Pooja Ramesh Singh v. J&K Bank (2026): a decision built on AI-hallucinated precedents is 'no decision in the eyes of the law'

Decided on 2 July 2026, Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. is the Supreme Court's most emphatic statement yet on artificial intelligence in the courtroom. A Bench of Justices P.S. Narasimha (who authored the opinion) and Alok Aradhe was confronted with an insolvency admission order that a tribunal had propped up with six "authorities" — most of which did not exist. The Court set aside both the tribunal's order and the appellate affirmation, and used the occasion to lay down a rule of institutional hygiene: an adjudication contaminated by hallucinated material is void, whatever its ultimate correctness on the merits.

The facts

The appellant is a suspended director of Essel Infraprojects Ltd. (EIL), which had stood as corporate guarantor for loans that Pan India Utilities Distribution Company Ltd. (PIUDCL) had taken from Jammu and Kashmir Bank. When PIUDCL defaulted and its accounts were classified as non-performing, the Bank invoked the guarantee and filed an application under Section 7 of the Insolvency and Bankruptcy Code, 2016, before the NCLT, Mumbai, seeking to initiate the corporate insolvency resolution process against EIL as corporate debtor.

By order dated 28 August 2024, the NCLT found debt and default, admitted the application, appointed an interim resolution professional and declared a moratorium under Section 14 of the Code. The suspended director's appeal to the NCLAT — which argued that a demerger and amalgamation had shifted EIL's liabilities and that a renewed sanction letter of 18 November 2017 had relinquished the guarantee — was dismissed on 11 September 2025.

The AI problem

Before the Supreme Court, senior counsel Ms. Madhavi Divan pointed to something the appellate tribunal had missed: the citations on which the NCLT had relied were fake and non-existent, and even where a citation was accurate, the paragraphs excerpted from it could not be traced to the reported judgment. The tribunal, in short, had reasoned from precedents that a machine appeared to have invented.

The Court undertook an independent examination of the six authorities the NCLT had marshalled and set out the results with unusual precision:

• ICICI Bank Ltd. v. Urban Infrastructure Real Estate Ltd., (2019) 16 SCC 528; V.S. Dempo & Co. Ltd. v. Reliance Communications Ltd., (2021) 10 SCC 176; and Sarbjit Singh v. Union Bank of India, (2022) 7 SCC 464 — non-existent citations;
• State Bank of India v. M/s Shree Ram Urban Infrastructure Ltd., 2020 SCC OnLine SC 341 — a wrong citation of an existing reported judgment (the Court noted in a footnote that the correct cause title for that citation is M. Subramaniam v. S. Janaki, (2020) 16 SCC 728; 2020 SCC OnLine SC 341), coupled with a non-existent paragraph;
• Everest Kento Cylinders Ltd. v. Union of India, (2015) 2 SCC 1, and Canara Bank v. N.G. Subbaraya Setty & Anr., (2018) 16 SCC 228 — correct citations but with non-existent paragraphs attributed to them.

Crucially, the responsibility did not lie with counsel. Respondent bank filed an affidavit stating that the alleged judgments had not been cited by its counsel at the bar, and that the tribunal had obtained these "precedents" through its own research. The fabricated authorities then escaped the scrutiny of the first statutory appellate tribunal altogether. As the Court observed, today's courts implicitly trust lawyers on the precedents they cite — a trust that becomes untenable if judges must independently verify the authenticity of every judgment placed before them.

What the Court held

The consequence, the Court held, is categorical. A judicial process tainted by material that is "unreal, fake, and do not exist at all" cannot stand: a decision based on fake and hallucinated material is no decision at all, and amounts to a subversion of the rule of law.

The rule is strict by design. Such decisions are to be set aside "even if an iota of fake or hallucinated material enters the decision-making process," because anything less would compromise the sanctity of adjudication. The Court declared zero tolerance for both the Bar and the Bench in citing, referring to, or relying on such material: it is misconduct for an advocate to cite AI-generated judgments without verification, and a serious lapse for a judge to rely on them. The Court was careful to add that its judgment casts no shadow on the legitimate use of AI — only on presenting or relying on fabricated material as if it were a court precedent.

The reasoning

Justice Narasimha framed the danger in vivid terms. Dependency on technology, he wrote, has never troubled the dispensation of justice, because courts have historically absorbed new tools into their systems. AI is different, and transformative, because it is not merely an aid to work but "an alternative to our own thinking, reasoning and even decision making." The by-product of AI — the production of fake, non-existent and hallucinated material used as precedent — was likened to the release of methyl isocyanate "in the province of law and justice: invisible, insidious, and catastrophic by the time anyone notices."

The remedy for this, the Court accepted, is not judicial declaration alone but public policy and enforceable rules, a process it noted was already under way. The judgment situates the concern within a wider international moment: it records that the UK's Solicitors Regulation Authority approved the first purely AI-driven law firm, Garfield Law, in 2025 — after satisfying itself that the firm managed the risk of "AI hallucinations" and would not itself propose case law — and it references Anthony Malcolm Cork & Anor v. Mark Smith, [2026] EWHC 1199 (Ch), where a fabricated statutory quote generated by a firm's internal AI slipped past supervising partners before the High Court noticed it.

Recognising that a prohibition without accountability is toothless, the Court directed the Bar Council of India, as the apex statutory body, to constitute a committee to deliberate on advocates submitting fake and hallucinated material as precedents, and to prescribe both guiding principles to prevent such occurrences and the disciplinary action that will follow a violation.

The outcome on the merits

Having found the process irretrievably tainted, the Court set aside both the NCLT order of 28 August 2024 and the NCLAT order of 11 September 2025, and restored the Section 7 application (RCP (IB) 6/MB/2023) to its original number. Importantly, it expressed no opinion on the merits of the guarantee, demerger or default dispute — those are for the adjudicating authority to decide afresh. The NCLT was directed to dispose of the restored application expeditiously, preferably within two weeks, with the parties to maintain status quo in the meantime.

Why it matters

Pooja Ramesh Singh converts a growing anxiety into a hard rule of law. The principle is stark: contamination by hallucinated authority is not a curable irregularity to be weighed against the correctness of the result — it is a defect that voids the decision outright. That the fabricated citations here came from the tribunal's own "research," not from an errant advocate, makes the point sharper still, because it shows the failure can originate at the Bench and survive appellate review undetected.

For practitioners, the operational takeaway is unambiguous: every citation — whether surfaced by a research tool, a generative model, or a junior — must be pulled up in the reporter and read before it goes into a pleading or an order. The judgment sits alongside the Kerala High Court's parallel warning in Woodland House School v. Shakeel Ahmad Malik that AI may assist legal research but cannot replace judicial verification, and it is a reminder — visible in the Section 7 lineage from Elegna v. Edelweiss to Catalyst Trusteeship v. Ecstasy Realty — that the integrity of the process is as much a part of insolvency adjudication as the finding of debt and default itself.

Sources

• Supreme Court of India, Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. & Anr., 2026 INSC 668 (2 July 2026)