Elsevier

Artificial Intelligence

Volume 289, December 2020, 103387
Artificial Intelligence

Explanation in AI and law: Past, present and future

https://doi.org/10.1016/j.artint.2020.103387Get rights and content

Abstract

Explanation has been a central feature of AI systems for legal reasoning since their inception. Recently, the topic of explanation of decisions has taken on a new urgency, throughout AI in general, with the increasing deployment of AI tools and the need for lay users to be able to place trust in the decisions that the support tools are recommending. This paper provides a comprehensive review of the variety of techniques for explanation that have been developed in AI and Law. We summarise the early contributions and how these have since developed. We describe a number of notable current methods for automated explanation of legal reasoning and we also highlight gaps that must be addressed by future systems to ensure that accurate, trustworthy, unbiased decision support can be provided to legal professionals. We believe that insights from AI and Law, where explanation has long been a concern, may provide useful pointers for future development of explainable AI.

Introduction

The English essayist Charles Lamb famously wrote “He is no lawyer who cannot take two sides”. For many, the same is true of AI and Law programs. Arguing for just one side, or worse simply pronouncing for one side, is not enough. To make a convincing case in court, one must be able to offer reasons for one's own side, and to anticipate and rebut arguments for the other side. The first important AI and Law program, TAXMAN [86], set out to reconstruct the arguments of the majority and minority opinions in the famous tax case of Eisner v Macomber1 In TAXMAN, there was no interest in assessing and deciding between the two opinions: the purpose was simply to be able to argue for both sides. The point is that the outcome of a case is often not clear: in any serious legal dispute there are opposing arguments, and very often opinions differ as to who has the better of it. Decisions are reversed on appeal, and may be reversed again at the highest level of appeal. Even at the highest level, where the most gifted lawyers are the judges, consensus is, as in Eisner, far from invariable: an article in the Washington Post2 stated

“According to the Supreme Court Database, since 2000 a unanimous decision has been more likely than any other result — averaging 36 percent of all decisions. Even when the court did not reach a unanimous judgment, the justices often secured overwhelming majorities, with 7-to-2 or 8-to-1 judgments making up about 15 percent of decisions. The 5-to-4 decisions, by comparison, occurred in 19 percent of cases”.

Although this article was in fact arguing that consensus was the norm, the results still indicate disagreement in the significant majority of cases, and the narrowest of majorities in nearly a fifth of cases. Consensus may be the most likely result of the ten possible (the quorum is 8), but disagreement remains far more likely, and 5-4 the second most likely result. Given that even the most expert people can disagree, it would not be reasonable to accept a judgement from a machine unless backed up with convincing reasons.

In recent years there has been some research directed towards the prediction of case outcomes using algorithms applied to large data sets (e.g. [5] and [88]), but for most of its history AI and Law has been far more interested in modelling the reasoning to explain the outcome (and to offer reasons for alternative possible outcomes) than in predicting the outcome itself. AI and Law therefore offers an interesting area in which to explore methods for the explanation of AI programs,3 as advocated in the most recent Presidential Address to the International Association for AI and Law [129]. In this paper we will review a number of approaches. Before we do so, however, we will consider some general points about explanation, especially in law.

Apart from the centrality of argumentation to legal reasoning, the intellectual challenge of modelling legal reasoning and the availability of, in the form of opinions on cases, a large volume of examples, there is another important reason why explanation is vital for artificial intelligence applied to law. This is the right to explanation [54]. In a legal dispute there will be two parties and one will win and one will lose. If justice is to be served, the losers have a right to an explanation of why their case was unsuccessful. Given such an explanation, the losers may be satisfied and accept the decision, or may consider if there are grounds to appeal. Justice must not only be done, but must be seen to be done, and, without an explanation, the required transparency is missing. Therefore explanation is essential for any legal application that is to be used in a practical setting.

In his recent illuminating survey [90], Miller gives four main findings of features of explanation. These are:

  • Explanations are contrastive. As well as explaining why a particular classification is appropriate, a good explanation will also say why other classifications are not. This is often done using counterfactuals and hypotheticals: “if x had been true, then the classification would have been A, not B”.

  • Explanations are selective. Rarely is a logically complete explanation provided, but rather only the most salient points are presented unless more detail is required by the recipient of the explanation. The assumption is that there will be a considerable degree of shared background knowledge, and so the explanation need only point to some fact or rule as yet unknown to the recipient.

  • Explanations are rarely in terms of probabilities. Using statistical generalisations to explain why events occur is unsatisfying since they do not explain the generalisation itself. Moreover, the explanation typically applies to a single case, and so would require some explanation of why that particular case is typical.

  • Explanations are social. Explanations involve a transfer of knowledge, between particular people in a particular situation and so are relative to the explainer's beliefs about the explainee's beliefs.

