The Supreme Court chose an unusually grave metaphor for what happens when artificial intelligence’s falsehoods enter judicial decisions. Fabricated precedents, it said, are “like the release of methyl isocyanate in the province of law and justice: invisible, insidious, and catastrophic by the time anyone notices”. Methyl isocyanate was the gas that killed thousands in Bhopal in 1984.The comparison appears in the court’s July 2 judgment in Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. A bench of Justices P.S. Narasimha and Alok Aradhe set aside an insolvency order of the National Company Law Tribunal (NCLT). The order rested on six supposed Supreme Court decisions that were either non-existent or carried invented passages. The National Company Law Appellate Tribunal (NCLAT) had affirmed it without detecting the false authorities.The court declared that a decision tainted by even an “iota” of hallucinated material is “no decision in the eyes of the law”. It directed the Bar Council of India to frame rules and disciplinary consequences for lawyers who place unverified, AI-generated authorities before courts.But the record disclosed an awkward mismatch. Jammu and Kashmir Bank told the Supreme Court in an affidavit that its lawyers had not cited any of the six authorities. The judgment records that the tribunal obtained them through its “own research”.The disciplinary direction is aimed at advocates. The immediate institutional failure occurred within the adjudicatory process itself.A genuine dispute, contaminated by fictitious lawThe proceedings arose from a Rs 200 crore credit facility granted by Jammu and Kashmir Bank to Pan India Utilities Distribution Company Ltd. Essel Infraprojects Ltd had furnished a corporate guarantee, and mortgaged land at Gorai in Mumbai as security.After the borrower’s default, the bank moved against Essel under Section 7 of the Insolvency and Bankruptcy Code, claiming approximately Rs 87.43 crore.Essel’s defence was not that the loan had never been advanced, or that the borrower had not defaulted. It argued that a court-approved demerger and subsequent amalgamation had passed the relevant assets and liabilities to other group entities. Its guarantee, it contended, was consequently no longer enforceable.In its August 28, 2024 order, the Mumbai bench of the NCLT rejected that argument. It admitted the insolvency application, appointed an interim resolution professional and imposed the moratorium under Section 14 of the Code. The order was signed by judicial member Reeta Kohli and technical member Madhu Sinha.On September 13, 2024, the NCLAT directed that no further steps be taken under the admission order while the appeal was pending. That interim protection matters. Essel stood formally admitted to insolvency, but the consequences of admission were halted within a little over two weeks.The NCLAT eventually dismissed Essel’s appeal through its September 11, 2025 judgment. It held that clause 8 of the guarantee deed preserved Essel’s liability despite any absorption or amalgamation. In doing so, it reproduced the six authorities the NCLT had cited.The Supreme Court has not decided whether the tribunals were right about the guarantee. It restored the bank’s application for a fresh hearing, preferably within two weeks, and expressly left the merits open. The scandal lies in the materials through which the tribunals purported to reach their conclusions.Case names, citations and passages that did not existThe six authorities displayed different species of fabrication.The supposed decision in State Bank of India v Shree Ram Urban Infrastructure Ltd, bearing the citation 2020 SCC OnLine SC 341, carried the wrong case name. That citation in fact belongs to M. Subramaniam v S. Janaki, an unrelated decision. The paragraph attributed to the purported banking judgment does not exist.Everest Kento Cylinders Ltd v Union of India and Canara Bank v N.G. Subbaraya Setty were genuine cases with genuine citations. But the passages quoted from them occur nowhere in the reports.The citations assigned to ICICI Bank Ltd v Urban Infrastructure Real Estate Ltd, V.S. Dempo and Co. Ltd v Reliance Communications Ltd and Sarbjit Singh v Union Bank of India do not exist at all.This was no stray typographical slip. It combined fictitious case names, genuine citations attached to the wrong cases and passages invented to prop up a specific proposition. These are the signatures of an AI hallucination: output that resembles authoritative legal writing but collapses on verification.The NCLAT did not originate the false authorities. Its failure was appellate. It reproduced the NCLT’s list and affirmed the decision without checking whether the precedents existed.What does ‘own research’ mean?The most consequential passage in the judgment is also among its briefest.The bank’s affidavit stated that its counsel had not cited the disputed decisions. The adjudicating authority, the Supreme Court recorded, had obtained them through its own research.That finding rules out the simplest account: an advocate slipping fictitious authorities past an overburdened tribunal. It does not, however, establish who generated the material.The judgment identifies no AI platform, user account or prompt. It does not say whether the research was done by either member, a law researcher, the registry or someone else assisting the bench. Nor does it find