By Deepabali Datta
For 73 paragraphs, Deepak Raheja vs Omkara Assets Reconstruction Pvt Ltd, 2026 INSC 1134, reads like any other insolvency judgment. It deals with a Rs 600-crore loan facility, two hotel companies operating the JW Marriott in Bengaluru and the Crowne Plaza in Pune, and a dispute over when a default occurs. On October 15, 2026, a Supreme Court bench of Justices Dipankar Datta and Augustine George Masih dismissed the appeals and allowed the corporate insolvency resolution process to proceed.
Then comes paragraph 74.
The Court records that the respondents had objected to the appellants’ compilation of judgments, describing it as “replete with AI generated non-existent judgments”. Asked to explain, the advocate-on-record blamed his clients and claimed ignorance of the compilation’s contents. Senior counsel for the appellants subsequently took responsibility and sought forgiveness on their behalf.
The Court had proposed exemplary costs against the appellants, but dropped them after learning that the first appellant had died. The appeals were nevertheless decided on merits, despite the respondents’ objection that the appellants’ conduct disentitled them from a hearing on the merits.
It is a single paragraph at the end of a lengthy commercial judgment. But its significance extends well beyond the case itself. An advocate-on-record is the lawyer formally responsible for filings before the Supreme Court. When he disclaims knowledge of his own filing over fabricated judgments, the question is no longer about one compilation. It is about where accountability lies when a machine invents the law.
A PROBLEM ACROSS THE SYSTEM
Deepak Raheja was the latest stop in a journey fabricated precedents had made through India’s justice system in 2026.
In February, the problem surfaced in a trial court. In Gummadi Usha Rani vs Sure Mallikarjuna Rao, 2026 SCC OnLine SC 341, a court in Andhra Pradesh, while deciding objections to a commissioner’s report, had relied on four Supreme Court judgments that did not exist. Each carried a plausible title and a precise-looking SCC citation.
On February 27, Justices PS Narasimha and Alok Aradhe of the Supreme Court observed that reliance on such fabricated authorities was not merely an error in decision-making. “It would be a misconduct and legal consequence shall follow,” the Court said.
In July, the problem reached two tribunals. In Pooja Ramesh Singh vs Jammu & Kashmir Bank Ltd, 2026 INSC 668, the same bench found that orders of the National Company Law Tribunal and its Appellate Tribunal relied on authorities that did not exist or contained paragraphs no judge had written.
According to the bank, its counsel had never cited the fabricated authorities; they had entered the proceedings through the tribunals’ own research. The Supreme Court declared zero tolerance “for the Bar as well as the Bench” and held that a decision founded on such material “is no decision in the eyes of the law”. It also asked the Bar Council of India to frame guiding principles and disciplinary consequences.
In September, the problem surfaced in a customs adjudication. In Vijay Ghanshyam Gadiya vs Union of India, 2026 INSC 947, a penalty exceeding Rs 425 crore on a diamond importer rested on judgments that did not exist and genuine judgments cited for propositions they did not support.
The Supreme Court described the material as “a hallucination of AI”. Justices Dipankar Datta and Sheel Nagu quashed the penalty, remitted the matter to another officer of the same rank and left action against the author to the appointing authority.
“AI may well serve as training wheels,” the Court observed, “but entrusting it with the pilot’s seat would be both imprudent and dangerous.”
In October, the problem reached the Supreme Court itself.
The message across all four cases is unmistakable. The problem is no longer confined to what lawyers file. It extends to what courts and adjudicating authorities rely upon in deciding disputes.
HOW A MACHINE INVENTS A JUDGMENT
Generative AI tools can produce plausible legal text without verifying whether the authority exists. Asked for a Supreme Court judgment supporting a particular proposition, they may generate precisely what a genuine citation looks like: a case name, year, volume, page number and even a paragraph of reasoning, regardless of whether the case exists.
That is what makes a hallucinated citation dangerous. It is a counterfeit note printed on the right paper, with the right watermark and the right serial number. It can survive a casual glance and fail only when someone checks it against the original.
In a common law system, that is no small failure. Precedent is the foundation of legal argument, and under Article 141 of the Constitution, the law declared by the Supreme Court binds every court in India.
In Pooja Ramesh Singh, the Court likened the problem to a gas leak: invisible, insidious and potentially catastrophic by the time anyone notices.
THIS IS NOT AN ARGUMENT AGAINST AI
None of this is an argument for abandoning artificial intelligence. Tools that translate, summarise and organise records can assist courts carrying some of the world’s heaviest caseloads. In Gadiya itself, the Court accepted that the use of AI as an assistive tool cannot be denied.
The distinction is between assistance and abdication.
