may just as well say that Criminal Lawyers do not fall within the genus ‘Lawyer,’ but if one were brave enough to make such a bald assertion, then one would surely have to justify it. The authors of MD fail to do so.
MD is then concerned about the comment of ZH that in the earlier article the authors had ‘…not examined any question of abstract thinking, open texture or the philosophical underpinnings of the law.’[13] MD claims that either we do not discuss these issues or that we fail to find them important. We shall take up this point later, but suffice to say at this point that MD’s claims are mischaracterisations of our arguments. Our comment was in relation to the difficulty in producing ESs for laypersons, where the difficulties are so far insuperable. These difficulties included word-sense disambiguation,[14] lack of ‘understanding’ and difficulties in representing world knowledge, as well as those points raised in the above quotation. We were simply making the point that we cannot at this stage build ESs for the general public, and that we must aim our sights lower. MD criticises this as either not ‘intelligent’ or only indexing. Well might we say, ‘So what?’ If we can provide more useful indices or can represent norms so that lawyers can use them, have we not then provided a useful and worthwhile system? More importantly, we cannot say that from these small acorns great oaks will not grow. These are the first steps in building truly intelligent systems, and MD really cannot at this stage say that these indexing systems will never lead to intelligent systems. We just do not know.
MD is most concerned about the use of rules in LESs, and proposes contract law as an example of an area where the authors of MD can show that no rules exist. They use this example because it is a blackletter law area, and one where AI and Law researchers believe that rules exist. How wrong we were. MD shows us that there is a movement away from the old formulation of offer and acceptance. This, MD asserts, shows that there is no rule that there must be offer and acceptance. In essence, the argument is really Moles’ ‘One cannot separate law from fact’ argument,[15] put in slightly different form and rechauffé. We may dispatch this argument with a range of shots:
1. It is a simple matter to recast the rule. Let us take the old rule ‘One cannot use estoppel as a sword but only as a shield.’ MD explains that this is no longer the case. Let us now say that the rule is ‘One can (in certain circumstances laid down by the High Court in Waltons’ Stores) use estoppel as a sword and a shield.’ How is this problematical?
2. Prior to Waltons' Stores, how did lawyers[16] acting for an aggrieved party approach a problem where there was a representation by one party, relied upon by the aggrieved party to his/her detriment? Applying High Trees they asked whether the aggrieved party sought to use promissory estoppel as a cause of action (that is, as a sword). If so, they said that this was not permissible, as there was quite clearly a rule[17] that one could not use promissory estoppel as a sword. The same rule applied at the time when the plaintiff in Waltons' Stores approached the lawyers. However, the lawyers saw that these facts and the current trend in law was away from the rule, and so suggested that it might be worth taking the matter all the way to the High Court. Suppose that those lawyers had in their possession a LES which said that there was no cause of action because one could not use promissory estoppel as a sword. Would the outcome have been any different? We suspect not. As in the actual case, the lawyers would simply have said that the facts were such and the trend in the law was such that the rule expressed in the computer was no longer correct. We seek to build systems for lawyers who then are charged with the responsibility, as they have now, of deciding whether the facts or other circumstances are such as to reject the computer generated advice. As we have said before, ‘The LES does not make value judgments: this is the province of humans.’[18]
3. MD argues that we cannot look at law as ‘rules + exceptions’ as this gives the appearance of consistency which does not exist. This principle of law not being ‘rules + exceptions’, it says, is admirably noted by the new edition of Cheshire & Fifoot on Contract. What this fails to recognize is that what authors can place in their books, knowledge engineers can place in their LESs. MD notes that Cheshire & Fifoot says ‘if the court is sufficiently determined to see relationships in term of offer and acceptance, it can find them anywhere.’ The book, while it does mention this, does give the lawyer and law student some indication of the rules of offer and acceptance, because in the majority of cases this analysis will suffice. Why then can not the perceptive knowledge engineer include this particular piece of advice in the LES? We may insert a rule in our LES that fires whenever the matter is one relating to offer and acceptance, a fact we allow the lawyer to determine simply by asking whether there is any question of an offer. The outcome of that rule firing may be as simple as flashing the following message on the screen:
‘Bob Moles, along with Cheshire & Fifoot, says that if the court is sufficiently determined to see relationships in term of offer and acceptance, it can find them anywhere. Give consideration to whether there is any likelihood of an offer being found here. I’m sorry that I cannot answer this question, but he law is in a state of flux and you are the human being—I’m just a stupid machine’
The authors of MD may claim that this is not intelligent, but this depends on one’s definition of ‘intelligent.’ This is an example of a system which recognizes its limitations and advises the user accordingly. Few lawyers have this capacity, and fewer still would be prepared to admit faults so readily. Even if one decides that this system is not ‘intelligent’, we return to the reason why we are building these systems: if we only seek to help lawyers, just as textbooks do, then a system which provides the same information but in a more user-friendly, accessible and relevant form, might well be called ‘intelligent.’
