Singapore’s Chief Justice Sets a Line for AI in the Judicial Function
Sundaresh Menon’s graduated framework puts AI strongest at gathering information, weaker at analysis and weakest at writing reasons in Judicial Function. The ordering travels well beyond a courtroom.
The Chief Justice of Singapore used a lecture in Brisbane this month to set out where artificial intelligence should and should not sit inside adjudication. Sundaresh Menon delivered AI and the Judicial Function: The Limits of Artificial Judgment at the 28th Annual Australian Judicial Officers Association Colloquium on 9 October, and the Supreme Court of Singapore circulated a summary on 30 October. The courts’ own account places the lecture as the second of five marking the bicentennial of the Second Charter of Justice.
The framework he proposed is more useful than most governance writing on AI because it says where the line falls and why, rather than asserting that humans must stay in control.
The organising principle
Menon’s rule is proportionality to the judicial function: the closer AI comes to exercising judgment, the greater the need for restraint. He grounds the analysis in three requirements of sound adjudication, judicial integrity, procedural fairness and institutional effectiveness, and observes that most of the case for judicial AI rests on the third. Scale, speed and baseline reliability promise courts that serve more people faster with outcomes generally correct. Effectiveness alone cannot justify every use, and that is a stricter reading of the judicial function than most institutions adopt when they buy a tool.
Disaggregating adjudication
The practical contribution is the instruction to break adjudication into its working parts and ask what each part exists to do. On that test the case for AI is strongest where a court acquires information, weaker where a model analyses material or proposes an outcome, and weakest where it formulates the reasons for a decision.
That ordering is the argument. Acquiring information produces a checkable output, because a reader can compare a summary against the document it summarises. Analysing material introduces interpretation. Writing reasons is where a judge takes responsibility for the decision, and a reason generated by a model is not the judge’s reason even when it reads like one. Menon’s closing line is that AI is inescapably artificial, cannot bear judicial responsibility and cannot replicate the human encounter at the heart of a fair hearing.
The risks worth naming
He identifies three risks in human-machine collaboration: over-reliance, skills fade and hidden institutional delegation. The third is the one institutions discuss least. A court that adopts a vendor’s tool inherits that vendor’s assumptions about what counts as relevant, how a document should be summarised and which precedents surface. The delegation leaves no signature, which is what makes it hidden.
His answer is not retreat. He argues for governance that preserves meaningful human judgment and control, and offers five practical observations on what that entails. The summary does not list them; the court publishes the full speech and links to it from the release.
Why it travels
The same ordering, in fact, applies wherever a professional is accountable for a decision a model helped produce. A radiologist, for instance, reading a flagged scan, an auditor testing a model-selected sample and a claims adjuster approving an automated recommendation each sit somewhere on Menon’s scale. The scale tells them something specific: the further a model moves from gathering evidence toward forming the conclusion, the more explicit the human control has to be, and the less a general promise of human oversight is worth. Menon’s judicial function test is at bottom a test about where accountability attaches, and accountability does not transfer to a model however fluent its output.
What to watch
One marker is whether the remaining lectures in the series extend the framework to the profession and to legal method, since Menon’s own account of AI’s effect on the rule of law covers those alongside the judiciary’s core work. Another is whether regulators outside law borrow it, because a test built for the judicial function is unusually clear about what it is protecting. A third is whether any court publishes the governance arrangements he describes, because a framework without a worked example remains an argument rather than a practice.