JUDICIAL DECISION-MAKING AND EXPLAINABLE ARTIFICIAL INTELLIGENCE

AuthorShaun LIM1 LLB (Hons) (National University of Singapore); Advocate and Solicitor (Singapore); Research Assistant, Centre for Technology, Robotics, Artificial Intelligence & the Law, Faculty of Law, National University of Singapore
Publication year2021
Published date01 December 2021
Date01 December 2021
I. Introduction

1 Advancements in artificial intelligence (“AI”) techniques with demonstrable results have led to a boom in AI research, development and marketing, especially in fields dominated by specialist professionals whose knowledge was thought to be impossible for AI to replicate, such as in medicine and law.2 In law itself, AI solutions are being

commercially offered for applications such as due diligence, legal analytics, outcome predictors, document automation and practice management.3 The Singapore judiciary, long an enthusiastic supporter of technological solutions in the courts, has itself pioneered many legal technological innovations over the years, such as the Integrated Case Management System4 for criminal case management, the Automated Court Documents Assembly5 service for certain categories of cases such as bankruptcy and Magistrate's Complaints, the Community Justice and Tribunals System6 for the Small Claims Tribunal, Community Disputes Resolution Tribunal and Employment Claims Tribunal cases, and even the Speech Transcription System,7 a voice recognition system for instant transcription of court proceedings.

2 The final frontier of the use of technology in courts then appears to be the actual rendering of judicial decisions itself, although few countries appear prepared to replace judges with AI. Even China, which has been devoting significant resources into developing technologically-advanced courts and legal assistants, appears to consider the use of AI in actual judicial decision-making to be unthinkable, saying that the expertise of judges is irreplaceable, not least by AI.8 It is of course possible to suggest several reasons for such a sentiment, such as societal confidence or lack thereof in AI, the need for the accountability that accompanies a societal presence, or the relative ease of training humans in common-sense reasoning as compared to AI. But in isolation, such a bald statement, made without substantiation, sounds merely like judicial exceptionalism, especially in light of the rapid advances that AI has made in recent years, and only exacerbates the risk of judiciaries being caught flat-footed if or when the spectre of AI judicial decision-making becomes reality.

3 Whether or not AI judicial decision-making comes to pass, the pace of AI-driven change in the legal sector suggests that it may be prudent for judiciaries to begin considering now what AI judicial decision-making may look like. There is already a need for judiciaries to begin grappling with the use of AI decision-making in related fields, such as the use of algorithms in administrative decision-making.9 Even if AI judicial decision-making fails to materialise, this exercise in thinking about what truly drives judicial decision-making may provide valuable lessons and thought leadership for a legal sector that increasingly augments its services with AI. And if AI judicial decision-making does indeed become reality, judiciaries will have drawer plans that they can rely on in taking charge of shaping the use of AI in the courts. It behoves them to ensure that courts continue to deliver the justice that society needs, even if not the justice that society sometimes may expect, and if AI were one day to make judicial decisions, judges must be there every step of the way to make sure that justice is not just done but is seen to be done.

4 This article is split into three parts. The first part deals with the current state of AI as applicable to the legal sector and developments in the field of AI. As is the case with every dynamic field, it will be quite difficult to be comprehensive, and this article will not purport to be so. Nonetheless, a taster in the form of general principles will suffice for present purposes. The second part breaks down what judicial decision-making entails, as far as is possible for laypersons to do so, in the hope of distilling the essence of judicial decision-making that any AI judicial decision-making solution must be able to replicate. The third part draws upon the analyses of the former two parts to consider what implementations of AI will satisfy the requirements of judicial decision-making, and therefore what is required to put together such an implementation of AI. It also considers if any aspects of judicial decision-making could change in response to the use of AI in judicial decision-making.

