By: Chakradhar Reddy
Six Supreme Court judgments. That is what the National Company Law Tribunal (NCLT) relied on in August 2024 when it admitted an insolvency application filed by Jammu & Kashmir Bank Ltd. (the Bank) against Essel Infraprojects Ltd. (Essel).[1] On paper it looked like careful work: settled authorities, neatly cited, all pointing the same way.
Then the Supreme Court checked them. Three of the six did not exist at all: no such case, no such citation. The other three were genuine judgments, misdescribed: passages attributed to them appear nowhere in their text, and in some instances the citation or the case name was wrong. Not one of the six held up.
Nobody seems to have slipped these into the record on purpose. According to the Bank’s affidavit, its counsel never cited them; they appear instead to have entered the record through the Tribunal’s own research. On appeal, the National Company Law Appellate Tribunal (NCLAT) reproduced all six in a single paragraph, apparently without verifying them.[2] Two forums, one imaginary body of law.
How a machine invents a case
Anyone who has used a generative AI tool has seen the trick. Ask for a precedent and you get one: a court, a year, a bench, a confident paragraph. The trouble lies in how these systems work. They generate text by predicting the most likely next word from patterns in what they have read. Unless connected to a reliable source, they do not look anything up, and even tools that do retrieve sources can misstate or misattribute what they find. A citation in the familiar “(year) volume SCC page” format looks right because it resembles thousands of real ones. Researchers call the result hallucination. It is a well-documented limitation of the technology, not an occasional glitch.[3]
What makes it dangerous is the tone. The output sounds exactly the same whether it is true or invented. It fails in the voice of competence.
No decision in the eyes of the law
On 2 July 2026, in Pooja Ramesh Singh v. Jammu & Kashmir Bank Ltd., the Supreme Court set aside both the NCLT and NCLAT orders, and it did so for a single reason: the fabricated authorities.[4] The language was blunt. A decision resting on hallucinated material, the Court held, is “no decision in the eyes of the law”, and it makes no difference whether the fake material had a direct or indirect bearing on the outcome.
That second part is what makes the case so striking. The result might well have survived without the phantom judgments. The debt and default do not appear to have been seriously in dispute. Essel’s principal defence, that its liability had passed to another company through a demerger and amalgamation, had been rejected by the NCLAT on an independent ground: Clause 8 of the guarantee deed provided that the guarantee would survive any such absorption.[5] That is a reasoned, fact-based answer standing on its own feet. The Supreme Court nonetheless set the orders aside, expressing no view on the merits.
The Court also directed the Bar Council of India to constitute a committee to frame guiding principles to prevent a repeat, with disciplinary consequences for violations, and restored the application to the NCLT for fresh adjudication, preferably within two weeks.
What “zero tolerance” actually means
Commentators have described the approach as a zero-tolerance rule, and the label is apt. It is a bright line. If a decision relies on hallucinated, AI-generated material, it is set aside. There is no inquiry into how much weight the fake authority carried and no room to save an order because the rest of the reasoning looks sound. On the Court’s formulation, that consequence follows even if an “iota” of fake or hallucinated material enters the decision-making process.[6]
Two ideas sit behind the rule. The first is integrity. Fabricated authorities, the Court said, undermine the integrity of the adjudication and its processes, which is why it set the orders aside without asking whether the fake citations changed the result. The second is trust. Courts and tribunals ordinarily rely on the precedents that lawyers place before them, and the Court asked what it would mean for a judge to verify every judgment cited.[7] A rule this strict tells everyone in the chain that checking must happen before the order is signed, not after it is challenged.
It also reaches beyond lawyers. Here the fake authorities appear to have come from the Tribunal’s own research, not from counsel, so the burden of verification sits with the decision-maker, and with those who assist in research, as much as with the Bar.
A rule, or a standard?
Not everyone agrees that a bright line is the right tool. Earlier in 2026, the Andhra Pradesh High Court had taken a different route. A fake citation, it held, does not by itself vitiate an order if the law stated in the order, and its application, are correct; it becomes fatal only where the principle applied is not the law of the land.[8] In June 2026, the High Court of Jammu & Kashmir and Ladakh reasoned similarly, declining to interfere on that ground alone, while insisting that precedents suggested by an AI tool must be independently verified from authentic sources before a judge relies on them.[9] The Andhra Pradesh decision is itself under challenge. In March 2026, the same Bench that later decided Pooja Ramesh Singh issued notice in that challenge, observing that a decision based on non-existent judgments is not an error in decision-making but would be “misconduct”, and that “legal consequence shall follow”.
Because the Supreme Court’s ruling binds all courts and tribunals under Article 141 of the Constitution, the more forgiving approach of the two High Courts can no longer safely be relied upon where a decision actually rests on hallucinated material. Some questions nonetheless remain open: whether the rule applies with the same rigour where a fake authority is merely recorded in a party’s submissions rather than relied upon, and what accountability follows for the individual decision-maker. The pending Andhra Pradesh challenge, in which the Court has sought the assistance of the Attorney General, the Solicitor General, the Bar Council of India and an amicus curiae, may supply some of the answers.
