Mind, Machine, And Metaphor

An Essay On Artificial Intelligence And Legal Reasoning

Alexander E. Silverman

• Publisher: Taylor & Francis
• Year: 2019
• Print ISBN: 9780367158088
• eBook ISBN: 9780429722943
• Edition: 1
• Topic: Social Sciences
• Subtopic: Sociology
• Index: Social Sciences

Mind, Machine, and Metaphor is a rich, original, and wide-ranging view of legal theory in the context of artificial intelligence (AI) research. It is essential reading for legal theorists and for legal scholars and students of AI with an interest in each other's fields.

1

Introduction

People write about artificial intelligence (AI) and law from any of several viewpoints. They may write as computer scientists, describing the development and implementation of a legal expert system or similar computer program, often in considerable technical detail. They may write as practicing attorneys, discussing the application of particular computer programs or systems in actual law practice. They may write as commentators on legal doctrine relating to AI, especially intellectual property law. Finally, they may write as legal theorists, speculating on the jurisprudential implications of AI.

This essay takes the legal theorist’s viewpoint. Attempts to apply so-called classical AI approaches to legal problem solving have met with limited success. Classical AI approaches embody brittle models of law and of human reasoning, and the limitations of these approaches are in essence the limitations of their models. Legal theorists and jurists often subscribe to the same sorts of models and encounter the same sorts of problems. New directions in AI research solve some of the problems of the classical approaches while raising new issues of their own. These new directions can illuminate new directions for jurisprudence as well.

The essay’s organization is as follows: Chapter 2 is an overview of AI technology, focusing on issues relevant to later discussion. Also included is a discussion of legal expert systems and an introduction to the mind-machine metaphor. In Chapter 3 I use the well-known vehicle-in-the-park problem to illustrate the limits of classical AI and to develop concepts of indeterminacy, open texture, and essential vagueness. The centrality of the categorization problem to AI and to legal reasoning is discussed. Chapter 4 is an exploration of whether the concept of Kuhnian paradigm shifts has meaning in AI and law or, for that matter, in science. Discussions of legal pragmatism and the nature of theory set the framework for the analysis. In Chapter 5 new metaphors of law are developed, the most important being that of law as connectionist AI system embodied in society. The entailments of the metaphors are considered; in particular, legal rules are seen as best-fit approximations to the law. In Chapter 61 focus on what the Al-law analogy suggests about human legal reasoning. The notion of the rule of law as something other than the law of rules is suggested, and the relationship between judicial opinion writing and human explanation capability is considered.

This essay takes seriously the notion that metaphor, rather than being mere trope, is essential to human thought and language, and in particular to the process of jurisprudential theory-building. Changes in theory are linked to changes in conceptual or metaphorical framework. In developing new metaphors of law, this essay does not purport to describe what law is. Rather, it illustrates how law may be seen in new and potentially useful ways.

Thomas Grey writes that theory is good for getting us out of trouble as we attempt to solve problems. He goes on to note that theory can also be a medium of play—deep play, but play nonetheless.¹ It may well be that this or any jurisprudential essay is no more (or less) than a game. On the other hand, sometimes playing a game can be therapeutic.

2

Artificial Intelligence and …

There is such a thing [as human nature], and it is not entirely tractable.

—Melvin Konner, The Tangled Wing (1982)

Approaches to AI

Roughly speaking, AI is the attempt by computer scientists to model or simulate intelligent behavior on computers. An overview of the field is beyond the scope of this essay, but it is helpful to review some concepts that will be important later on. The classical and connectionist approaches, which compose the mainstream of AI research, are discussed in this section, along with the newer fuzzy approaches. (The more radical interactionist approaches are taken up in the last section of the chapter.)¹

Classical and Connectionist Approaches

AI researchers have adopted two main sets, or families, of approaches, the classical and the connectionist approaches. Classical approaches include rule-based expert systems, frame systems, blackboard systems, logic programming, and various other related approaches. They served as the basis of the vast majority of AI research from the early 1960s through the mid-1980s and continue to serve as the basis for most AI research today. Connectionist approaches, also called parallel distributed processing or artificial neural network approaches, attempt to model or to pattern themselves after the processing mechanisms of human and animal brains. Examples include backpropagation networks, Hopfield networks, and adaptive resonance models. Connectionist approaches were first tried as early as the 1940s but nearly died out in the late 1960s. They have enjoyed a resurgence within the past decade and, according to their proponents, have provided useful methods for solving certain problems that had been difficult or intractable under classical approaches.²