Furthermore, rather than having such a stupid message flashed on the screen, we might be able to provide even more ‘intelligence.’ We have the technology now to provide ‘intelligent’ information retrieval depending on the existing facts. Cases, comment and reviews are available in response to the user’s query, all of which relate only to the relevant issues. This provides even more useful access to information for lawyers, and well well earn the sobriquet ‘intelligent.’
The argument of MD is moreover predicated on the assumption that rules are limited entities, with no representational capabilities beyond simple IF-THEN rules. We have no problems with defining different rules systems based on non-monotonic logics—logics which allow for answers beyond strictly yes-no. Examples such as fuzzy logics, abductive logic or Polya’s plausible reasoning approach[19] take our discussion beyond the limited way in which MD envisages rules and rule-based reasoners.
4. Even if we were to dismiss the above three arguments, we would still not have dismissed all of AI and Law. MD’s reasoning applies only to rules, and whether we can extract rules from cases. The current research into LESs is not concerned with rule based reasoners; rather the excitement is in the field of case-based reasoners. These expert systems, although only in their research models, seek to apply previously decided cases to the case at bar. These do not apply rules, but attempt to argue from experience by using analogy, and mirror aspects of the way lawyers argue, particularly in the common law jurisdictions.
MD continues to discuss the rule question, in much the same vein, but this time labelled ‘Rules as “reasons”.’ The thesis propounded, and supported by reference to Moles’ other works, is that rules are only a form of shorthand, and do not exist of themselves. This argument does not destroy the validity of even the maligned rule-based reasoners, such as the logic programming commented upon in Moles’ earlier work.[20] Much less does it disturb the current research into case based systems.
Let us look at an example to show how this ‘rules as reasons’ principle works, and why it should cause us to lose little sleep. A lawyer is consulted by her client. The client needs to know whether he is in breach of an environmental pollution statute.[21] The lawyer will consult the law, as expressed in the legislation, text, case, or whatever. From this she may derive a rule, or it may be explicitly defined. In this case let us say the rule is ‘Thou shalt not spew any cyanide gas into the atmosphere.’ The client says that he has indeed spewed the offensive gas into the atmosphere. The first thought the lawyer has is, ‘Well, that’s that. He’s guilty. Game over.’ However, as soon as she mentions this to her client, the client says, ‘But it was an unavoidable accident.’ Aha, she thinks, I do recall that there was a principle that unavoidable accidents were a defence to even strict liability defences. Our diligent theoretical lawyer then has recourse to case, texts and so forth, from which she formulates an argument on her client’s behalf. Moles is correct in saying that the rule simply provides the shorthand for the reasoning. However, in our view this proves nothing, and is anything but fatal to the application of AI to techniques to law.
Three main responses appear.
The first is that the lawyer always has the power and the prerogative to consult the cases and factors which may be important for the case. This is true whether our theoretical lawyer is using her own vague memory of strict liability defences from law school, or a text, or, Heaven forfend, a rule-based legal expert system. The control is not taken from the lawyer’s hands, and on the contrary we would argue much greater control is given to the lawyer. By having access to more rules, better information retrieval and faster access to a range of materials, the LES gives the lawyer increased control, and hence the client receives better advice. This control must, according to MD be unimportant, since when we made this point previously[22] MD dismissed these as purely stupid indexing systems: ‘This [an intelligent information retrieval system] may make it a quick and efficient retrieval or indexing system - but where is the intelligence or expertise?’[23] Once again it appears that the authors of MD have confused the type of systems we seek to build. We are not seeking to build an artificial lawyer, we seek only to build intelligent systems as lawyer aides. Even if MD claims that these systems are not ‘intelligent’, and even if we were to accept the authors’ judgement (which we do not), the systems we build have an immediate use. More importantly, the systems we build now are more likely to lead to improved, truly intelligent machines than beginning from some abstract theoretical position which is subject to criticism. A fotriori, it strikes us as difficult to begin from MD’s theoretical position when it is impossible to verify that this position is correct one.[24] Better then to build the system and see if it works or not, and leads us to ideas to help us in building better systems.
The second answer is dependent not so much on one’s view of the technology as on jurisprudence. In the case above, it is quite clear that the rule ‘Thou shalt not spew any cyanide gas into the atmosphere’ is not the only necessary rule which we must consider. However, it is not at all clear, at least to the authors, why we could not represent the ‘law’ or as MD would have it, the ‘reasoning’ with a series of rules. These rules would have to include the main one, ‘Thou shalt not spew etc’, but might also include a range of exception rules, such a strict liability defences, definitions of atmosphere, and so on. We can define as many rules as we need to to represent what we understand to be the law. This LES would be a representation of an expert’s or several experts’ view of the law, but then so is each piece of advice. The question is really whether the advice is as good as one is likely to get from an expert and whether we can deliver it as value for money. If the choice is between cheap advice from the LES and no advice at all then the LES must have served its purpose[25]
Alternatively, if we accept MD’s assertion that we cannot separate law from the facts of each case, we could have each case represented in the form of rules. For example, we could represent the salient facts of Donoghue v Stephenson [26] in a production rule system as:
IF you drink a bottle of ginger beer
AND IF you are female
AND IF you ingest part of a slug/snail
AND IF you suffer injury from the ingestion
...etc...
