The Supreme Court’s recent decisions on AI-generated legal authorities raise a fundamental question for the legal profession: when technology produces a judgment that does not exist, who bears responsibility for allowing it to enter the courtroom?
Introduction: When a Citation Looks Real but the Law Is Not
Imagine a legal proposition that appears perfectly sound, supported by a Supreme Court judgment, complete with a case name, citation and paragraph number. An advocate places it in a written submission, or an adjudicating authority relies upon it while imposing a substantial penalty. The proposition is persuasive, the reference looks authentic, and the document passes through the ordinary process of legal scrutiny. There is only one problem: the cited judgment does not exist, or the judgment exists but says nothing of the kind attributed to it. This is no longer merely a hypothetical concern. Artificial intelligence tools can assist lawyers in locating authorities, summarizing lengthy judgments and organizing legal arguments. Yet these tools can also produce convincing but fabricated citations, invent judicial observations or attribute an entirely incorrect legal principle to an existing decision. The danger lies not simply in technological inaccuracy, but in the possibility that fabricated law may influence a real judicial or quasi-judicial determination.
Recent decisions of the Supreme Court of India have brought this concern into sharp focus. In Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd., decided on 2 July 2026, the Court addressed judicial orders resting on AI-generated fake and hallucinated precedents. Two months later, in Vijay Ghanshyam Gadiya v. Union of India, the Court set aside a customs penalty order after finding that it relied on non-existent authorities, fake citations and legal propositions unsupported by the cases cited.
Read together, these decisions offer an important lesson. Artificial intelligence may assist legal work, but it cannot replace the independent verification, professional judgment and accountability on which the administration of justice depends.
The September 2026 Decision: A ₹425-Crore Customs Penalty Set Aside
In Vijay Ghanshyam Gadiya v. Union of India, 2026 SCC OnLine SC 1741, decided on 2 September 2026, the Supreme Court considered a challenge arising from an Order-in-Original dated 8 October 2025. The Additional Commissioner of Customs, Surat, had imposed a penalty of approximately ₹425.27 crore under Section 114 of the Customs Act, 1962, in connection with the alleged misdeclaration of natural diamonds as lab-grown diamonds to secure a lower tariff.
The challenge to the penalty had been dismissed by the Gujarat High Court on 20 January 2026. Before the Supreme Court, however, the appellant questioned the reliance placed in the original adjudication order on judgments and articles allegedly generated using artificial intelligence. The Court undertook an independent examination of the material. It found that the authorities relied upon included cases that were either non-existent or supported by fake citations. Further examination revealed that certain cases did exist, but did not lay down the legal propositions attributed to them. The Court considered these to be instances of AI hallucination. The significance of the decision lies in what followed. The Supreme Court held that reliance on such dubious material was fatal to the sustainability of the penalty order. It set aside both the Gujarat High Court’s order and the original customs order, restoring the proceedings for fresh adjudication by an officer of the same rank other than the officer who had passed the original order.
The Court did not merely identify an error in legal research. It found that the use of fabricated or misrepresented legal authorities had compromised the adjudicatory process itself. At the same time, the Court left it to the appointing authority to consider, in accordance with law, whether action against the officer who authored the original order was warranted. The distinction is important. The immediate consequence established in this case was the setting aside of the impugned decisions and a direction for fresh adjudication. The Court did not hold that a party had lost its case because its advocate had used AI. Rather, the decision concerned an adjudicating authority’s reliance on dubious legal material in an order imposing a substantial penalty.
The Earlier Warning: Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd.
The September decision gains additional significance when considered alongside Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd., 2026 INSC 668, decided on 2 July 2026.
That case arose from insolvency proceedings involving Essel Infraprojects Ltd. The National Company Law Tribunal had admitted an application under Section 7 of the Insolvency and Bankruptcy Code, 2016, and the National Company Law Appellate Tribunal subsequently dismissed the challenge. In the course of its reasoning, the NCLT relied on six judgments that were reproduced in the appellate proceedings. Before the Supreme Court, it was pointed out that the cited authorities appeared to be fake or non-existent and that some paragraphs attributed to genuine judgments could not be found in the original decisions. The Supreme Court independently examined the citations. It found that three citations were wholly non-existent, while others were genuine references accompanied by paragraphs that did not exist in the judgments cited.
