A note for deeper search and reflection as the Supreme Court's Artificial Intelligence Committee receives public comments on the draft 'Regulations for Use of Artificial Intelligence (AI) in Courts, 2026' until 15 July 2026.

1. Something curious happens whenever an artificial intelligence tool invents a judgment that does not exist. The profession pounces. The offending lawyer is hauled up, the show-cause notice issues, the bar buzzes, the story is run with a barely concealed delight, and a thousand commentators pronounce, with the satisfaction of a prophecy fulfilled. That the machine is after all a 'stochastic parrot'—a mimic without a mind, a statistical conjurer of plausible falsehoods. The episode enters the institutional memory as proof of something. And the question I wish to ask is: proof of what, exactly?

2. I do not defend hallucinated citations. A fabricated authority placed before a court is an impediment to the administration of justice, whether its author is silicon or flesh, and the lawyer who files it without verification deserves what follows. My question is not about the fabrication. It is about the velocity and the relish of our reaction to it—and about the strange, monastic silence that surrounds errors of an altogether graver order when they are committed by human beings wearing robes.

3. Let me put a real case on the table — a reported one. In Shanti Devi v. State of Haryana, Criminal Appeal No. 2861 of 2025, decided on 6 August 2025, the Supreme Court examined a murder said to have been committed in 1997. The trial court convicted a mother and her son, along with a third accused and sentenced them to imprisonment for life in 2003. The High Court, exercising its full appellate jurisdiction — reappreciating evidence, hearing counsel, applying its collective judicial mind — affirmed the convictions in 2024. Two tiers of the judicial hierarchy, operating with every safeguard our system provides, concurred that these persons should spend the remainder of their lives in prison. Then the Supreme Court read the record and found that there was nothing in it: the extra-judicial confessions on which the prosecution principally rested were, in the Court's own words, "a cock and bull story created by the police". It acquitted the mother who had appealed. And then it did something remarkable: it acquitted her son, Rajbir, who had never appealed to the Supreme Court — who had, after the High Court had confirmed his conviction, for whatever reason of poverty, despair or ignorance, accepted his fate — because in deciding the appeal before it, the Court had inevitably and comprehensively traversed his case too, and having seen that the foundation was non-existent for all, it refused to let the accident of a missing appeal keep an innocent man in prison. From the invention of that story to its judicial burial, the hallucination sat on the record, wearing seals and signatures, for twenty-eight years. Nor is it an isolated specimen. Only months earlier, in Nusrat Parween v. State of Jharkhand, 2024 INSC 955, decided on 10 December 2024, the Court found that the prosecution had failed to prove even the chain of circumstances, that the trial court and the High Court had clearly erred in certifying that fiction as proof beyond reasonable doubt — and, invoking Article 142 of the Constitution, it extended the acquittal to a co-convict, Abdul Rahman Khan, who had likewise never appealed to it.

4. And here is the part that should keep us awake at night: nothing was concealed. The falsehood was not buried in some forensic subtlety that only the Supreme Court's instruments could detect. It lay on the open face of the record, for twenty-eight years, waiting for anyone to simply read it. Consider what the file itself said. Three accused persons — a mother, her son, and another woman — were each supposed to have walked up to three separate strangers and spontaneously confessed to murder, in strikingly similar fashion, as though the guilty in Haryana travel door to door making announcements. The plot, stated plainly: three murderers in search of unconnected persons to make a confession of gruesome murder. Two courts accepted it as proof beyond reasonable doubt — and on its strength sentenced a mother and her son to spend the rest of their lives in prison. One of those strangers was never even examined by the prosecution. Another did not know the accused by name — and yet somehow named her. The forensic report on the recovered weapons came back negative; the science said no; and both courts relied on the recoveries anyway. No advanced degree was needed to see through this. A moment's honest pause — would three human beings really do this? — would have collapsed the story at the first hearing. But nobody paused. The trial judge read the chargesheet and completed the pattern. The High Court read the trial judgment and completed the pattern. Each court did precisely what we accuse the machine of doing: it took a prompt and produced the most expected continuation — fluent, well-cited, internally coherent, and false. We have a name for that when silicon does it. When it is done under oath, in open court, over a mother and her son who then vanish into prison for the better part of their remaining lives, we call it "concurrent findings of two courts".

5. There are two lessons in that judgment, and they pull in opposite directions. Both deserve to be stated plainly, because the current debate over artificial intelligence in our courts insists on hearing only one of them.

6. The first lesson is genuinely glorious, and it is this: no artificial intelligence system in existence, and none on any visible horizon, would have done what the Supreme Court did for the men who never appealed to it. An algorithm optimised for docket disposal decides the case before it and closes the file. The parties not on the array are, computationally speaking, out of scope. It took a human mind—several human minds—to notice the men who were not in the room; to feel the moral absurdity of freeing an innocent mother while leaving her equally innocent son to suffer on the technicality of a missing memorandum of appeal; and to reach past the pleadings to do complete justice to persons who had never asked for it. That is compassion. That is moral imagination. That is the capacity to be haunted by someone else's silence. It is not computable, it is not approximated by any large language model however many parameters it carries, and it is the true and sufficient reason why human primacy in adjudication is worth defending. When the Supreme Court's draft AI Regulations insist that no machine shall ever adjudicate, this judgment is the argument that clinches it—and it clinches it on grounds of conscience.