The cost of the bright-line approach is visible in this very case. The Bank’s application, admitted in 2024, has to be heard again, so the delay falls not on whoever let the hallucinations through but on the litigants whose dispute is still waiting. For an insolvency regime built around strict timelines, that is an uncomfortable result.
And yet the Court’s instinct is easy to understand. If the people who decide cases cannot be trusted to check what they cite, it is the court’s authority that suffers. A rule without exceptions is also a loud message, and judges, clerks and lawyers will all hear it. Both sides are protecting something real: public confidence in adjudication on one hand, the litigant’s time on the other. Each concern is legitimate, but where the two collide, public confidence is the harder to rebuild.
Not only India’s problem
Courts elsewhere confronted the issue earlier, but largely in relation to lawyers rather than judgments. In Mata v. Avianca, a US federal court in New York imposed a US$5,000 penalty on the plaintiff’s attorneys and their firm. Its finding of bad faith rested not merely on the citation of AI-generated cases, but on the attorneys continuing to stand by them after their existence was questioned.[10] In England, a Divisional Court of the High Court in Ayinde warned that placing false material before a court with the intention that it be treated as genuine may, depending on the person’s state of knowledge, amount to contempt of court and, in the most egregious cases, the criminal offence of perverting the course of justice.[11] Those courts focused on the individuals responsible. India’s Supreme Court went further and set aside the decision itself, while also directing the Bar Council towards individual accountability.
None of this means that courts want AI kept out. The Draft Regulations for Use of Artificial Intelligence in Courts, 2026, released for public consultation by the Supreme Court’s AI Committee in June 2026, treat legal research, precedent retrieval, citation verification and summarisation as permissible uses, and call for structured training for judges, advocates and court staff.[12] The draft would also require parties and their advocates to declare any AI-assisted preparation of pleadings, documents or evidence. The Kerala High Court has gone further for its district judiciary, prohibiting the use of AI tools to arrive at findings, grant relief or draft orders and judgments.[13] The message is consistent: use the tool, but do not hand over the checking.
What this means for practice
For lawyers, the lesson is dull and non-negotiable. Pull the judgment from an authoritative database. Read the paragraph you are citing in the judgment itself, not in a summary and not in a chatbot’s version of it. Confirm that the case says what your draft says it says, and that it remains good law. AI is useful for a first pass or for finding leads, much like a keen junior who has never once said “I’m not sure”. You would not file that junior’s work unread, and the same goes here. It is also prudent to keep a short record of how each authority was verified, to have a written policy on which tools may be used and for what, and to be open with clients about where AI is used in their matter. Placing a fabricated authority before a court is not a technical slip: it may expose an advocate to disciplinary proceedings for professional misconduct under Section 35 of the Advocates Act, 1961, in addition to whatever the Bar Council’s forthcoming principles prescribe. The Essel case will probably be remembered less for insolvency law than for what it shows about shortcuts. Six unverified judgements , three of which never existed, cost two forums their orders and sent the Bank back to the start, nearly two years after its application was admitted. Checking them would have taken far less.
[1]Pooja Ramesh Singh v. Jammu & Kashmir Bank Ltd., (2026) 268 Comp Cas 709 : 2026 SCC OnLine SC 1258 : 2026 INSC 668 (Civil Appeal No. 11950 of 2025, decided 2 July 2026, P.S. Narasimha and Alok Aradhe JJ). The NCLT order is Jammu & Kashmir Bank Ltd. v. Essel Infraprojects Ltd., 2024 SCC OnLine NCLT 14718 (28 August 2024).
[2]Pooja Ramesh Singh v. Jammu & Kashmir Bank Ltd., 2025 SCC OnLine NCLAT 1464 (11 September 2025).
[3]Ziwei Ji et al., “Survey of Hallucination in Natural Language Generation” (2023) 55(12) ACM Comput. Surv., Article No. 248.
[4]Pooja Ramesh Singh (n 1).
[5]NCLAT order (n 2); see also Pooja Ramesh Singh (n 1).
[6]Pooja Ramesh Singh (n 1).
[7]Pooja Ramesh Singh (n 1), para 16.
[8]Gummadi Usha Rani v. Sure Mallikarjuna Rao, (2026) 1 ITCC 1 : 2026 SCC OnLine AP 194. The Supreme Court’s order issuing notice in the SLP against this decision (SLP (C) No. 7575 of 2026), including its observation on “misconduct”, is reported at 2026 SCC OnLine SC 341.
[9]Woodland House School v. Shakeel Ahmad Malik, 2026 SCC OnLine J&K 421 (June 2026).
[10]Mata v. Avianca, Inc., 678 F Supp 3d 443 (SDNY 2023).
[11]R (Ayinde) v. London Borough of Haringey, [2025] 1 WLR 5147 : [2025] EWHC 1383 (Admin) (Divisional Court).
[12]Artificial Intelligence Committee, Supreme Court of India, Draft Regulations for Use of Artificial Intelligence (AI) in Courts, 2026 (published for public consultation on 3 June 2026), Regns 17(1), 19(1)(d), 43(3), 49.
[13]High Court of Kerala, Policy Regarding Use of Artificial Intelligence Tools in District Judiciary (19 July 2025).
