India's Supreme Court draws the line for AI in courtrooms
Setting aside two company law tribunal orders built on AI invented citations, the Court told lawyers to verify, disclose, and bear personal liability for unverified AI output.
Setting aside two company law tribunal orders built on AI invented citations, the Court told lawyers to verify, disclose, and bear personal liability for unverified AI output.
In an Essel Group-linked insolvency case, the Supreme Court of India set aside orders passed by two lower tribunals on July 2 after finding they rested on case citations that did not exist. The precedents had been produced by an AI tool and never checked. The bench, in Pooja Ramesh Singh v Jammu and Kashmir Bank Ltd., 2026 INSC 668, then used the order to write the rule for AI in courtrooms: every citation must be verified, AI assistance must be disclosed, and the lawyer who signs the filing vouches for the output.
The two tribunals are the National Company Law Tribunal (NCLT), India's company-law court of first instance, and the National Company Law Appellate Tribunal (NCLAT), the body that hears appeals from it. Both had cited fabricated judgments in rulings in the Essel Group's ongoing corporate resolution, an Indian media and infrastructure conglomerate. The Supreme Court's order set their conclusions aside. The bench treated the issue as zero-tolerance: hallucinated cases, the AI term for confident output the model invents because it does not exist, are not clerical slip-ups. They are counterfeit legal authority, on the same plane as a forged document. The Daily Pioneer captured the framing in one line: an invented precedent is not a typo, it is counterfeit currency of the law.
For AI-assisted work in courtrooms, three rules now apply. First, every AI-generated citation has to be checked against a primary source before filing: case number, court, date, the standardized neutral citation that identifies the judgment, and the reporter or database where it can be read. Second, lawyers must disclose when AI has materially helped draft a pleading, a written submission, or a case compilation. Third, when the AI is wrong and the lawyer did not check, the lawyer is personally liable. Not the tool. Not the platform. The Indian Lawyer's analysis reaches the same conclusion. The order is the immediate cause; the Bar Council of India has been directed to frame binding professional norms for AI use in filings and research.
The order is also, on its face, pro-AI. India's judiciary is heavily overburdened, with case backlogs that no human bench can clear without help. The Court accepted AI as a courtroom tool. It drew the line between AI as a research assistant, useful for transcription, translation, cause-list management, case clustering, legal research, summarising pleadings, and contradiction detection, and AI as a substitute for the human act of verification. The Hindustan Times analysis lists the same constructive use cases.
Two practical questions follow. The first is what counts as a 'primary source' in 2026. Verdictum, Indian Kanoon, the Supreme Court's own e-resources, and the IBBI's public order portal, which hosts the underlying record for the Essel-matter corporate entity in this case, are the obvious ones. The harder question is what 'materially helped draft' means when a model has read the entire record and proposed a structure the lawyer then rephrases. The Bar Council draft will have to answer that, and lawyers will be the ones testing it.
The second question is portable. The ANI wire report and the India Today summary both carry the same warning to anyone whose work produces authoritative output: medicine, audit, compliance, financial advice, journalism. Borrow a model, and you take on the verification. The Supreme Court of India has now written that down in an enforceable form, and the Bar Council's draft rules are the next thing to watch.