A tribunal cited six cases. Three didn't exist.
On July 2, 2026, the Supreme Court of India set aside orders from the National Company Law Tribunal (NCLT) and the National Company Law Appellate Tribunal (NCLAT) in Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. (2026 INSC 668), after finding the tribunals had relied on citations that were, in the Court's words, "non-existent, fake and hallucinated material generated through artificial intelligence tools" (LiveLaw, judgment text via Indian Kanoon).
The underlying dispute was mundane: J&K Bank had extended ₹200 crore in credit facilities to Pan India Utilities Distribution Company, secured by a corporate guarantee from Essel Infraprojects Ltd. When the loan defaulted, the bank moved under Section 7 of the Insolvency and Bankruptcy Code, claiming ₹87.43 crore was owed. The NCLT's Mumbai bench admitted the insolvency plea on August 28, 2024; the NCLAT upheld it on September 11, 2025. Only on appeal to the Supreme Court did Senior Advocate Madhavi Divan, representing suspended director Pooja Ramesh Singh, show that six of the judicial decisions underpinning the order were fabricated — cases like State Bank of India v. Shree Ram Urban Infrastructure and Sarbjit Singh v. Union Bank of India either didn't exist or didn't say what the order claimed they said. It took two tiers of adjudication before anyone checked.
"No decision in the eyes of law"
Justice P.S. Narasimha, writing for a bench that also included Justice Alok Aradhe, reached for a striking comparison: AI-generated fake precedent, he wrote, is "like the release of methyl isocyanate in the province of law and justice — invisible, insidious and catastrophic." The Court's operative holding is a bright line: "A decision of a Court or an adjudicating authority based on material which is fake and hallucinated is no decision at all," and that taint voids the order "even if an iota" of fabricated material entered the reasoning — regardless of whether it actually changed the result. The NCLT has now been ordered to rehear the Section 7 application from scratch, on an expedited two-week timeline, with status quo preserved in the meantime (Bar and Bench).
Critically, the Court did not blame the technology. It directed the Bar Council of India (BCI) to constitute an expert committee to frame guiding principles on AI use by advocates, including disciplinary consequences, and to treat the issue "with utmost seriousness." That framing — misuse as a professional-conduct failure under the existing Advocates Act, 1961, rather than a case for new AI-specific court legislation — is the part worth paying attention to.
Not a one-off
This is not India's first brush with the problem, which makes the tribunal-level failure here more alarming, not less. In December 2024, Bengaluru's Income Tax Appellate Tribunal recalled a ₹669-crore order after finding it cited Supreme Court and Madras High Court rulings that did not exist. In October 2025, the Bombay High Court quashed a ₹27.91-crore tax assessment resting on three nonexistent precedents. Delhi High Court warned against "fictional case laws" as far back as August 2023, and Manipur High Court flagged a government pleader's undisclosed ChatGPT research in May 2024 (iPleaders). What changed in July 2026 is that the failure reached the apex court in a live commercial dispute worth crores, after passing through two specialized tribunals unchallenged — evidence that hallucination-checking habits have not caught up with hallucination-generating tools, even among people whose job is to catch exactly this.
The steelman for going further
There's a real case for a harder line than "let the Bar Council figure it out." Insolvency proceedings move fast, on tight statutory clocks, and NCLT benches are chronically under-resourced relative to caseload — precisely the conditions in which a specialist tribunal member might lean on an AI summarization tool for background research without an independent verification habit. A skeptic could reasonably argue that self-regulation by the same profession generating the problem is thin protection, and that mandatory citation-verification tooling, court-side, would close the gap faster than a disciplinary body that has historically moved slowly on far simpler misconduct complaints. The BCI's track record on internal accountability isn't spotless, and "guidelines," without teeth or a deadline, can sit in committee for years.
Why the Court's approach is still the right one
Even so, the Supreme Court's calibration is closer to correct than a prescriptive AI-use ban would have been. The failure here was not that AI tools exist in Indian legal practice — it's that nobody at two levels of adjudication verified citations against a primary source before signing an order, which is a professional-competence failure that predates generative AI and would exist with a sloppy paralegal's memo just as easily. Banning or heavily restricting AI research tools in litigation would not fix that; it would just remove one convenient scapegoat while leaving the actual verification gap untouched. Placing the burden on the Advocates Act's existing misconduct machinery — reprimand, suspension, removal from the roll under Section 35 — targets the actual point of failure (an advocate or bench member who filed unverified material) rather than the tool they used to generate it. That's proportionate, and it avoids importing a chilling, tool-specific liability regime into every other corner of Indian civil practice where AI-assisted drafting is now routine and mostly harmless.
The test now is whether the BCI actually delivers. A committee mandate with no public deadline, watched by a profession that has absorbed several similar warnings since 2023 without changing its habits, risks becoming exactly the kind of toothless guideline this ruling was supposed to prevent. The Supreme Court gave the Bar the right tool. Whether it uses it is a different question — and the next fabricated citation that reaches the apex court, rather than getting caught at the tribunal, will be the answer.