The various classical approaches, although dissimilar in many respects, share a great deal in common with one another. Classical AI systems are written in programming languages such as LISP, PROLOG, or variants of these. Typically they express information as verbal statements or lists of word-like symbols, e.g., DOG, CAT, VEHICLE; such symbols represent the most primitive level of information in the system and are not “understood” by the computer except as they relate to other such primitive symbols.³ Classical approaches thus may be said to represent knowledge in a relatively low-density, coarse-granularity fashion. They are designed to represent fairly directly what has been called “structured information,”⁴ that is, information that can be formalized as rules or at least as explicit propositions or statements. They are primarily symbolic rather than numerical in their computations, although they may employ probabilistic techniques to qualify the certainty or uncertainty of their results. Conceptually, their algorithms are serial, although parallel computational hardware may be employed to increase speed in particular implementations. Classical approaches model or mimic cognitive phenomena that occur in humans on time scales of one second or more. They employ deductive and inductive inferencing. They engage for the most part in monotonie inferencing.⁵

Connectionist approaches show a different set of commonalities. Connectionist systems are composed of massive numbers of densely interconnected units that act in parallel. “A single unit may correspond to a neuron, a cluster of neurons, or a conceptual entity related in a complex way to actual neurons.”⁶ Connectionist systems store information by altering the strengths of the connections between the units. Their knowledge representation may thus be said to be relatively high-density and fine-grained. Moreover, network behavior is an emergent property. Knowledge is not stored in particular units but is distributed throughout the network. “Learning” takes place via algorithms that use numerical information available locally at the unit level; local changes in connection strengths give rise to learning at the network level. Connectionist systems are designed to represent “unstructured information,” that is, information that has not been or cannot be formalized. The systems learn from examples (so-called supervised learning) or from experience (unsupervised learning), rather than from explicit statements or rules.⁷ They are capable of generalizing their knowledge to new situations. They exhibit graceful degradation of performance when units fail or when input data are noisy. They match or classify patterns holistically rather than relying on deductive or inductive reasoning. Conceptually, their algorithms are parallel, although they are most often implemented through simulations on serial hardware. They model or mimic cognitive phenomena that occur in humans on time scales of less than one second.

An important but sometimes overlooked point about connectionist systems: Basically, a connectionist network constructs or estimates a multidimensional mathematical function that represents the network’s target knowledge domain. The network may be said to be a very sophisticated curve-fitting program. A designer need not, and often cannot, specify the form of the knowledge representation function in advance; rather, the designer specifies the network geometry and learning procedure, and the network itself constructs the function as it learns from examples or experience. The storage of data as a knowledge representation function contrasts with the pigeonhole- or mailbox-like memory of a classical AI system.

Connectionist systems’ ability to generalize to new situations may be viewed as being a result of the mathematical properties of the knowledge representation function. The network constructs a function based on a training data set,⁸ attempting to find the best fit to sampled data. The function so constructed runs smoothly between sample points. The network thus responds easily to new data falling between sample points. Of course, the representation of data as a smooth function also creates the potential of confusion or cross talk. The function is an abstraction of the training set, and a network may fail to construct a function sufficiently complex to permit the distinction of every sample from every other sample. Such confusion is reminiscent of the memory confusion human beings sometimes exhibit.

Hoopla, Hype, and Hacks

Workers in artificial intelligence, unlike most scientists, almost never acknowledge their difficulties and are highly sensitive to criticism.

—Hubert Dreyfus and Stuart Dreyfus, Mind over Machine (1986)

Various AI researchers, both classical and connectionist, have from time to time overstated the performance of their computer programs, especially as regards the ability of these programs to produce intelligence comparable to human intelligence. Such overstatement may or may not be typical of most AI workers and probably is heard less frequently today than in years past. Some of the researchers’ hype may have been deliberate, made during sales pitches to customers or grant agencies. Nonetheless, the popular press has sometimes believed the hype. And unfortunately, the researchers have sometimes believed it, too.

