Pooja Ramesh Singh v. J&K Bank [2026] INSC 668: India's Supreme Court sets aside a tribunal's own AI-hallucinated 'precedents' — 'no decision in the eyes of the law'

The problem

Generative AI can fabricate case law that reads like the real thing, and when fabricated authorities enter a judicial record the decision built on them is corrupted at the root. The unusual and serious feature of Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. is that the fake authorities were not slipped in by a careless advocate — they were relied on by the tribunal itself. The Supreme Court of India opened its judgment by recording that “this is yet again a case where the Tribunal relied on non-existent, fake and hallucinated material, generated through Artificial Intelligence (AI), as if it were a precedent in support of its judgment” (source). Reporting the decision, the independent legal newsroom Bar & Bench put the same fact plainly: the Supreme Court “set aside a National Company Law Tribunal (NCLT) judgment on Essel Infraprojects insolvency after finding that the tribunal had relied on non-existent, fake and hallucinated judgments/ precedents generated through artificial intelligence (AI) tools” (source).

What was built

There is no deployment to celebrate here — this is an AI-governance honest-negative, and the “system” is an unnamed generative tool whose output was treated as law. The matter itself was an insolvency dispute: J&K Bank filed a Section 7 application under the Insolvency and Bankruptcy Code before the NCLT, Mumbai, against Essel Infraprojects (the corporate guarantor), the NCLT admitted it on 28 August 2024, and the NCLAT dismissed the appeal on 11 September 2025. What unravelled both orders was that the authorities underpinning them did not exist. The Court recorded that “Respondent No. 1 has filed an affidavit indicating that the alleged judgments relied on by NCLT were not cited by its counsel at the bar,” that “the so-called precedents relied on by the adjudicating authority were obtained through its own research,” and that “the fake, non-existent judgments escaped scrutiny by the first statutory appellate tribunal” (source). The appellant’s own senior counsel had flagged that the citations were “fake and non-existent, probably AI-generated” (source). The Court reached for a deliberately alarming metaphor, warning that fake AI-generated material used as precedent “is like the release of methyl isocyanate in the province of law and justice: invisible, insidious, and catastrophic by the time anyone notices” (source) — a characterisation Bar & Bench independently reported as the Court warning that such precedents are “invisible, insidious and catastrophic” to judicial determination (source).

The outcome

The remedy was to set the tainted orders aside and to lay down a rule, not to fine anyone. On the tainted decision, the Court was categorical: “We have no hesitation in declaring that such a decision is no decision in the eyes of the law, irrespective of whether such material had a direct or indirect bearing on the decision-making,” adding that “such decisions are to be set aside even if an iota of fake or hallucinated material enters the decision-making process, as it would violate the sanctity of adjudication” (source). On the professionals, it drew the line at both Bar and Bench: “It is necessary for Courts to adopt a zero-tolerance mode for producing, citing or using AI-generated precedents without verification. It is a misconduct on the part of an advocate to cite such judgments without verification. Equally, it is a serious lapse if a judge relies on such a fake or hallucinated AI-generated material as precedents in support of the determination” (source). It nonetheless affirmed AI’s rightful place, framing its “resolve to adopt AI technology in aid of adjudication, while at the same time asserting and declaring total and absolute control over adjudication, with a human in the loop at every stage” (source). For enforcement it turned to the regulator, directing “the Bar Council of India, being the apex statutory body, to constitute a committee and deliberate on this issue of members of the bar submitting such fake and hallucinated material before the Court as if they are precedents of law” (source). No monetary penalty was imposed on any party or advocate; the NCLT and NCLAT orders were set aside and the Section 7 application was restored for fresh disposal (source). The independent court-monitoring publication Supreme Court Observer, restating the ruling in its own editorial voice, records that “The Court adopted a zero-tolerance standard for the bar and the bench against citing unverified material, while permitting responsible, human-controlled AI assistance” and that “The Court directed the Bar Council of India to constitute a committee to formulate disciplinary guidelines for advocates submitting fake or hallucinated legal precedents” (source).

The weak-source honesty line, where you meet the claim: the primary here is the apex-court judgment itself, but it was read this session from a reproduced full-text copy rather than from an authoritative court host — Indian Kanoon, which carries the judgment, Cloudflare-blocks this session’s tools and even the Wayback save (HTTP 520), so the byte-tie to an official Supreme Court / eSCR capture is the one open verification step; the neutral citation (2026 INSC 668) and appeal number were independently confirmed on CaseMine, and three independent legal publications (Bar & Bench, LiveLaw and Supreme Court Observer) restate the core findings. The three non-primary sources are Tier 2 reports, not the court record, and are used only to corroborate facts already stated verbatim in the judgment (the set-aside, the AI-hallucination finding, the “invisible, insidious and catastrophic” warning, the advocate-misconduct holding, the zero-tolerance declaration and the Bar Council direction), never to carry a fact the judgment does not.

Verification status: CHECKING — court-origin apex ruling, AI-hallucinated material found by the court itself; handed to the checker, not verified.