THEN you may recover damages.
Now, this form of representation has a number of drawbacks. For one, it does not represent the essential rule (that word again) laid down in that case that one has a duty of care to one’s ‘neighbour.’ We can work around this problem by including such a rule which the lawyer extracts from the case as the ratio decidendi. It also is computationally inelegant and would be extremely slow to process, though it is computationally tractable.[27] It possesses one other disadvantage: it really does not represent law in the way in which lawyers think of it. This, however, is a criticism of the jurisprudential model which MD adopts. If law were thought of this way then only cases on all fours with the case at bar could be used, and as we know lawyers use analogy widely. Researchers in AI and Law recognize this, which is but one reason why case based systems are now coming to the forefront of research.[28] Such systems use hypotheticals and precedents in analogical reasoning about the instant case.
Antithetically, this form of representation does have the advantage that it can represent all the factors in a case, a concern which MD expresses in its discussion of rules as only shorthand for reasoning from factors. Whether representing all factors is necessary, or even correct, is not of concern here. The point is: even simple rule-based reasoners can provide what MD asks for. We can represent the same information in a number of other more useful and speedier ways,[29] but there is no theoretical reason why we could not use this representation.[30]
Finally, we could build into our system a rule which requires the system to ‘argue at all costs’, a rule which MD seems to believe characterises ‘real’ lawyers. Once again we see that if MD can express a concern we can code it into a rule.
There is one final reason why this form of jurisprudence should not unduly concern us. Though MD is at pains to inform us that lawyers use this form of reasoning to arrive at a conclusion, it may not be necessary for a computer to do likewise. As we have previously noted, champion chess players do not reason the way computer chess programs operate. Nonetheless, computers are fast making up the gap in quality of play, as measured in international competitions between grand masters and the best computer chess program, ‘Deep Thought.’ So too, it may be that a similar brute force approach in building LESs may succeed, though the authors of MD will no doubt again claim that it is not very ‘intelligent.’ This may be true, but at least it gets the job done.
The question of consensus
MD believes that using consensus as the basis for building legal expert systems is flawed.[31] After noting the commonplace assertions that the law only deals with disputes and that sometimes we get what we want though it is not condoned in law, MD asks whose knowledge we are modelling? It says:
‘If the expert systems people are modelling “consensus” as the basis for their systems, then we would suggest that they should carry out some real investigation of what it is, and what it is based on. The suggestion that they are modelling the consensual knowledge of experts merely shifts the problem, rather than answer it. Which experts?’[32]
This argument could equally be made for books as for computer systems, a point we have mentioned before. It is furthermore anything but a new argument in ES design, and applies equally to other ESs as with LESs. In building, say, a medical ES to diagnose heart disease we have recourse to the expert knowledge of a cardiac surgeon or preferably a group of cardiac specialists. In the main, the symptoms of heart disease are not contentious. We insert the non-contentious heart attack symptoms into our ES knowledge-base, and no-one raises any objections. However, there is a great deal of contention about a number of other subsidiary symptoms which may indicate heart disease according to some specialists and not according to others. Here we face the problem of consensus which MD posits as fatal to LESs. However, as in the heart disease ES, a LES which models the expert knowledge/opinion of any one or more specialists is sufficient, provided we accept that each specialist, and each ES, may have a different ‘opinion’. These differing opinions will occur at the edges of our knowledge/ES model, which is no different in medicine as in law.[33] But when the patient arrives with pains in the chest and left arm, and his heart is not beating then all of our experts and our ESs will diagnose, ‘heart attack.’ We can perform the same exercise for legal expert systems
In essence what we argue for here is the acceptance of the following two principles:
1. Expert systems cannot, in general terms, be better than the experts which create it. However, in medicine a large number of experts may build a system which diagnoses better than any one of them.[34] Though we have a way to go, it is possible and indeed probable that we will begin to see legal expert systems built by teams of lawyers and knowledge engineers. These expert systems may well be better, more comprehensive, more knowledgeable and more ‘intelligent’ than any one individual legal expert; and
2. If we can cheaply build a LES which does the legal equivalent of diagnosing a heart attack from obvious symptoms, (say for example, a cheap will-maker, a cheaper litigation institution adviser, a quick-and-dirty bankruptcy adviser, etc) then we have performed a useful function. Again, it is a question of value for money.
