Should Artificial Intelligence Decide Cases? The Question Courts Can No Longer Avoid

Through the side door

Artificial intelligence has already entered the courtroom, mostly without invitation. Judges around the world have been embarrassed by lawyers filing AI-fabricated citations — confident, fluent, and entirely invented. Court systems are piloting AI transcription, translation and case-management triage. Litigants-in-person increasingly arrive with AI-drafted submissions, which are sometimes better than what they could afford from the profession and sometimes dangerously wrong. All of this is the side door: AI assisting the humans who run the system.
The harder question is coming through the front door, and no serious judiciary will escape it this decade: should any part of adjudication itself — the weighing, the finding, the deciding — be delegated to a machine?

The case for

The argument for is not frivolous, and pretending otherwise wins nothing. Court backlogs in South Asia are measured in millions of cases and decades of delay; justice delayed on that scale is justice denied on that scale. Human judging is also demonstrably inconsistent — studies across jurisdictions find that outcomes vary with factors that have nothing to do with the merits. A system that decided small, routine, high-volume matters instantly, uniformly and cheaply would, its advocates say, deliver more justice to more people than the status quo delivers now. For traffic penalties and undisputed money claims, the question “why not?” deserves a better answer than tradition.

The case against

The better answer begins here: judgment is not computation. A judge gives reasons, can be questioned, can be appealed, and is morally and constitutionally accountable for the decision; a model gives outputs, and accountability dissolves into vendors, training data and version numbers. Systems trained on past decisions inherit past biases with perfect fidelity while laundering them through a machine’s apparent neutrality — and they freeze the law’s capacity to grow, for no algorithm trained on yesterday would have decided Donoghue v Stevenson or opened the door to public interest litigation. Deeper still lies the right to be heard: Article-level guarantees of fair procedure assume a mind that can genuinely attend to the litigant’s case. An unheard litigant before an unhearing machine has received a process, not a hearing.

Where the line belongs — for now

Our provisional view: AI as a tool for judges — research, transcription, translation, anonymisation, case management, even first-draft summaries a judge then owns entirely — deserves cautious welcome, with disclosure and verification duties to match. AI as a judge does not. But honesty requires admitting that the line between assisting and deciding is thinner than it looks: a “draft” judgment that a tired judge signs unchanged has, in substance, been decided by the machine. Drawing and policing that line is the constitutional task of the decade. We invite responses from readers on both sides — practitioners, technologists and judges alike. This column exists to host the argument, not to end it.