A litigant sitting outside a courtroom will likely never be told that an algorithm touched their file - unless it "materially" shaped the outcome.

The Word Doing All the Work Isn't "AI." It's "Non-Derogable." Most coverage of the Supreme Court's draft AI regulations reads them as a list of what technology is allowed to do in a courtroom. That's the wrong list to focus on.

The regulations permit AI for administrative tasks in fairly ordinary language - "actively seek opportunities," "demonstrably improve access to justice." What's unusual and what the exam should actually test, is the second list: the one marked "absolute and non-derogable." What "Non-Derogable" Actually Forecloses Most regulatory permissions in Indian administrative law carry an escape hatch - a competent authority can relax a rule, grant an exemption or phase in an exception.

The draft regulations deny themselves that option for a specific set of functions: risk scoring for flight risk, predicting recidivism, evaluating bail eligibility, assessing witness credibility and profiling a party's likely future conduct.

The text is explicit that "no authority can later permit them" - not the Apex Body, not a future amendment, not even the Chief Justice of India acting alone. That is a stronger commitment than most statutory bars, which Parliament or a rule-making body can usually revisit.

The reason this matters is what it protects against. India has no COMPAS-style controversy yet - the kind of racially skewed recidivism-scoring tool that triggered years of litigation and criticism in American courts. The regulations are pre-empting that failure mode before it has a domestic case to point to, which is unusual for Indian administrative rule-making, which typically reacts to a crisis rather than anticipating one.

The Litigant Who Isn't Told Here is where the regulation's protective architecture has a real gap. The disclosure duty triggers only when AI "materially assists" a judge - a threshold the regulations don't define with precision.

A litigant whose transcription, translation or preliminary research was AI-generated may never be told, because those uses sit below the "material assistance" bar even though they touch the same case file. The rule protects against AI deciding outcomes; it does not promise transparency about AI touching the process that leads to that outcome.

That gap is not necessarily a flaw - full disclosure of every administrative AI touch could bury litigants in irrelevant technical detail. But it does mean the real boundary these regulations draw isn't "AI versus no AI." It's "AI that assists a human decision" versus "AI that would replace one" - and the second category is fenced off completely, while the first is fenced off only from outcome-determination, not from silence about its own use.

For the exam, the insight worth carrying is this: when technology enters a governance function anywhere - courts, welfare eligibility, policing - the design question that actually matters isn't what the tool can technically do.

It's whether the rule permitting its use can be relaxed later by someone with the authority to relax it. A "non-derogable" bar, once written, is a structural commitment that outlasts the specific technology it was written for.