MD ends its discussion of consensus with the concern that each time a practitioner listens to a case he/she must examine all the facts and factors in determining the best course. It says,
‘In judging whether any particular formulation is suitable, for the case at hand, we are involved in another complex process of selection and formulation. It is this process which is central to the task of lawyering and which is so little understood by some of the people involved in this debate.’[35]
We disagree that we do not understand this proposition: we are extremely mindful of it. Which is precisely why we advocate the development of LESs. Computers are brilliant at information retrieval and other brute force computational problems. In any given period, the computer can recall, process, and discard or accept far more information than a human can. Why not use the computer in this way? MD’s argument that this filtering process precludes computers because there is too much ‘selection and formulation’ is fallacious.
MD goes on to discuss[36] ‘researchers [who]…are much more aware of the environment within which their computers have to operate’[37] unlike the authors who have a ‘certain mind-set’ and worse, two-thirds of whom have committed the sin of being mathematicians![38] In fact, MD is somewhat misleading, for it only really mentions one researcher, Stamper. Here, we are in agreement and believe that Stamper’s work on jurisprudence in AI and Law is important, though his research projects could be re-coded into relatively short PROLOG programs. This perhaps gives an indication that Stamper’s work is nothing special in terms of its jurisprudential approach, and further that this debate is both irrelevant and sterile.
Our only other comment would be that it is important that some researchers look primarily at jurisprudential questions and others at practical implementations. One without the other provides little use and limits the long term viability of the field. We agree with MD’s concluding comments that ‘…one’s practice may be altered by a well-informed theory.’[39] We would like to think that we, along with MD and a range of others, have given some regard to theory in line with our special aims. The body of this article, we hope, shows that we have considered the relevant questions, and disposed of the irrelevant ones.
Perhaps the two main criticisms which we can level at the authors of MD is that they have misconceived the purpose behind our creation of legal expert systems, and that they are unaware of the other work in this field. To discuss the first issue, as we have explained[40] LESs have a narrow application and are developed for a range of reasons. If we constrain our efforts to producing ‘intelligent’ legal tools then there really is little in the way of a jurisprudential debate. These systems are, as we have said time and time again, little more than systems which can assist lawyers, and in some cases provide answers which lawyers would provide. The systems are at present stupid. MD presupposes that the systems about which we are arguing are supposed to be intelligent, as laypersons understand the expression. They are not. No researcher has chanced upon any methodology which promises to advance these machines any more than an infinitesimal step towards a truly intelligent system. In future, the field of AI and Law holds promise in producing much more intelligent systems: ones which reason like lawyers and can display the reasoning technique of lawyers to non-lawyers. Whether we will ever have LESs or Legal Artificial Intelligences which can operate at the same level as human lawyers is pure speculation. We reserve our judgment.
The second aspect which MD fails to grasp is that rule-based systems are now old-hat. As we have explained, logic programming and other rule-based paradigms do not reside at the forefront of the current research. This research focuses on case-based reasoning, machine-learning, neural networks and other promising technology. In one of our previous articles[41] we criticised Moles for failing to discuss technologies based upon paradigms other than logic programming. MD’s response is that we have missed the point and that
‘...one could plainly see from the title that the declared intention was to examine one particular approach, it would have been quite inappropriate to engage in some general survey of work based on other approaches.’[42] [their emphasis]
We accept that point, but we reiterate that Moles’ original article failed to address the other fields and tarred all research with the logic programming brush. MD’s discussion of rules, and the rather extreme jurisprudential approach taken by MD in relation to the use of rules in law, again fails to look at other approaches. Granted, MD discusses Stamper’s work but ignores all of the other work investigating knowledge representations in law and jurisprudence.[43] MD criticises all of AI and Law, however much it may say that it confines its arguments to logic programming and rule-based reasoners.
Having spent an unconscionably long time discussing legal expert systems in general terms, it may be useful to return to logic programming. Though we have been at pains to point out that there are a number of other methodologies and paradigms which are being implemented with often remarkable degrees of success, we should not overlook logic as a means of the formal representation of legal concepts. Logic is an incredibly powerful tool for expressing legal relations, and it so happens, is particularly amenable to computerisation. Though MD accuses us of using methodologies which fit the computer rather than law, we beg to differ.
Let us look at some of the benefits of logic and how it can be used in legal representations.
Logics are formal systems which are used by mathematicians, computer scientists and AI researchers to define and prove certain propositions. Lawyers also use forms of logic which may not be as formal as the mathematicians may desire. For example, the legal rule may be:
‘If you drive while drunk then you will lose your licence.’
This rule may be put into a formal logic system[44] as:
drink(x) & drive(x) -> licence_loss(x)
which simply means:
if person x drinks and also drives this implies that person x will suffer a licence_loss
where: drink(x) means person x has a level of alcohol in the blood above a certain limit;
drive(x) means person x is driving a type of vehicle, in this case a car, truck or motorcycle
licence_loss(x) means person x’s right to drive on public roads will be revoked.