Path to green

The court-stated facts need no subject confirmation: the origin is a published, citable Supreme Court of India judgment (2026 INSC 668) delivered on 2 July 2026. Three independent legal publications carry the core findings firsthand — Bar & Bench (fetched this session, the set-aside and AI-hallucination finding), LiveLaw (2026 LiveLaw (SC) 653, organic Wayback capture 20260703063029 predating this session) and Supreme Court Observer (its SCOLR report, which restates the zero-tolerance declaration and the Bar Council of India committee direction in its own editorial voice) — which together place both critical claims at two independent secondaries above the court-origin primary. What the independent public record still needs to yield is an authoritative capture of the primary judgment the checker can reach directly — the Supreme Court’s own judgment portal / eSCR PDF, or an organic Wayback capture of the Indian Kanoon page (currently Cloudflare-blocked) — plus a post-judgment currency check that no review or recall petition disturbed the ruling and whether the Bar Council of India committee direction has produced any rule. No green badge is sought; verified is human-only.

How this was verified

• Method: Every quoted finding was read verbatim from a source fetched this session and confirmed byte-present in the locally saved source copy. No number, quote or finding is paraphrased, rounded or recalled.
• Primary anchor: The court-origin apex judgment, Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. & Anr., 2026 INSC 668 (Supreme Court of India, Narasimha and Aradhe JJ, 2 July 2026), read from a reproduced full-text copy; neutral citation and Civil Appeal No. 11950 of 2025 independently confirmed on CaseMine.
• Independence: The set-aside and the AI-hallucination finding are restated firsthand by two independent legal newsrooms — Bar & Bench (“Supreme Court sets aside NCLT judgment for relying on hallucinated AI-generated citations”, 2 July 2026) and LiveLaw (2026 LiveLaw (SC) 653, organic Wayback capture 20260703063029). A third independent publication, Supreme Court Observer, restates the zero-tolerance declaration and the Bar Council of India committee direction in its own editorial voice (locally saved copy, byte-present), so both critical claims sit at two independent secondaries.
• Standard: Court-origin honest-negative. The outcome is the published set-aside, the declaratory holding and the Bar Council direction, so the public record — not any subject’s confirmation — is the proof; green never depends on the tribunal, the bank or counsel agreeing.
• Open item: An authoritative capture of the primary judgment reachable by the checker (Supreme Court / eSCR PDF, or an organic Wayback capture of the Indian Kanoon page), and a post-judgment currency check on any review/recall and the Bar Council committee — verification steps, not gates on the court-stated facts.
• Verified on: 2026-08-20 (maker draft; checker pass pending).

Related case files

• Re A, B, C, D [2026] EWFC 71 (B) — an England & Wales family court names an unregistered barrister who put four AI-hallucinated authorities before it — the closest peer on a court’s non-monetary response to fabricated authorities, but here the hallucinated ‘precedents’ were relied on by the tribunal itself, not merely filed by an advocate.
• Noland v Land of the Free (California Court of Appeal) — first published opinion imposing a $10,000 sanction for AI-fabricated citations — another apex/appellate court treating the published opinion itself as part of the deterrent, but pairing it with a monetary sanction the Indian court declined to impose.
• Murray Wamba Wemba v Victoria [2025] FCA 731 — Federal Court of Australia, indemnity costs for AI-fabricated citations — a superior court in another common-law system reaching for costs rather than a rule-making direction to a bar regulator.
• Asif v Minister for Immigration [2026] FedCFamC2G 1402 — a completely fabricated AI ‘confession’ reaches an Australian court, above-scale costs — another 2026 court confronting wholly fabricated AI material entering the record, and calibrating a proportionate judicial response.

Sources

1. [Tier 1] Supreme Court of India, Narasimha and Aradhe JJ · Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. & Anr., 2026 INSC 668 / 2026 LiveLaw (SC) 653, Civil Appeal No. 11950 of 2025 · judgment · 2 July 2026 · https://indiankanoon.org/doc/113338666/
2. [Tier 2] Bar & Bench · “Supreme Court sets aside NCLT judgment for relying on hallucinated AI-generated citations” (S N Thyagarajan) · 2 July 2026 · https://www.barandbench.com/news/litigation/supreme-court-sets-aside-nclt-judgment-for-relying-on-hallucinated-ai-generated-citations
3. [Tier 2] LiveLaw · “2026 LiveLaw (SC) 653 | Pooja Ramesh Singh versus Jammu and Kashmir Bank Ltd & Anr” · 2 July 2026 · https://www.livelaw.in/sc-judgments/2026-livelaw-sc-653-pooja-ramesh-singh-versus-jammu-and-kashmir-bank-ltd-anr-539623
4. [Tier 2] Supreme Court Observer · “Reliance on AI-generated fake precedents” (Supreme Court Observer Law Reports / SCOLR) · July 2026 · https://www.scobserver.in/supreme-court-observer-law-reports-scolr/reliance-on-ai-generated-fake-precedents/