Can artificial intelligence replace human judges?

As framed, this question is incomplete. After all, human judges play a variety of roles and discharge a wide range of functions. A judge who presides over a criminal jury trial does something different from a judge who conducts a bench trial in a civil case. A judge who uses Dissomaster to calculate support payments in a marital dissolution case does something different from a judge who rules on a bail application. Trial judges do something different from what appellate justices do. So to ask whether AI can replace human judges is to condense a series of complicated questions into one question.

But there are some approaches to the general question which will help us if, and when, we try to address the specific ones. I offer two approaches which have to do with the role of language in asking and answering questions; and a third approach which is intended to remind us of the role some believe is essential for the law.

1. Can computers play chess?

When I was a philosophy student at Berkeley in the early 1960s, my friends and I used to meet in our favorite coffee shop to discuss what seemed to us to be a philosophical question: can computers play chess? Yes, my friends and I considered this to be an interesting philosophical question, and we argued about it endlessly. At that time, we focussed on the pure size of the problem – the number of alternative sequences of moves was too large for computers in those days to ingest, much less to digest; and on the difficulty of formulating algorithms which could be invoked at any position for either player, and which would generate “moves” which would both comply with the rules and advance the strategy. It seemed unlikely, but not theoretically impossible - so we argued.

Ultimately, when it came time for us to conclude our discussions because we were graduating and going to different places to live, I offered a resolution that my friends found satisfying: “it may be possible for computers to play chess, but the inescapable fact is that humans do play chess!” That observation made my friends smile. Some disagreed with me: humans have to be taught to play, just as computers do. So what difference is there really? But others agreed with me that my observation cast the whole discussion in a new light.

That was in the 1960s. Since then, AI has changed its shape from linear processes to simultaneous processes operating on “big data.” And not only have machines surpassed human players of chess and Go: a couple of years ago, the MuZero program running on the TPU processor was shown to be able to learn the rules of chess without being taught them specifically (again, by operating on big data and abstracting the rules from millions of games), and then, to learn how to surpass human players. So the technical limitations we believed existed 50 years ago are gone. But the fact remains that unless they are pushed and prodded to do so, computers do not play chess.

2. Can computers write great poetry? The burden of proof.

About five years later, while running my small software company, which developed “expert systems” – a special form of artificial intelligence – to optimize large-scale telecommunications networks, I organized a university-style debate attended by my colleagues and some of our customers, on the question of whether computers can write great poetry. The first speaker advocated for the position that such a thing was impossible, and the second speaker advocated for the position that yes, computers can write great poetry.

After each speaker had given his opening presentation, the other party was permitted to ask questions. The first question addressed to the “impossibility” speaker was: “What, exactly, is it that you contend computers cannot do?” They cannot create texts with the special rhythms required of great poetry. “Oh, yes, just give us a few months and we will adjust our programs to make sure that the texts they produce will have the proper rhythm.” But these texts also need to rhyme. “OK, just give us time and we will make sure the texts have rhythm and rhyme.” But then all the computer will be writing is doggerel: real poetry requires an inner meaning. “And what exactly is that?”

And then, the first question addressed to the “yes” speaker was: “What is it that you contend computers have done that qualifies as great poetry?” The texts they’ve generated have rhythm and rhyme. But great poetry requires more than that. What else do you contend puts them over the line?” The poetry has humor and charm. The fact that the listener might smile doesn’t mean that the poem is humorous, or charming. How does the computer itself feel about what it has written? Or, does it have feelings at all?

This debate made clear the dialectical nature of the issue. The party who assumes the burden of proving the proposition cannot carry that burden without convincing us of a detailed definition of “great poetry.” But the same goes for the party who assumes the burden of disproving the proposition: he also must identify in a convincing way some property of great poetry which the computer cannot produce.

3. Legal processes and religious services: justice as mystery.

It is difficult not to sense the similarities between legal processes – trials particularly – and religious services. The courtroom, like the religious sanctuary, is separated from the workaday outside world and seeks to maintain a sense of what might be called “holiness” – or at least, reverence. Both settings are structured by rituals. In both settings there are groups of people with specified roles who are entitled to speak only at designated times. In both settings, there is an unspoken effort to foster a unity of feeling among all the participants: a respect for justice. And these superficial similarities are rooted in something much deeper: the notion developed through rabbinic Judaism, the last “mystery religion” of the Alexandrian Age, and Islam, that justice is a pathway to God, the central mystery, and that law is its secret art. On this view, human judges are participants in that mystery, and unless they share the sense of reverence, and strive to foster it, they are not really doing their job.

So judging involves – and requires – sharing some inner sense, a sense of something important.

But one of the irritating trends in the development of Anglo-American law in the 20th century has been the elevation of the notion of contract to a foundational position in our law. The law of contract with which we deal today – contemporary contract law – was developed in the last decades of the 19th century and is basically a Victorian conception. Starting with Henry Sumner Maine’s wonderful book on Ancient Law, published in 1861, which traces the development of law from status to contract; and proceeding through the magisterial treatises of Corbin and Williston, as well as through the writings of Oliver Wendell Holmes, we see contemporary contract theory enshrined as the apotheosis of centuries of legal development (most of it European). In contracts, according to these authors, man’s free will and power of self-determination is given full expression, and society recognizes the individual’s freedom of choice and enforces the bargains individuals make. This body of law represents the adoption by the great legal thinkers of the Victorian Age of the unspoken assumptions and dogmas of industrialization, colonization, and commercial development.

Once this notion was ensconced in our thinking, Oliver Wendell Holmes could write in his 1897 lecture, “The Path of the Law,” that the fundamental work of lawyers was to predict how courts would act. He rejected the notion of mystery, and said that the object of the study of law is “prediction” – predicting how the courts will act. If that were in fact the fundamental enterprise of the law, then AI might actually carry out that task better than humans.

Conclusion

All three of the above ideas lead us to realize that sharing – an emotional commitment and shared aspiration – is central to the law. And that is something that AI cannot replace.

With regard to chess, there is something about the notion of “play” that is different from merely observing rules and trying to win. Play has to do with interrupting one’s workaday schedule, and seeking entertainment deliberately. Computers can depart from a programmed schedule, but they do not really seek “entertainment.” When we ask what the difference is between humans and computers playing chess, we come to the realization that the computer cannot “play” as we do, and so cannot share our sense of retreat from the world and immerse itself in the environment of the chessboard.

And when we consider how it might feel for an AI to replace a human judge, we realize that no matter how excellent the results might be, there would be no sense of the AI’s sharing or participation in the experience we undergo when we participate in a legal process.