However, even though this is in the statutes of Victoria, judges and magistrates have used their discretion to exclude what we may at first think is the obvious interpretation. For example, what of someone who has a level of alcohol in their blood above the proscribed limit and is driving, but is driving a bicycle which is fitted with an auxillary motor? Does the modified bicycle fall within this rule? This is the problem of ‘open texture’, a problem which we have previously mentioned and will not repeat here.[45] Nonetheless, the simple logic system we have used above can represent some extremely complicated concepts. This is in large part due to the fact that the system need not ‘understand’ any of the concepts used. For example, the system, whether it be our abstract logical system or a logic programming language, need have no knowledge of the idea of ‘drink’ or ‘drive’ or ‘loss of licence’. All it need be able to do is store facts and rules, and derive further facts from those given, using the given rules. Thus, the logic is extremely powerful in representing a whole range of concepts, including in this case, legal consequences.[46]
To use the above example, to determine who should lose their licence, we need merely determine the facts as to those people who were both drinking and driving.[47] In logic, if we were to attempt to determine which people would be liable to lose their licence,then we need merely move to the body (that part of the rule to the left of the -> sign) of the rule. We must determine those values of x for which both drink(x) has the value True and drive(x) has the value True. This simply means we must find those people who have been both drinking and driving.
Thus in both law, and logic, to determine those x for which licence_loss(x) we need only determine those x for which both drink(x) and drive(x). For both drink(x) and drive(x) this is merely a matter of verifying facts.
There may be other rules which lead to licence loss. These may include driving whilst having accumulated more than a given number of points in a certain period, or failing to give way at stop signs, driving in a careless manner, and so on. It is easy to include such knowledge in our rule base. By adding these rules we can eventually build a comprehensive ‘model’ of the area of law with which we are dealing. The model is a computer representation, in prolog, of the entire body of law. As we have discussed, we are not limited to using prolog, and may use other languages and methodologies, and that we can also begin to model law based on cases rather than explicit rules in legislative form.
We have demonstrated how logic can be used to help perform statutory interpretation. However, logic is not limited only to statutory interpretation, and may be used wherever rules can be derived from the law. Further, logic is particularly useful in handling a complex query and identifying which sub-queries need to be answered. We have also shown that the arguments about being unable to represent legal concepts in logic systems is simply incorrect. Since the logic system is only a means of representing relationships, we can represent any type of relationship in it, without having to ‘explain’ the nature of the actors to the system. This makes logic very powerful, since it is a very compact way to express ideas. It also means that logic can be used to represent law and produce complex and ‘intelligent’ legal expert systems, even though the logic system we use is relatively simple. As we hope we have shown, we should not write logic off as easily as MD do.
In conclusion then:
‘Definition: An economist is a person who says, “I know that it works in practice, but will it work in theory?” ’
The above might equally well serve as a definition of a jurisprude. It turns out that it cannot, for MD argue that it is impossible to know if it works in practice unless one has a sound theory to define the meaning of ‘works in practice.’ Failure to agree with this indicates a lack of respect for theory.[48]
Perhaps we are merely suffering from the terrible disadvantage of having received early training in subjects other than law (or perhaps jurisprudence), but we thought that theories were tested against reality. Wrong again. It turns out according to MD, that ‘empirical facts are…themselves the products of theoretical frameworks.’ Well, yes that is true to a point, but it does not follow that empirical facts are subordinate to theories. Bumblebees do fly, nuclear weapons do explode, expert systems do provide advice that is acceptable to professionals in many fields, and theories that lead to opposite conclusions are inadequate and must be re-evaluated.[49]
And when it comes to respect for theory, why is it that MD continue to ignore the theory which shows that production rules can be used to compute any computable function? If we understand their arguments correctly, then we will be able to build smart machines once we fully embrace their theories. But if we can build a smart machine, then it can be rebuilt using production rules. This is real theory, and it shows again that the debate over logic programming is misconceived and generally a waste of time.
Perhaps what MD really mean is that defining ‘works in practice’ to mean ‘the advice given is acceptable to professionals working in the field’ is not a satisfactory definition. If that is the case, then this debate is at an end because this is precisely what most AI researchers say that they are trying to achieve. MD may argue that the goal of AI researchers is misguided, that they are wasting their time, that their results will be useless, etc. But at least we can all be grateful that the logic programming debate is at an end.