Miller says that he believes “most research and practitioners in artificial intelligence are currently unaware” of these features. AI and Law, however, has long recognised these features, and made them an important part of its approach to explanation.

  • Contrastive explanations can be found in legal case based systems such as HYPO ([108] and [7]), quite possibly the most influential of all AI and Law programs [21]. Indeed the name HYPO is itself short for hypothetical: one of the main motivations of the system was to explore how the hypothetical variations on cases would change their outcome. Also there are explanations based on the weighing of pro and con reasons such as the Reason Based Logic of Hage [69] or the tool developed by Lauritsen [77].

  • Selective explanations were pursued by several AI and Law researchers. Often this was done through the use of argumentation schemes such as that of Toulmin [123], used in, for example, [83] and [26]. The idea was to present the key data items which gave rise to the inference and to suppress things that should be expected to be already known, such as “John is a man” or “67 > 65”, unless explicitly requested by the user.

  • Probabilities are rarely used in legal decisions. Even where Bayesian reasoning is used in AI and Law, the explanation is presented not in probabilistic terms but as scenarios [38], [130] or arguments [121]. A legal decision is supposed to determine what is true on the facts of the particular case. An 80% probability would mean that one in five cases would be decided wrongly, which would not be justice.

  • Legal explanations are inherently social, occurring in the context of courtroom procedure, and involving an interaction between plaintiff, defendant, judge and, possibly, jury. This is reflected in the popularity of dialogues as the vehicle of explanation in AI and Law, such as [71], [61], [16] and [128].

Thus AI and Law provides an excellent domain in which to study explanation of AI systems. In AI and Law explanation has a long history, is a mandatory feature of fielded applications in the legal domain, and AI and Law has long recognised the important facets of explanation identified in [90]. The rest of this paper will be structured as follows. Section 2 will give an overview of the main types of explanation used in AI and Law. The various types of explanation will then be described in more detail in sections 3, 4 and 5. Section 6 will look at interactive explanations, section 7 will consider efforts to explain machine learning in AI and Law and section 8 will look at some current research directions that may become influential in the future. Concluding remarks are provided in section 9.

Section snippets

Explanation in AI and law

In its early days, AI and Law followed two main approaches: cased based approaches such as TAXMAN [86], HYPO [108] and CATO [6], and rule based approaches including Gardner's account [57], approaches using production rules [114], and approaches using logic programming [115]. For some time case and rule based reasoning were seen as alternatives [35]. Each gave rise to distinctive styles of explanation: case based systems tended to explain by offering precedent cases as examples, while rule based

Explanation through examples: case based reasoning

Although there have been several approaches to reasoning with legal cases, including the use of prototypes and deformations [87] and semantic networks [40], by far the dominant approach has been the use of dimensions and factors [21]. This approach will therefore be the one considered in detail in this section.

Step by step explanation: rule based reasoning

For our example of this style of system we will consider logic programming in the style of [115], or, as applied to case law, [33]. For case law, this approach requires that a set of rules be derived from the precedents, encapsulating the knowledge that they represent. This does, however, require some degree of interpretation on the part of a knowledge engineer or domain expert. Moreover, the interpretation is subject to change, and the rules may require reconsideration in the light of new

Argumentation based explanation

All of the above explanations can be seen as arguments, reasons for adopting the conclusion. This is natural enough since a legal trial comprises both sides presenting their arguments. This being so it was sensible to look at ideas about argumentation from Informal Logic. This led to the notion of argumentation schemes, first that of Toulmin [123], and later the schemes proposed by Walton [131]. Also in the mid-90s the notion of abstract argumentation [55] emerged, and this too had an important

Interactive explanation

The desire to include selective and social elements led to interest in the use of interactive explanations through dialogues. Another attempt to improve the presentation of explanations was through the use of visualisations.

Explaining machine learning

Until very recently the use of machine learning in AI and Law to make and predict decisions was limited, primarily because the explanation facilities were unsatisfactory. The prevalent view was similar to that recently expressed by Robbins:

“the explanations given by explicable AI are only fruitful if we already know which considerations are acceptable for the decision at hand. If we already have these considerations, then there is no need to use contemporary AI algorithms because standard

Future directions

In this article, we have so far discussed various ways that have been used in existing AI-based legal systems to provide explanations at various levels. Compared to the earlier systems that used association rules and/or a limited set of human-engineered features, modern-day machine-learnt AI and Law systems automatically derive salient features from massive data collections in natural language using deep learning techniques and so pose a complex set of challenges with regard to explainability,

Concluding remarks

In this paper we have described the various traditional methods for explaining the reasoning of systems in AI and Law. Despite a recent upturn in interest in machine learning methods, such as [5], [47], [43] and [88], doubts remain about the quality of explanation produced by such systems without the guidance of human experts. Therefore traditional methods continue to be pursued in the development of practical systems [4], and methodologies to support such systems continue to be developed [3].