A lawyer may use a tool to locate a case, but must still read and verify it. A judge may use AI to organise a file, but must still decide the dispute through independent reasoning. The failure in each of this year’s cases was not necessarily the use of AI; it was the absence of verification.
WHO ANSWERS?
Deepak Raheja brings the question of accountability into sharper focus. At present, the law offers no single, comprehensive answer.
The first candidate is the lawyer. Every filing before the Supreme Court passes through an advocate-on-record, who is answerable to the Court for its contents. The signature is meant to function as a professional assurance that the material has been examined and verified.
An advocate-on-record who disclaims knowledge of his own filing undermines that assurance. Regulation 43(6) of the Supreme Court’s draft Regulations for Use of Artificial Intelligence (AI) in Courts, 2026, points in the same direction: a person submitting fabricated or inaccurate AI-generated material cannot plead its AI origin as a defence.
The second is the client. Clients may hand their lawyers AI-generated research, some of which may be fabricated. But a client is not an officer of the court and does not sign the filing. In Ayinde vs London Borough of Haringey, [2025] EWHC 1383 (Admin), the High Court of England and Wales made the point plainly: a lawyer cannot rely on a lay client for the accuracy of citations.
The third is the decision-maker. Gummadi Usha Rani and Gadiya demonstrate that judges and adjudicating officers can introduce fabricated authorities themselves, without either party having cited them. The Supreme Court has treated such conduct as potential misconduct, while the draft Regulations place accountability for AI-assisted decisions on the officer who makes them.
What remains missing is a consistent mechanism for investigating such failures and prescribing proportionate consequences.
The regulatory framework itself is unfinished. The draft Regulations, which would also require disclosure of AI-assisted material in pleadings and submissions, have not yet been notified. There is no indication that the Bar Council of India has issued the rules sought by the Court in July. The Kerala and Gujarat High Courts have AI policies, but these govern judges and court staff rather than the Bar.
Until these gaps are addressed, accountability will continue to depend largely on how individual benches respond to individual cases.
Deepak Raheja illustrates the difficulty. The proposed costs were directed against the appellants, but dropped after the first appellant’s death. Yet, if the lawyer is responsible for what he files, imposing costs on litigants could penalise the wrong party. The question of professional responsibility for the fabricated compilation remains unresolved.
A PROBLEM THE WORLD IS FACING
India is not alone. In Mata vs Avianca, 678 F Supp 3d 443 (SDNY 2023), a New York court sanctioned two lawyers and their firm for citing cases invented by ChatGPT.
In Ayinde, the English court warned of wasted costs, regulatory referrals, contempt and, in egregious cases, prosecution.
And in Heart and Soul Entertainment Ltd vs Deepak s/o Shivkumar Bahry, SLP (C) No. 3090 of 2026, the Supreme Court itself observed that the menace is rampant “not only in India rather throughout the world”.
Each new case brings a fresh judicial rebuke, yet the problem continues to recur.
WHAT A RESPONSE COULD LOOK LIKE
Several measures could be considered without discouraging legitimate use of technology.
The declaration contemplated in the draft Regulations could go a step further, requiring advocates to certify that every authority cited has been verified against an authentic source. This would make an existing professional duty explicit and traceable.
The disclosure requirement in draft Regulation 43(3) could be finalised and notified. Disclosure of AI use need not be a prohibition; it could serve as an alert, prompting the court to examine the material more closely where necessary.
Court registries could check citations in compilations against established law reports and legal databases before matters are listed. Such screening could largely be automated using reliable databases rather than generative tools. The AI Content Verification Authority proposed under draft Regulation 44 could support this process.
Judicial academies could give practical effect to the training envisaged in draft Regulation 49 by teaching judges and adjudicating officers how to verify every judgment and authority on which they rely.
The same discipline must extend to administrative decision-makers. The accountability envisaged for judicial officers under the draft Regulations could be supported by clear inquiry procedures. For departmental adjudicators, including customs and tax officers, the applicable safeguards could be clarified and reinforced through binding ministerial guidelines.
Consequences, too, should be graded and predictable: costs for carelessness, disciplinary action for repeated or deliberate misconduct, and a clear rule that whoever signs a filing or adopts an authority in a decision is responsible for verifying it.
THE COST OF A FALSE CITATION
No apology can restore the time a court spends chasing a judgment that does not exist. No disclaimer can repair a litigant’s trust in a decision built on invented law. And no software, however sophisticated, can take an oath to uphold the Constitution.
Some judgments never existed; others were made to say what they never said. The consequences were real.
Preventing a recurrence requires more than another judicial warning. Those who submit or rely on fabricated authorities must answer for doing so. Otherwise, the next judgment that never was may already be finding its way into a courtroom.
The post The Judgment That Never Was appeared first on India Legal.
