AI-Hallucinated Citations in Indian Courts and Tribunals: Incidents and Responses, 2023-2026

This entry tracks incidents in which authorities, quotations or legal propositions that cannot be verified appeared in an Indian court or tribunal record and where AI generation was admitted, found or strongly suspected. The Supreme Court uses "hallucinated" for such material in Pooja Ramesh Singh, which describes "fake, non-existent, and hallucinated material" being used as precedent and says it is concerned with the legal consequences, not the technical cause.¹ The draft Regulations for Use of Artificial Intelligence in Courts, 2026 define hallucination as plausible-looking outputs that are factually incorrect, fabricated or unsupported, including erroneous legal facts and case law.

Incidents in which judges merely consulted a chatbot, such as the 2023 Punjab and Haryana High Court bail order and the 2024 Tripura and Punjab and Haryana High Court judgments, are treated in a separate entry on that topic; they involved disclosed use of ChatGPT rather than fabricated authority. The overview also leaves out cases where a court relied on a non-existent statute without any suggestion of AI, such as a Delhi family court judgment citing a non-existent section of the Special Marriage Act, and the Bombay High Court's Rs 5 lakh costs on the State in June 2026 for relying on a non-existent Act in a soil-excavation dispute.

The following table lists the principal reported incidents in date order. "Response" is the immediate judicial response, not the later regulatory reaction.

Other episodes were reported with less detail. In February 2026 the Supreme Court was reported to have flagged an "alarming" trend of lawyers using AI to draft petitions after cases in which cited judgments and quotations proved non-existent or inaccurate, and in March 2026 a Ludhiana hosiery trader's PIL on the PM CARES Fund was reported to have been dismissed after he admitted using AI tools to draft it.

Who produced the fake material. In the adjudicator incidents, the likely source was the decision-maker's own research. In Pooja Ramesh Singh the bank swore that its counsel had not cited the NCLT's authorities and that they came from the NCLT's own research, according to the bank's affidavit as recorded by the Court.¹ In Faiz Enterprise the officer admitted drafting with AI.¹¹ In filer cases, the other side usually discovered the problem.

How it was caught. Opposing counsel (Greenopolis), the affected litigant (Pooja Ramesh Singh, Gummadi, Gadiya, Faiz Enterprise) or a reviewing court (Karnataka, Akasa) raised the issue; the sources reviewed do not say who first noticed the Buckeye Trust errors. The Supreme Court observed in July 2026 that courts and tribunals "implicitly trust lawyers when referring to precedents" and that it would be impractical to have courts verify every cited judgment.¹

What was fabricated. Reports describe invented case names with plausible citations, real citations with invented paragraphs, and real cases cited for propositions they do not contain. The Supreme Court's tables in Pooja Ramesh Singh and its findings in Gadiya show all three.¹²

Attribution to AI. In several incidents the link to AI is inferred. The Greenopolis order says only that precedents "do not even exist", while the attribution to ChatGPT comes from press reports,⁵⁶ and the Karnataka order does not mention AI.⁷ In Gummadi the Andhra Pradesh High Court "realized that the judgments are Artificial Intelligence (AI) generated", and in Akasa the Delhi High Court spoke of an "impression" that AI software may have been used.³¹⁰

The courts' reactions fall into three stages.

Procedural responses (2025). The ITAT recalled its order and the Karnataka High Court restrained the author-member from rehearing the case; the Delhi High Court allowed the petition to be withdrawn; a Karnataka High Court judge asked the Chief Justice to consider action against a civil judge for citing decisions "never decided by the Apex Court or any other Court".⁴⁵⁷

Supreme Court escalation (February to September 2026). On 27 February 2026, in Gummadi Usha Rani, a bench of Justices P.S. Narasimha and Alok Aradhe declared that a decision based on non-existent AI-generated judgments "is not an error in the decision making. It would be a misconduct and legal consequence shall follow", barred the trial court from proceeding on the Commissioner's report, appointed Senior Advocate Shyam Divan as amicus and sought responses from the Attorney General, Solicitor General and Bar Council of India.³ On 2 July 2026 the same Bench in Pooja Ramesh Singh set aside the NCLT and NCLAT rulings, held that such a decision is "no decision in the eyes of the law", called unverified citation "a misconduct on the part of an advocate", and directed the Bar Council to constitute a committee and prescribe norms and disciplinary action.¹ On 2 September 2026 Justices Dipankar Datta and Sheel Nagu applied it to the customs penalty in Gadiya, saying that AI "may well serve as training wheels but entrusting it with the pilot's seat would be both imprudent and dangerous".²

Rule-making. Kerala issued a policy in July 2025 barring AI from findings and judgments; Gujarat followed in April 2026 with personal liability for signatories, and the Punjab and Haryana High Court told district judges not to use AI tools for judgments or research. After Faiz Enterprise the Gujarat High Court made State tax instructions on verifying AI-assisted material enforceable by contempt.¹¹ On 3 June 2026 the Supreme Court's AI Committee published draft regulations that would require disclosure of AI-assisted filings and bar litigants from using AI as a defence for fabricated material. Each of these has its own entry in this wiki.

Responsibility for lawyers rests with the Bar Council of India and State Bar Councils. In June 2026 Bar & Bench reported what the BCI told Parliament: advocates who file AI-generated fake citations, fabricated quotations or non-existent judgments "cannot escape professional liability by blaming technology", and must verify all AI-generated legal research.¹² The Supreme Court's direction of 2 July 2026 asks the Council to turn this into a guiding principle with disciplinary consequences.¹ The Supreme Court Advocates-on-Record Association has proposed, in response to the draft regulations, that advocates certify personal verification of citations in the SLP certificate instead of a general AI-disclosure rule.

The sources reviewed did not show that the BCI had published its norms or that any advocate had been disciplined for AI-fabricated citations by October 2026. The Greenopolis respondents said they would take "appropriate steps", but no outcome is recorded.⁵

The cases show that the risk is not limited to junior lawyers. It reaches judges, tribunal members and revenue officers who decide large sums, and the Supreme Court has said that a single "iota" of fake material can vitiate a decision.¹ That standard, and the direction to set decisions aside, shifts costs onto litigants, as the Buckeye Trust assessee and the Essel Infraprojects appellant had to litigate for months to get fresh decisions.

Several questions remain open. The sources reviewed show no official register of incidents, so prevalence cannot be measured, and reports depend on parties spotting errors. The line between sanctionable AI use and an honest error is unsettled: the Karnataka civil judge incident and the Delhi family court case involved non-existent authority without proof of AI. The Supreme Court's draft regulations contemplate officers recording reasons to dispense with verification of AI output in some cases, which sits uneasily with the zero-tolerance line.

The adjudicator incidents also raise a due-process question, as the Buckeye Trust recall illustrates: when a flawed order is withdrawn, who decides the case again, and how the integrity of the process is protected.

As of 15 October 2026 the Supreme Court's inquiry in Gummadi Usha Rani remained pending, with the IT and Law ministries impleaded and the last recorded listing on 1 September 2026. The Bar Council of India's committee and norms, and the final Regulations for Use of Artificial Intelligence in Courts, had not been published in the sources reviewed. Reported incidents became more frequent through 2026, and this entry should be updated as the Supreme Court and High Courts decide further cases.