that either member deliberately fabricated a case or knowingly approved false material.Those distinctions matter. Accountability cannot rest on speculation. The person who generated the text, the person who failed to verify it and the members who signed the order may bear different responsibilities.But uncertainty about authorship cannot convert an adjudicatory failure into the autonomous act of a machine. AI pronounces no order and authenticates nothing with a signature. Responsibility for the judicial product remains human and institutional.The direction to the BarThe Supreme Court declared that lawyers who submit unverified, hallucinated cases commit professional misconduct. It asked the Bar Council of India to constitute a committee, formulate preventive norms and prescribe disciplinary consequences.There is ample reason for such action. In February, The Wire reported the court’s alarm at AI-assisted pleadings containing fictitious cases and invented extracts. One supposed precedent bore the implausible title Mercy v Mankind.Days later, the Narasimha-Aradhe bench itself took suo motu cognizance of an Andhra Pradesh trial court order resting on fabricated judgments. In October 2025, the Bombay high court had quashed a Rs 27.91 crore tax assessment founded on three non-existent decisions. The Bengaluru bench of the Income Tax Appellate Tribunal recalled a December 2024 order in the Buckeye Trust matter for the same infirmity, prompting Justice B.R. Gavai’s caution that AI must remain an aid, never a replacement.The BCI direction therefore addresses a systemic problem. It does not complete the accountability inquiry arising from the Essel case.The judgment calls a judge’s reliance on fake material a “serious lapse” and announces zero tolerance for Bar and Bench alike. Its only operative regulatory direction deals with the Bar.There may be an explanation. During the hearing on May 19, the bench said the citation issue was already being examined on the Supreme Court’s administrative side, and had been referred to a committee.The final judgment does not identify that committee, define its jurisdiction or state whether either NCLT member has been asked for an explanation. It is silent on whether the committee is examining this order alone, tribunals’ wider use of AI or the NCLAT’s failure of scrutiny.An ongoing administrative inquiry may justify judicial reticence while facts are gathered. It does not extinguish the public interest in knowing whether the inquiry concluded, what caused the failure and what safeguards followed.The Supreme Court’s own draft rules point to the answerThe judgment comes a month after the court published its draft Regulations for Use of Artificial Intelligence in Courts, 2026, whose contours The Wire has reported.The proposed regulations extend to courts, tribunals and statutory commissions performing adjudicatory functions. They define hallucination to cover fabricated precedents, statutory provisions and factual claims. They require human supervision, treat AI output as advisory and demand verification before use.The draft also bars a submitting party from pleading the AI-generated character of hallucinated material as a defence; responsibility stays with the person who submits it. It contemplates AI registers, an incident database, annual audits and documented lines of responsibility.The regulations remain a draft and cannot retrospectively fix culpability. But they embody the principle that should govern the response to this incident: responsibility must be traceable.That demands more than invalidating the order. The drafts, research histories and digital records should be preserved. It must be established who generated the authorities, who incorporated them, what checking occurred and how six defective citations survived two tiers of adjudication.The individuals concerned must have the opportunity to explain their roles. A careless research error, a supervisory failure, reckless adoption of machine output and deliberate fabrication carry different degrees of culpability. A fair inquiry should distinguish among them. Its outcome, subject to legitimate confidentiality, should be disclosed.Human in the loop, accountability in the systemThe Supreme Court is right that the decisive safeguard against hallucination is a human in the loop. The phrase becomes an empty assurance unless it identifies which human must verify what, at which stage and with what consequence for failure.The NCLT order passed through research, drafting, deliberation and signature. The NCLAT then reviewed it. Six fictitious or corrupted authorities survived every stage.A judgment can be set aside. The harder question is what happens to the people and processes that produced it.Zero tolerance should not mean automatic punishment or guilt by association. It should mean verifiable facts, an impartial inquiry and consequences proportionate to responsibility. It should apply with equal seriousness whether false law enters from the advocate’s lectern, a researcher’s computer or the adjudicator’s desk.Otherwise the system will discipline lawyers for the hallucinations they submit, while treating hallucinations generated within adjudication as technological accidents. The poison the court described would keep an unguarded route into the province of law.V. Venkatesan is a contributing editor at Supreme Court Observer. Views are his own.