A little debunking may be in order. It is not true that AI systems are intelligent in the sense that humans are, or that AI researchers are at this time anywhere close to building a computer as intelligent, say, as HAL in the movie 2001: A Space Odyssey. It is not true that the classical AI programs known as “expert systems” can truly replace human experts, or that they reason the way human experts do. Expert systems can occasionally outperform human experts and can be valuable aides to human experts, but they cannot do the sort of sophisticated, contextual, situated judgment that human experts do, usually automatically.⁹ Likewise, it is not true that the popular reference to connectionist systems as “artificial neural networks” means that connectionist systems are miniature brains. Nor is it true that the “neurons” that compose the computational elements of connectionist systems are anything more than highly abstracted representations of biological neurons, or that the “learning” of connectionist systems is anything more than a highly abstracted version of the learning that takes place in biological neural systems.

The term “artificial intelligence” itself is potentially misleading. The term is used at least three different ways. It is used to characterize computer science research related to the ultimate goal of creating a machine intelligence, that is, a computer program or system that exhibits intelligence comparable to human or animal intelligence. It is also used to characterize computer programs that attempt to model certain aspects of human intelligence so that such processes may be better understood in their own right. Finally, it is used to characterize computer programs that automate certain tasks that traditionally could be performed only by humans. These computer programs may operate according to principles having little to do with the way human cognition operates. Often, the programs’ designers have no pretensions about modeling human intelligence; they simply want to get the task done. Possibly, confusion among these different uses of “artificial intelligence” may have in some measure contributed to some AI researchers’ apparent self-delusions that their particular research projects were accurate models of human intelligence. More likely, the converse is true. (Despite the ambiguity, the term “AI” will be used in all of these senses, and perhaps others of which I am unaware, throughout this essay.)

Another reason for AI hype is the so-called first-step fallacy.¹⁰ When a research direction achieves an impressive success early on, people may tend to believe, falsely, that its success will continue. Certainly this was the case with classical AI, and critics argue that it is also the case with connectionist AI. In both cases, after first flushes of enthusiasm, difficulties have arisen, and at least in the case of classical AI, those difficulties appear to be fundamental.

It may be worthwhile at this point to review some of the limitations of both classical and connectionist approaches. The short version of the criticism levied against classical approaches is that they require a lot of adhoc tweaks (“hacks,” in computer jargon) in order to be made to do anything worthwhile. Although not entirely fair, this observation contains a large grain of truth. Classical rule-based expert systems, for example, may contain rules that employ ad hoc threshold or cutoff values used to discriminate between categories of input. As another example, there seems to be no general theory of how to design the frame data structures in classical frame-based systems. More generally, classical approaches tend to be brittle, prone to catastrophic failure, and do not respond well to new situations. They are seldom capable of learning from examples or from experience. They may be difficult to program in those cases where it is difficult to get human experts to articulate their expertise in the form of rules, which turns out to be very often. Indeed, it seems that human experts seldom use rules, and that step-by-step, rule-by-rule procedures are more characteristic of novice practice than expert practice. Experts employ sophisticated patterns of situational responses, patterns that for them have become nearly automatic. When asked, they must go out of their way to articulate how they do what they do.¹¹

In some respects, classical approaches’ limitations are connectionist approaches’ strengths. Connectionist systems have successfully performed certain pattern-recognition tasks such as converting printed English words to intelligible audible speech, tasks that humans do seemingly effortlessly and that classical AI approaches traditionally have done only with great difficulty and limited ability. Connectionist systems can display bidirectional associative memory. For instance, a LEXIS-like connectionist system designed to associate case names and keywords not only can recall the cases associated with a given keyword but can recall the relevant keywords if given the name of a case. Connectionist systems exhibit graceful degradation and learning and require no articulation of rules.

But connectionist approaches also require hacks. In a connectionist system, there are numerous design parameters that are chosen more or less ad hoc, including the number of units (nodes) in each layer of the network and the number of layers; the number of connections between nodes; the geometry of those connections; the initial distribution of connection weights prior to training of the network; the choice to use nodes that represent abstract neurons or nodes that represent more complex data (microfeatures), and if the...

Table of contents

• Cover
• Half Title
• Series page
• Title
• Copyright
• Dedication
• Contents
• Preface and Acknowledgments
• 1 Introduction
• 2 Artificial Intelligence and …
• 3 Simulated Vehicles in the Park
• 4 Paradigm Shifts
• 5 Law as Mind-Machine
• 6 Rules and Judgment
• 7 A Final Word
• Appendix
• Notes
• Bibliography
• About the Book and Author