5 A key caveat is that the observations in the rest of this article on the usability of AI in judicial decision-making below hold up only for current implementations of AI as weak AI.10 Paradigm changes in the field may invalidate these observations. For example, quantum-11

and photon-based12 intelligence technologies both have the potential to significantly increase the efficiency of AI algorithms. Without practical data as to the effects of these advances in technology, it will profit us little to speculate on possible changes; being aware that such a possibility exists suffices at this point.
II. The workings of AI
A. Overview

6 AI generally refers to the phenomenon of computers “learning” to perform a task better, and hence simulating the human capacity to learn and process information. The foundations of AI research lay in the Dartmouth Summer Research Project on Artificial Intelligence in 1956, a gathering of a group of scientists interested in the field of AI, from whence came the Dartmouth Proposal: “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it”.13 Currently workable AI applications are limited to replicating specific features of human intelligence, and are known generally as “weak” AI.14 This is in contrast to the type of AI more commonly seen in science fiction and known technically as “strong” AI — an all-round AI capable of human-like thought and interactions with other humans, but which is not currently achievable with the present state of the art.15

7 This pursuit of the replication of human intelligence is by no means monolithic. AI research centres around various techniques that show promise for replicating intelligence, and the viability of these techniques have waxed and waned depending on a multitude of factors, such as scientific interest, developments in allied mathematical fields, and technological advancements. Research into an early form of artificial neuron, known as the perceptron, stalled after it was argued that existing computers lacked the necessary processing power to handle

large networks of perceptrons.16 Attention then shifted to the building of expert systems, aiming to encode expert knowledge.17 Although expert systems were very successful when they worked, the sheer complexity of the rules that governed human knowledge began to pose a problem for the updating and maintenance of such systems, leading to their gradual abandonment in favour of modern techniques.18

8 Current AI research revolves around machine learning — the training of an algorithm with correlated sets of input-output data, in order to predict an output given a particular input.19 This is achieved through a range of statistical techniques such as linear and logistic regression,20 tree models,21 and artificial neural networks,22 building on the back of advancements in computing and digital storage technology. In the legal context, an AI could for example be trained to predict the outcome of the division of matrimonial assets by analysing the parties' actuarial data.23 At the point of optimisation, the AI is in theory capable of returning or predicting what the outcome of any given input that is within the parameters of the training data would be. The value of such an AI is in its ability to process far more information than humans will ever be able to. However, because such techniques adapt to large amounts of provided training data iteratively, their exact method of operation and the way in which they arrive at results is not immediately explainable, if

at all,24 although the degree to which this is true varies from technique to technique and implementation to implementation. For instance, simple decision trees, being essentially a cascading series of classifications, are relatively easy to understand, whereas the most complex form of neural networks, with multiple factors feeding into each other at different levels and degrees, can be very difficult to plot and visualise in a graphical manner.

9 The difficulty in explaining their results notwithstanding, these sophisticated statistical techniques are powerful and accurate enough that they have been experimentally and sometimes even commercially applied in situations where they are capable of delivering results that can exceed those of trained professionals. An example is the use of AI in radiology and imaging diagnosis,25 where the performance of AI algorithms was shown to exceed that of practising radiologists in the detection of pneumonia, with the promise of extending the same techniques to the detection of other diseases. Closer to legal home, an experiment showed that one AI contract review software beat lawyers for both time and accuracy in reviewing non-disclosure agreements, albeit in looking out for specific clauses or issues that the AI was trained for.26 While much work still needs to be done to improve the rigour of these applications for frontline deployment, these experimental results show sufficient promise for industry stakeholders to consider pursuing them in earnest, as evidenced by the surge in interest in AI applications in modern industry.27

10 The success of these AI techniques also owes much to the extremely modern phenomenon of “big data”.28 Human history has never before witnessed the capability not just to granularise so many aspects

of human existence, but also to analyse and utilise such massive flows of information. The use of big data permits AI to rapidly analyse trends within the...

To continue reading

Request your trial

VLEX uses login cookies to provide you with a better browsing experience. If you click on 'Accept' or continue browsing this site we consider that you accept our cookie policy. ACCEPT