7. But the second lesson is the one nobody wants to sit with. Ask the question in the machine's vocabulary: what is a hallucination? It is output that is plausible, internally coherent, confidently delivered—and false. By that definition, the trial court judgments in these cases were hallucinations. The High Court judgments affirming them were hallucinations. Two courts, across years of proceedings, generated elaborate, reasoned, formally impeccable texts confidently asserting the guilt of innocent people in cases that were horrible in their foundations. Those hallucinations did not embarrass a lawyer at a hearing. They consumed years of human liberty—years that no appellate correction can restore, years measured out in prison cells while the hallucination sat on the record bearing seals and signatures, and but for the fortuity of further appeals and an unusually attentive bench, they would have consumed entire lives.

8. Now perform the comparison our profession refuses to perform. When an AI tool hallucinates a citation, there is a register for it: the draft Regulations propose an AI Incident Database at every court's AI Secretariat (Regulation 39), annual audits (Regulation 38), disclosure endorsements (Regulation 43), grievance mechanisms (Regulation 52) — and, for good measure, a public AI Register recording every incident (Regulation 37) and an Annual Transparency Report tabulating them (Regulation 45). When two courts hallucinate guilt, there is nothing. No incident report is filed. No database records the event. No audit asks how the trial judge and the two appellate judges came to certify a bogus case as proof beyond reasonable doubt. No annual transparency report tabulates how many such cases exist—and every criminal practitioner in this country knows the acquittals I have described are only specimens. Indeed, the law reports supply their own confession: from Dandu Lakshmi Reddy v. State of A.P., (1999) 7 SCC 69, through Bijoy Singh v. State of Bihar, (2002) 9 SCC 147, Suresh Chaudhary v. State of Bihar, (2003) 4 SCC 128, and Pawan Kumar v. State of Haryana, (2003) 11 SCC 241, to Sahadevan v. State of Tamil Nadu, (2012) 6 SCC 403, the Supreme Court has had to build, brick by brick over a quarter of a century, a settled doctrine for extending acquittal to convicts who never appealed to it. Pause on what that means. A doctrine is not built for an event that happens once. The very existence of this jurisprudence is the institution's own admission that entire prosecutions certified by two courts collapse at times to require a standing rule for rescuing the men left behind. The judgment is celebrated, correctly, for the Supreme Court's intervention; the two hallucinations it corrected are filed away as the system 'working as intended', because the appeal succeeded. We measure the system by its last-gasp correction and not by the error that made the correction necessary.

9. Why the asymmetry? I suggest the answer is insecurity, and it is worth being honest about it. The machine's error comforts us. Every hallucinated citation is seized upon because it confirms the story we need to tell ourselves—that the machine is a parrot and we are the thinkers, that its fluency is fake and ours is real, that there remains something irreducible about us. And so we amplify its failures with the enthusiasm of a threatened incumbent. The judge's error, by contrast, indicts us. It cannot be laughed off as the babbling of a parrot, because it issues from the very faculty—trained human judgment—on which the whole edifice rests. So we do with it what institutions always do with evidence against themselves: we individualise it ('a bad judgment'), we proceduralise it ('that is why appeals exist'), and we move on. The machine's hallucination is treated as a property of the species; the judge's hallucination is treated as an accident of the instance. That is not analysis. That is self-protection.

10. And notice what the honest comparison actually yields. It does not yield the conclusion that machines should judge—paragraph 5 forecloses that, and I stand by every word of it. It yields something more uncomfortable: that the case for human primacy rests on compassion and moral agency, not on accuracy, because on accuracy the record of purely human adjudication is nothing to defend a monopoly with. Once that is admitted, the regulatory implications follow naturally. Seize upon AI hallucinations, certainly—verify, disclose, sanction. But do it with the humility of an institution whose own hallucinations wear robes and carry seals, and whose gravest errors are punished, if at all, decades late and only when some further appeal is filed. Build the AI Incident Database that Regulation 39 promises—and let its existence prompt us into asking why no equivalent has ever been thought necessary for the incidents that put innocent men in prison for life. Demand explainability from the machine—and remember that the two most consequential documents in the case I have described were fully 'explained', beautifully reasoned, and false.

11. The men who never appealed to the Supreme Court walked free because human beings on the bench possessed something no machine possesses. That is the glory of the judgment. The reason they needed to walk free at all is that human beings on two earlier benches confidently certified a fiction. That is its warning.

The sooner we admit that our anxiety about the machine is really an anxiety about ourselves, the sooner we can regulate it like adults: welcoming the tool where it exceeds us, confining it where it cannot follow us, and reserving our fiercest vigilance not for the errors that embarrass lawyers, but for the ones that bury the innocent.

K V DHANANJAYAdvocate, Supreme Court of India