But, unfortunately, that can hardly be what MD mean since much of the article criticises rule based systems for being unable to perform like a practising lawyer.[50]
Some of these arguments show a failure to understand what rule-based reasoning systems can do as opposed to what the Imperial College systems actually do.[51]
‘As a lawyer, one of the most important things you will probably want to know, before you engage in this process, is how much money your potential client has.’[52]
This seems a strange criteria to use, but there is clearly no difficulty in building a rule based machine that exhibits this particular lawyer-like behaviour. Nor is there any obstacle to defining different strategies based on the resources available.[53] MD seems to argue that rule based machines cannot be built which produce arguments when all of the precedents are on the other side.[54] This is simply wrong. Here is such a machine:
IF [All precedents favour the other side] THEN
Argue changed social conditions AND
Former precedents are wrong
The real question is whether we can build a machine which does this as well as a practising lawyer.[55] The same observations apply to their other criticisms. It is easy to build machines that ‘do’ each of the tasks that they mention, but the real question in each case is if it is possible to build machines that perform these tasks at an acceptable level.
We do not know the answer to that question, but if such machines can be built using MD’s theories or any other theory then they can also be built using production rules and logic programming. To argue otherwise shows quite an astonishing lack of respect for theory.
How shall we address the question of whether it is possible to build a machine which gives legal advice which is comparable to that of a practising lawyer? It is possible that jurisprudential theory will point the way—we hope that it undertakes this task, for it has been woefully inadequate to provide any guidance so far. Our money is on the less glamorous process: evaluate existing technologies by rigorous testing, determine the capabilities and deficiencies, make incremental improvements and repeat the process. And by ‘evaluate’ we mean to compare the advice against that given by human practising lawyers.
We are not saying that there is no place for jurisprudes in this new school of work. On the contrary, we welcome their involvement since we believe that future intelligent legal systems will need this assistance. If jurisprudes would sit down and take the trouble to appreciate the work already undertaken as well as the technology available, and then point out the difficulties, then we believe that all our work would progress all the quicker. This unfortunately is rare in this field. Rather, we have jurisprudes decrying development in the field, on the basis that they are unhappy with our lack of respect for jurisprudence. MD may quickly claim that it never said to AI researchers, ‘give up, the problem is beyond you.’[56] However, the import of MD’s doctrine that ‘we can never extract law from its facts’ gives little assistance to us poor technicians: jurisprudentially-illiterate automata toiling our thankless hours in the trenches. MD may claim that it seeks to help us work better[57] but fails to deliver on that claim.
In this, the Defendants’ Answer to the Plaintiffs’ Reply, we have sought to explain our position, why it is we have adopted the approaches we follow, and, we hope, answered some of the concerns aired by the MD Reply. We believe that these pleadings have narrowed the issues sufficiently, and that there will be no need for a Rejoinder and a Rebutter, or God forbid, a Surrejoinder or Surrebutter. The debate is essentially a sterile one, and there is a very practical way of resolving it: we must produce more and better systems and let History judge for itself.
[*] Daniel Hunter lectures in information technology law and artificial intelligence and law at the Law School, University of Melbourne. He has published a number of articles on the law of computers, and with Dr Zeleznikow has previously written on Artificial Intelligence and Law. He is co-editor of Computers & Law, the journal of the Australian and New Zealand Societies for Computers and Law. He is a Barrister and Solicitor of the Supreme Court of Victoria and High Court of Australia, and has worked as a computer programmer and in practice as a solicitor.
[*] Alan Tyree is the Landerer Professor of Information Technology and Law at the University of Sydney. He is a co-founder with Graham Greenleaf and Andrew Mowbray of the DataLex Project which has conducted research into legal AI systems and advanced database design. He has recently received a National Teaching Development Grant from the Committee for the Advancement of University Teaching to implement a cost-effective form of computer assisted learning for a number of law subjects.
[*] Dr John Zeleznikow is the Head of the Legal Reasoning Group and a Senior Lecturer in the Department of Computer Science and Computer Engineering, La Trobe University, Melbourne. His current area of research involves integrating rule based and case based reasoning to legal expert systems. He has published numerous articles on the construction of advanced legal expert systems, and has taught Artificial Intelligence and Law at the Law School, University of Melbourne.
[1] Moles, R.N. ‘Logic Programming - An Assessment of its Potential for Artificial Intelligence Applications in Law’, (1991) Volume 2(2) Journal of Law and Information Science 137-164 ; generating the response, Zeleznikow, J. and Hunter, D. ‘Rationales for the Continued Development of Legal Expert Systems’, (1992 ) Vol 3 (1) Journal of Law and Information Science 94-110 and Tyree, A., ‘The Logic Programming Debate’, (1992 ) Vol 3 (1) Journal of Law and Information Science 111-116; both of which provided the impetus behind Moles, R.N. and Dayal, S. ‘There is more to life than logic’, (1992) Vol 3(2) Journal of Law and Information Science p188-218
[2] Moles,R.N., op.cit (hereafter ‘Moles’). There was also some comment of a similar nature in the report on the major AI and Law conference contained in Brown, D., ‘The Third International Conference on Artificial Intelligence and the Law’, (1991) Volume 2(2) Journal of Law and Information Science p233-239
[3] Sergot, M.J., Sadri, F., Kowalski, R.A., Kriwaczek, F., Hammond, P. and Cory, H.T., ‘The British Nationality Act as a Logic Program’,(1986)Volume 29(5) Communications of the ACM 370-386
[4] Zeleznikow, J, & Hunter, D., op.cit. (hereafter ‘ZH’)
[5] Tyree, op.cit. (hereafter ‘Tyree’)
[6] For an overview of these researches see ZH, p107-110
[7] Moles, R.N. and Dayal, S. ‘There is more to life than logic’, (1992) Vol 3(2) Journal of Law and Information Science p188-288 (hereafter ‘MD’)
[8] For those who worry that the litigation analogy is not only far from perfect but also dangerously divisive (see for example MD, p189), we would just comment there is a conflict between our opposing viewpoints. Indeed, MD refers to narrowing the issues, ‘as with the lawyers’ exchange of pleadings’ (MD, p189). Happily for us all, no adjudicator sits in judgment of us. The old aphorism seems to hold true here as elsewhere: ‘The reason why battles between academics are so hard-fought is because there is so little at stake.’
