When the fake law reaches the bench

Almost every AI citation scandal so far has followed the same script. A lawyer, pressed for time, asks a chatbot to do the research, files whatever it produces, and gets caught when someone actually checks the citations. The judge issues a sanction, the fabrications get struck, and the real body of law carries on undisturbed. The damage is embarrassing but contained; it stops at the lawyer who cut the corner.

India's Supreme Court ruling on July 2, 2026 in Pooja Ramesh Singh v. Jammu and Kashmir Bank is a different and darker thing, because the invented law did not stop at a lawyer's desk. It worked its way into the reasoning of two tribunals and became the foundation of their orders. The court that was supposed to catch the mistake had instead built on top of it.

The dispute underneath

The underlying fight was a corporate insolvency matter connected to the collapse of Essel Infraprojects, the sort of case that moves through India's National Company Law Tribunal (NCLT) and its appellate body, the National Company Law Appellate Tribunal (NCLAT), without ever troubling the front pages. These are the specialist forums that decide who gets paid, in what order, when a company goes under. Their orders carry real financial weight for creditors, debtors, and everyone standing in the queue behind them.

When the matter reached the Supreme Court, the justices did what appellate judges are supposed to do: they checked the authorities the tribunals had relied on. Several of them did not exist. Some of the cited judgments were wholly fictitious, complete with plausible-sounding names and holdings. Others were real decisions that had been quietly doctored, salted with quoted passages that appear nowhere in the actual rulings. The signature of a large language model asked to supply supporting precedent is unmistakable: everything looks correct, the case names sound right, the reasoning fits the argument, and none of it is real.

Methyl isocyanide in the province of law

The Court did not reach for restrained language. It compared the injection of fabricated AI judgments into the legal system to "the release of methyl isocyanide in the province of law and justice" - an unmistakable invocation of the Bhopal gas disaster, about the most loaded industrial-catastrophe metaphor an Indian court could choose. The point of the comparison was that a poison, once released into the record, does not stay where it started. A fake precedent cited once can be cited again, relied on by the next tribunal, and quoted in the next order, until nobody can tell which parts of the reasoning rest on real law.

From there the Court set out a rule with no wiggle room. A judgment contaminated by even a trace of hallucinated authority, it held, is not a flawed decision to be tidied up on appeal; it is "no decision in the eyes of the law." That framing matters. It means the remedy is not to sever the bad citations and salvage the rest. The entire order falls, because a court cannot know how much of the reasoning leaned on authority that was never real. The NCLT and NCLAT orders were set aside on that basis, and the matter sent back to be decided on law that actually exists.

Why the tribunal angle is worse

There is a meaningful difference between a lawyer citing fake cases and a tribunal relying on them. When a lawyer does it, the adversarial system has a backstop: opposing counsel and the judge are supposed to catch it. When the tribunal's own order carries the fabrications, that backstop has already failed. Whoever fed the hallucinated precedents into the process - a party's submissions, a research tool used in chambers, or some combination - the result is an official ruling of the state resting on invented law, with the imprimatur of a judicial body attached.

The Supreme Court's response treats the provenance as beside the point. It did not spend its energy assigning blame for how the fake citations got in. It declared that any decision touched by them is void, full stop, which puts the obligation on every tribunal to verify what it signs rather than to explain, after the fact, that the bad law came from somewhere else. That is the same non-delegable-duty principle that Western courts keep restating in lawyer-sanction cases, pushed up a level to the bench itself.

A directive, not just a rebuke

The Court did not stop at throwing out the orders. It asked the Bar Council of India, the body that regulates the country's lawyers, to examine the problem and frame norms for the use of AI in legal work. That turns a single insolvency appeal into something closer to national policy groundwork. India has a vast, overburdened legal system where generative tools are an obvious temptation for anyone drowning in filings, and the apex court has now put the profession's regulator on notice that guardrails are expected rather than optional.

The pattern keeps escalating

The Vibe Graveyard has catalogued a long run of AI hallucination cases, most of them ending in fines and reprimands for individual lawyers. What separates the Indian ruling is where the fabricated law was found and how far the court went in response. This was not a $10,000 sanction against a solo practitioner who trusted a chatbot; it was the highest court in the country vacating the work of two specialist tribunals, comparing the underlying failure to an industrial disaster, and demanding the legal profession's regulator respond.

The lesson courts keep relearning is that a language model does not retrieve law, it generates text that resembles law, and the resemblance is good enough to fool people who should know better - including, now, people wearing robes. Verification is not a nicety that can be skipped when the docket is full. When it gets skipped at the tribunal level, the cost is not one embarrassed lawyer. It is a judicial decision that the Supreme Court has to declare never legally happened at all.