Proponents of artificial intelligence have long championed its use in litigation as a means to achieve frictionless access to justice and consistency in outcomes. They see the deployment of large language models (LLMs) in courts as a natural culmination of our century-long effort to make the judge more than just the mouth that pronounces the word of the law—but a cog in a system that should, at least collectively, travel past the idiosyncrasies of individual men.
Despite the concerted efforts of the administrative State to systematise, they say, a judgement today is as impelled by an individual judge’s pathologies as it is by law. This turn towards systematisation, which had taken root in the late 19th century with our engagement with law as a coherent body of principles uncovered through systematic analysis of case law and legislation, would only be strengthened by the use of AI.
They claim that the impending shift towards broader adoption of AI would finally create a system that is resistant to our many pitfalls. An LLM does not tire by the day’s end, nor does it carry the irritation of a previous matter into the present, or bring arguments from home into the court. It does not discriminate between lawyers based on their standing, nor can it be a lawyer’s champion on the Bench.
Whatever may be held against it, AI, when trained properly, the thinking goes, it may help reduce overt arbitrariness, if not totally eliminate it from adjudication. However, their central claim is not that AI would ever be a better judge than humans. Rather, the argument reveals the continued push for creating the type of judge that we always wanted and never managed to produce: general, indifferent and blind to the person standing before it.
While there may be some truth to this, the behaviour of these automatons is still wildly unpredictable: their functioning is learned rather than specified, meaning the manufacturer lacks the control to monitor their inference—the phase where a trained AI model processes our prompts through its neural network to calculate and generate an output—and the operator lacks control over their actions.
They have no compass to navigate life, even if they were ever to make sense of it. And unlike us, whose ethical judgement arises from reasoning, social conditioning and empathy, AI is a product of pure logic dictated by statistics. Which brings us to the question of integrating AI into litigation and high-stakes decision-making.
Let us take one of the oldest and most-used devices in ethics and moral science: the trolley problem, which has become a subject in its own right. A runaway trolley will kill five people, but you have the option to divert it so that it kills one. Generations of students have been asked whether they would pull the lever to divert the trolley, or they would push a heavy man onto the tracks to stop it, or they would jump off to save themselves the trouble. We have heard hundreds of answers to this question, and each one is produced by the same sharp, resistant intuition the thought experiment was designed to manufacture.
What if the trolley were to be an autonomous car? Every assumption immediately goes for a toss. The machine does not decide alone. It has been hardwired through training and datasets. The company had chosen certain traits to dominate others. Its engineers and the company’s supplier ecosystem would have pushed to bring the machine into the market. As such, when a car is set to make the decision, it is not facing the choice that humans clamber with—the decision was already made years back and only the consequences are in the here-and-now. In many ways, such decisions are not responses to the live moral conundrum but results of distribution. They are rules—and, in the case of LLMs, ones whose basis we cannot trace.
This applies equally to decision-making. When AI gets involved in adjudication of disputes, even if it were only to assist, what it does can hardly be termed ‘deciding’. What it does is more analogous to legislating. The answer is already settled even before the case makes it to court, by a machine that seeks out patterns and applies it more so determined by pattern than by facts and conduct of parties. Unlike humans, these decisions are not based on morals, social conditioning or empathy. When AI encounters a real-life case, it faces data for classification, not a moral dilemma.
We lawyers have always recognised this danger and have provided it a fitting name: the rule against fettering discretion. It is a form of illegality where an authority granted with discretion adopts a policy but shuts its ears to individual cases. Such authority must remain willing to listen to individual applicants and to depart from the policy where the facts demand it.
American legal philosopher Lon Fuller located this rule at the root of adjudication itself: what distinguishes adjudication from every other mode of reasoning is that the affected party participates by proof and arguments addressed to the person deciding. What AI removes is not the discretion but the power to distinguish, the fact of presence and contestable language.
Courts, including those in India, have been slow in adopting AI for these very reasons. Even where guidelines have been issued, the courts have promoted AI only as research aids and sorting hats. None of which means that AI should be kept at bay. What is important is that those at the steering wheel be aware that the sorting hat should come with holes.
Saai Sudharsan Sathiyamoorthy | Advocate, Madras High Court
(Views are personal)