Copy of the email as sent to The Member Secretary, AI Committee, Supreme Court of India
I am writing to submit comments on the draft Regulations for Use of Artificial Intelligence in Courts, 2026, published on 3 June 2026.
I support the draft’s stance of active adoption over restriction. At the current stage of the technology, prohibiting the use of AI in courts is comparable to prohibiting the use of smartphones: the underlying capability is now mature enough that a restrictive default is no longer the correct posture, and the draft’s presumption in favour of responsible adoption is the right starting point.
My central comment does not concern hallucination rates as such. These rates are declining, and current systems are, for many practical purposes, approaching a near-deterministic level of accuracy. The concern I wish to raise is distinct from model accuracy and concerns the human element of this framework instead.
Over the course of human evolution and experience, people are trained to associate and read signals that the mind maps to categories such as legitimate, fraudulent, credible, dangerous, and so on. A large language model is capable of mimicking the signal of what appears legitimate to a human reader, even where it is operating in good faith, without there necessarily being any actual substance or guarantee underlying that appearance. Even where an AI system’s output is correct, nearly correct, or fully accurate, it remains difficult to train the human mind to maintain the discipline required not to defer to, or gloss over the contents of, output that carries this appearance of legitimacy.
This concern is distinct from the automation bias documented in other reliable, largely deterministic systems. In those cases, the system being relied upon is, in the overwhelming majority of instances, genuinely reliable and deterministic, and the risk lies in over-trusting a system that is in fact trustworthy. Here, the system is reliable-looking, and in some instances may appear more reliable than it in fact is. The distinction between a system being reliable and a system merely appearing reliable is the crux of the concern, and it is not addressed by provisions that rely solely on the discipline or professional judgment of the human reviewer.
The efficiency gains from AI adoption are real, and so is the risk of an output appearing efficient or authoritative while being incorrect. In low-stakes contexts, this risk is tolerable. In a context as high-stakes as a court of law, however, the appropriate response to error should not be limited to retrospective revision after a fault has occurred. There should instead be a deterministic graph or model operating underneath the system that automatically verifies citations to the relevant point of law or fact. In this context, retrospective revision is not a matter of inconvenience; it is not merely the equivalent of an annoying afternoon of correction. It may directly determine a litigant’s livelihood.
I would ask the Committee to consider mandating a deterministic graph or parsable format underlying AI systems used in Court processes: specifically, a computable, parsable representation of historic court data, as distinct from the current corpus of scanned or loosely formatted documents.
Such a structured representation would serve two distinct functions:
(a) It would improve the underlying AI systems themselves, by grounding them in better-structured, more reliable source data rather than unstructured or scanned material; and
(b) It would separately enable retrospective, automated verification between a court submission and its use before the Court, by allowing citations contained in that submission to be verified against the structured dataset.
I note that this proposal addresses the concerns above from the data end of the application, rather than solely through review obligations placed on the human reviewer at the point of use. I would suggest this is a necessary complement to, and not a substitute for, the Human-in-the-Loop requirements already present in the draft.
Comments on the Draft Regulations for Use of Artificial Intelligence in Courts, 2026