[9] MD, p190
[10] Also known as Natural Language Processing (NLP).
[11] MD, p191
[12] See for, example, Winston, P.H., Artificial Intelligence, Addison Wesley, Third Edition, 1992.
[13] MD at 191 quoting ZH at 97
[14] Part of the work of researchers involved in NLP.
[15] Seen before in the earlier Moles article.
[16] By this we mean those poor unfortunates in practice as barristers and solicitors who do not have the benefit of a deep jurisprudential understanding, and instead must interpret the law as they find it. It is the experience of one of the authors and all of the practitioners surveyed that jurisprudence plays little part in day-to-day legal practice. Whether this is good or bad is irrelevant—it is merely true.
[17] Which the lawyers extracted from the cases itself.
[18] ZH, p101
[19] Polya, G, Patterns of Plausible Inference—a guide to the art of plausible reasoning, (Princeton University Press, 1954).
[20] See the earlier Moles article.
[21] Or for that matter, statutory regulation, common law doctrine, leading case, etc etc. The difference is immaterial since all in some way are ‘law’ to which the lawyer must have access and upon which the lawyer must advise.
[22] ZH, p96
[23] MD, p191
[24] Witness the ongoing and unresolvable debates between the positivists, rule-sceptics and the various other jurisprudential schools, who argue over how many lawyers can dance on the head of a pin.
[25] Again a point one of us has raised before: Tyree p115. See also Berman, D.H. and Hafner, C. ‘The potential of artificial intelligence to help solve the crisis in our legal system’, (1989) vol 32(8) Communications of the ACM 928-938. Strangely, MD raises the question of access to justice as question in determining the outcome of cases (MD, p199) and yet fails to recognise the potential benefits that LESs possess.
[26] [1932] AC 562
[27] This representation is not ruled out by problems with what is called ‘np-completeness’ which plague other areas of human endeavour.
[28] See for example HYPO (Ashley, K. D., ‘Arguing by Analogy in Law: A Case Based Model, (1988) Analogical Reasoning, pp 205-224 and Ashley, K. D., Modelling Legal Argument- Reasoning with Cases and Hypotheticals, (Cambridge, MA: Bradford/MIT Press, 1990); CABARET (Rissland, E.L. and Skalak, D.B. ‘CABARET: Rule Interpretation in a Hybrid Architecture’ International Journal of Man Machine Studies, 34(6) 1991, pp. 839); GREBE (Branting, L.K., ‘Building Explainations from rules and structured cases’, International Journal of Man Machine Studies, 34(6) 1991, pp 797 - 838 ); FINDER (Tyree, A. L., Greenleaf, G. and Mowbray, A., ‘Legal reasoning: the problem of precedent’, Proc. Conf. AJAI, Sydney, November 1987, pp 419-432); PROLEXS (Walker,R.F., Oskamp,A., Schrickx,J.A., Opdorp,G.J., Berg,P.H. van den , ‘PROLEXS: Creating Law and Order in a Hetorogeneous Domain’, International Journal of Man Machine Studies 35(1) 1991, pp. 35-68); IKBALS (Vossos, G., Zeleznikow, J., Dillon, T. (1990c), ‘Combining Analogical and Deductive Reasoning in Legal Knowledge Base systems - IKBALS II’, in Legal Knowledge Based Systems- Aims for Research and Development, (Koninklijke-Vermande 1991) pp 97-105.
[29] Hence the research into the use of standard AI architecture in law domains. Such architectures include frames, schemas, lists, logics, and a range of newer AI architectures such as object orientation, distributed AI systems and blackboard systems. These are simply tools to make representation and operation more efficient.
[30] An issue which we have dealt with before in the discussion of the proof that production rules models are equivalent to Turing Machines, and hence if computers can eventually reason as lawyers there is no reason why the computer could not be a production rule system. See Tyree p 113.
