ACCOUNTABILITY BETWEEN BAR AND BENCH FOR AI-HALLUCINATED PRECEDENTS

 

On 2 July 2026, in Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. (2026 INSC 668), the Supreme Court of India set aside the orders of the National Company Law Tribunal (NCLT) and the National Company Law Appellate Tribunal (NCLAT), both of which had admitted an application filed under Section 7 of the Insolvency and Bankruptcy Code, 2016, against Essel Infraprojects Ltd., a corporate guarantor for the defaulted loans of Pan India Utilities Distribution Company Ltd. The NCLT had admitted the application, and the NCLAT had upheld that decision, with both tribunals relying on six cited precedents in doing so.

When the Supreme Court of India examined those six citations on its own, it found something troubling. Three of the precedents did not exist at all, while two others were genuine judgments whose cited paragraphs could not be found anywhere in the reported decisions. The Supreme Court has since remanded the matter to the NCLT for fresh disposal, while making clear that it expresses no opinion on the merits of the underlying insolvency proceeding.

The concern raised by this judgment extends well beyond a mere citation error. Judgments that rest on case law which was never law at all do not simply damage confidence in the judicial process; they impose real costs, in both time and money, on the litigants who are made to bear them.

The Supreme Court of India has rightly treated the use of fabricated precedents as a serious matter. Yet, in seeking to prevent a recurrence, the Court has taken a path that is easier than it is fair, since it places the entire burden of prevention on advocates without holding the judicial process to the same standard it now demands of the Bar.

This distinction matters because, on the Court's own findings, the fabrication did not originate with the advocates at all. Paragraph 1 of the judgment records that the Court was dealing with a tribunal relying on such material "yet again," a phrase that points to a recurring pattern rather than a single lapse. Paragraph 16 goes further, recording that the precedents relied upon were "obtained through its research" and had not been submitted by counsel, and that the fabricated judgments had "escaped scrutiny by the statutory appellate tribunal." The fabrication, in other words, originated in the NCLT's own research process and passed undetected through the NCLAT's appellate review as well.

Yet the consequences that follow from this finding run in only one direction. An advocate who cites material that does not exist is guilty of professional misconduct and may face disciplinary action, and the Bar Council of India has accordingly been directed to examine the conduct of the advocates involved. No corresponding direction has been issued in respect of the tribunals, and no inquiry has been ordered into the research process that produced the fabricated citations in the first place.

It might be argued that this asymmetry is simply a matter of institutional structure, since judicial accountability operates through separate channels, such as contempt jurisdiction, and a civil appeal may not be the appropriate forum in which to impose consequences on a tribunal member. That explanation, however, sits uneasily with the Court's own language. The judgment describes the tribunal's reliance on the fabricated material as "a serious lapse," a description that itself invites scrutiny of that conduct, yet the judgment stops short of naming any mechanism through which such scrutiny might be carried out, even as it simultaneously directs the Bar Council of India to act against the advocates concerned. That is not institutional restraint; it is a choice, and it shows precisely where the Supreme Court of India was prepared to assign consequences, and where it was not.

A useful point of comparison is the position in England and Wales, where judicial guidance on the use of artificial intelligence was updated in October 2025. That guidance is addressed not to advocates but to judges themselves, and it extends as well to clerks, legal advisers and other members of the judicial support staff. It proceeds on the understanding that an AI-generated error is a risk borne by the judicial process as a whole, rather than one to be managed solely through the discipline of counsel. The judgment in Pooja Ramesh Singh, by contrast, addresses only the Bar, leaving unexamined the very process, within the NCLT itself, in which the error actually arose.

If this gap is not closed, the present case is unlikely to be the last of its kind, and litigants in India will continue to lose time and money to fabricated case law with no corresponding safeguard on the judicial side. The direction issued to the Bar Council of India in paragraph 9 of the judgment is a necessary step, but it cannot substitute for action on the other side of the courtroom. The Supreme Court of India, or the High Courts in the exercise of their supervisory jurisdiction, should accordingly direct every tribunal and appellate body within the country to build a verification function into its own processes, whether through existing law clerks or through a role created specifically for that purpose, so that cited authorities are verified before they are incorporated into a judgment. Such a step would mirror the safeguard that the guidance in England and Wales now provides through its clerks and legal advisers, and without it, the judiciary risks repeating the very error it has just condemned.

On the next occasion, there may be no advocate left to blame, and until the Bench is held to the same standard of verification that the Supreme Court of India has now set for the Bar, the "zero tolerance" it has declared will govern only half the courtroom.

 Share