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References (142)

  • N. Aletras et al.

    Predicting judicial decisions of the European Court of Human Rights: a natural language processing perspective

    PeerJ Comput. Sci.

    (2016)
  • V. Aleven

    Teaching case-based argumentation through a model and examples

    (1997)
  • K.D. Ashley

    Modeling Legal Arguments: Reasoning with Cases and Hypotheticals

    (1990)
  • K.D. Ashley et al.

    Automatically classifying case texts and predicting outcomes

    Artif. Intell. Law

    (2009)
  • K. Atkinson

    Introduction to special issue on modelling Popov v. Hayashi

    Artif. Intell. Law

    (2012)
  • K. Atkinson et al.

    Parmenides: facilitating deliberation in democracies

    Artif. Intell. Law

    (2006)
  • K. Atkinson et al.

    Realising ANGELIC designs using Logiak

  • D. Bahdanau et al.

    Neural machine translation by jointly learning to align and translate

  • L. Baldini Soares et al.

    Matching the blanks: distributional similarity for relation learning

  • P. Baroni et al.

    An introduction to argumentation semantics

    Knowl. Eng. Rev.

    (2011)
  • T. Bench-Capon

    Neural networks and open texture

  • T. Bench-Capon

    Specification and implementation of Toulmin Dialogue Game

  • T. Bench-Capon

    Representation of case law as an argumentation framework

  • T. Bench-Capon

    Persuasion in practical argument using value-based argumentation frameworks

    J. Log. Comput.

    (2003)
  • T. Bench-Capon

    Try to see it my way: modelling persuasion in legal discourse

    Artif. Intell. Law

    (2003)
  • T. Bench-Capon

    Representing Popov v Hayashi with dimensions and factors

    Artif. Intell. Law

    (2012)
  • T. Bench-Capon

    HYPO's legacy: introduction to the virtual special issue

    Artif. Intell. Law

    (2017)
  • T. Bench-Capon

    Before and after Dung: argumentation in AI and Law

    Argum. Comput.

    (2020)
  • T. Bench-Capon et al.

    Dimensions and values for legal CBR

  • T. Bench-Capon et al.

    Lessons from implementing factors with magnitude

  • T. Bench-Capon et al.

    Persuasion and value in legal argument

    J. Log. Comput.

    (2005)
  • T. Bench-Capon et al.

    Argument-based explanation of the British Nationality Act as a logic program

    Inf. Commun. Technol. Law

    (1993)
  • T. Bench-Capon et al.

    A dialogical model of case law dynamics

  • T. Bench-Capon et al.

    Developing heuristics for the argument based explanation of negation in logic programs

  • T. Bench-Capon et al.

    A case study of hypothetical and value-based reasoning in US Supreme Court cases

  • T. Bench-Capon et al.

    Back to the future: Dimensions revisited

  • T. Bench-Capon et al.

    Towards a rule-based representation of open texture in law

  • T. Bench-Capon et al.

    PLAID: proactive legal assistance

  • D.H. Berman

    Developer's choice in the legal domain: the Sisyphean journey with DBR or down hill with rules

  • D.H. Berman et al.

    Representing teleological structure in case-based legal reasoning: the missing link

  • R. Bevan et al.

    Extracting supporting evidence from medical negligence claim texts

  • F.J. Bex

    Arguments, Stories and Criminal Evidence: A Formal Hybrid Theory

    (2011)
  • T. Bolukbasi et al.

    Man is to computer programmer as woman is to homemaker? Debiasing word embeddings

  • L.K. Branting

    Data-centric and logic-based models for automated legal problem solving

    Artif. Intell. Law

    (2017)
  • L.K. Branting et al.

    Scalable and explainable legal prediction

    Artif. Intell. Law

    (2020)
  • L.K. Branting et al.

    Semi-supervised methods for explainable legal prediction

  • G. Brewka et al.

    Abstract dialectical frameworks revisited

  • S. Brüninghaus et al.

    Predicting outcomes of case based legal arguments

  • B. Buchanan et al.

    The MYCIN Experiments of the Stanford Heuristic Programming Project

    (1984)
  • D.L. Chen

    Judicial analytics and the great transformation of American law

    Artif. Intell. Law

    (2019)
  • Cited by (101)

    • Minimality, necessity and sufficiency for argumentation and explanation

      2024, International Journal of Approximate Reasoning
    • Directive Explanations for Actionable Explainability in Machine Learning Applications

      2023, ACM Transactions on Interactive Intelligent Systems
    View all citing articles on Scopus

    This paper is part of the Special Issue on Explainable AI.

    View full text