[31] MD also takes the opportunity of misrepresenting one of the seminal researchers in the application of jusrisprudence to AI and Law. It would be remiss of us not to mention this misrepresentation, for it gives a false impression of the work of Richard Susskind. MD says that the researchers in AI and Law are modelling consensus, and that this modelling of consensus ‘…is a central aspect of the work of Richard Susskind Expert Systems in Law (1987) Oxford University Press. The whole of his approach is based on the search for consensus.’ MD p195, n29. While this is true, the consensus of which Susskind speaks is not consensus in the legal source or legal representation, but rather consensus in jurisprudence. Susskind writes, ‘It would indeed be embarrassing for all concerned with jurisprudence if it transpired that we had to admit to computer scienists that, though we have been speculating about the nature of law and legal reasoning for well in excess of two thousand years, no matters of controversy have been settled, no agreement attained, in consequence of which legal theory has little to offer to the development of techniques of legal knowledge engineering. Such an admission is, of course, not necessary if legal theorists confine the bulk of their attention, as I do in this book, to identifying consensus within jurisprudence.’ Susskind op.cit. p35 It seems that the authors of MD claim that there is no consensus at all in jurisprudence, an assertion which Susskind disproves.
[32] MD p196-197
[33] Interestingly, the Hartian doctrine claims that law has a core with an umbra surrounding it. This accords fairly satisfactorily with most practical applications, though clearly would be rejected by Moles.
[34] Note the experience with MYCIN, Shortliffe, E.H., Computer Based Medical Consultations: MYCIN, New York, Elsevier, 1976.
[35] MD p197
[36] We skip MD’s discussion under the heading ‘ “Rules” “Categories” and “Consensus” as constructs.’ This section of the paper simply applies the principles explained earlier in the paper, which we have mentioned and answered. The primary thesis of this section appears to be that the authors of MD can always find an argument for any client, in the face of any precedent of lack of it. The reductio ad absurdum of this form of reasoning is that Bob Moles can win every case he takes. The discussion in any event ignores the fact that Ashley’s HYPO (Ashley, op.cit.) amongst others has sought to emulate the lawyer’s technique of arguing from both sides of a given case.
[37] MD p205
[38] MD p205, n.56
[39] MD p219
[40] ZH, p96-99, Tyree, p113-4
[41] ZH, p 103
[42] MD, p189
[43] To name but a few examples, there has been a great deal of discussion on the following jurisprudential questions in the standard AI and Law literature:
Toulmin’s structure of argument, see Lutomski, ‘The design of an attorney's statistical consultant’, Proceedings of the Second International Conference on Artificial Intelligence and Law, (ACM Press 1989); Marshall, ‘Representing the Structure of Legal Argument’, Proceedings of the Second International Conference on Artificial Intelligence and Law, (ACM Press 1989); Dick ‘Representation of Legal Text for conceptual retrieval’ Proceedings of the Third International Conference on Artificial Intelligence and Law, (ACM Press 1991); Bench-Capon, Lowes and McEnery, ‘Argument based explanation of logic programs’ (1991 ) Vol4 No 3 Knowledge Based Systems .
Arguing both sides of a case, seeAshley, K., op.cit; Rissland, E.L. and Skalak, D.B., op.cit.
Jurisprudence, AI and Law generally, see Susskind, op.cit.
[44] In First Order Predicate Calculus, one common type of logic system.
[45] ZH, p97-103
[46] For a detailed description of logic in law, see Zeleznikow & Hunter, Building Intelligent Computer Aided Legal Information Systems: Destroying the Myths, chapter 6, accepted for publication, Kluwer Law & Taxation Publishers.
[47] It may be simplistic to say that determining whether someone was drinking is a mere fact. In Victoria, Australia, anyone who has a blood alcohol level of above .05% is deemed to be drinking. So we might write this as a rule blood_alcohol_level(x,y) & (y > .05) -> drinking(x). But even then a lawyer could argue that blood_alcohol_level is not automatically verifiable, because the machinery measuring blood alcohol level might be faulty. Nevertheless, we shall assume, for simplicity, that blood alcohol testing machines are accurate and give a legally undebatable ruling as to whether a person has been drinking under the Victorian Motor Traffic Act.
[48] This is a very strange charge to lay against two mathematicians and a computer scientist! One wonders what kind of mathematics, if any, is taught to potential lawyers.
[49] Or are reduced to the status of religion or superstition.
[50] See MD, p197-204.
[51] Or rather, what the authors of MD suppose that those systems do.
[52] MD, p199.
[53] This type of reasoning is an essential feature of the machines that Mead and Johnson have built.
[54] MD, p200.
[55] That is: Does it ‘work in practice’?
[56] Tyree p 114, quoted in MD p189.
[57] MD, p190.
AustLII:
Copyright Policy
|
Disclaimers
|
Privacy Policy
|
Feedback
URL: http://www.austlii.edu.au/au/journals/JlLawInfoSci/1993/5.html