Crucially, the Court recorded that the bank had filed an affidavit stating that its counsel had not cited those judgments before the tribunal. The fake authorities appeared to have entered the decision-making process through the tribunal’s own research. The citations had also escaped scrutiny at the appellate stage. The Court set aside the orders of both the NCLT and the NCLAT and restored the insolvency application to the NCLT for fresh disposal on merits. It adopted a zero-tolerance approach towards the production, citation or use of AI-generated precedents without verification. It also directed the Bar Council of India to constitute a committee to consider appropriate guiding principles and disciplinary consequences for advocates who submit fake or hallucinated material as precedent.
This decision is particularly instructive because it demonstrates that the problem cannot be reduced to an advocate’s misuse of a chat bot. Fabricated legal material can enter adjudication through several routes, including research conducted by a tribunal or court itself. Responsibility for verification therefore cannot be confined to one participant in the litigation process.
Does an AI-Generated Citation Automatically Defeat a Case?
The title of this article raises a practical concern: can an AI-generated citation cost a party its case? The answer requires a distinction between the risk posed by fabricated authorities and the precise consequences established by the judgments.
In Pooja Ramesh Singh, the Supreme Court set aside the NCLT and NCLAT orders because the judicial process had been tainted by fake and hallucinated precedents. In Vijay Ghanshyam Gadiya, the Court set aside the customs penalty order and the High Court’s order affirming it because the adjudicatory decision relied upon dubious legal material.
Neither decision should be inaccurately described as establishing that every use of AI in legal research invalidates a proceeding, or that every error in a citation automatically results in dismissal of a party’s case. The central concern is the introduction of fictitious or materially misrepresented authorities into the process of adjudication. The consequences may nevertheless be serious. Depending on the circumstances, fabricated authorities may undermine the credibility of submissions, distract the court from the real issues, necessitate further hearings, contribute to an adverse order or expose a professional to disciplinary scrutiny. Where the fabricated material has entered the reasoning of a decision, the resulting order may face a challenge on the basis recognized in these Supreme Court decisions.
The precise legal consequence will depend on the facts, the role played by the fabricated material and the applicable law. The decisions do, however, make clear that fake precedents cannot be treated as harmless drafting mistakes merely because they were generated by software.
The Duty of Counsel: Verification Cannot Be Outsourced
For advocates, the most immediate lesson is that the use of AI does not dilute the duty to the court. A lawyer remains responsible for the authorities cited in a pleading, written submission, opinion or oral argument. The fact that a citation was produced by an AI tool does not establish that the case exists, that it remains good law or that it supports the proposition for which it is cited.
Legal research requires more than identifying a case name and a plausible-looking citation. The cited judgment must be located in a reliable source and read in its relevant context. The court, bench, date, citation, paragraph numbers and actual ratio must be checked. Counsel must also consider whether the decision has been reversed, distinguished, overruled or affected by a later statutory amendment or binding precedent.
The Supreme Court’s observations in Pooja Ramesh Singh are especially significant in this regard. The Court identified the citation of unverified AI-generated judgments by an advocate as professional misconduct. Its direction to the Bar Council of India to consider guiding principles and disciplinary consequences signals that the issue extends beyond the quality of legal drafting to professional accountability.
That does not mean that every inaccurate citation should automatically be treated as deliberate dishonesty. An inadvertent error, a negligent failure to verify and a knowing attempt to mislead may involve materially different circumstances. The appropriate response must follow the applicable legal and disciplinary framework. Nevertheless, describing the error as a technological malfunction cannot, by itself, discharge the professional responsibility to verify a legal authority before relying on it.
Beyond Advocates: The Responsibility of Adjudicating Authorities
The two decisions also carry an institutional message. Pooja Ramesh Singh demonstrates that fabricated authorities may enter a judicial order even when counsel has not cited them. Vijay Ghanshyam Gadiya demonstrates the consequences when dubious authorities form part of an administrative adjudication order.
The duty to ensure that a decision rests on genuine law cannot therefore be discharged merely by expecting litigants to identify errors. Courts, tribunals and statutory authorities must retain control over their reasoning and verify the materials on which their decisions are based. A citation appearing in an AI-generated summary or research note is not a substitute for examining the judgment itself. This is particularly important in quasi-judicial proceedings, where administrative authorities may determine substantial financial liabilities, regulatory consequences or individual rights. The use of technology to improve speed and efficiency cannot come at the expense of a reasoned, legally sustainable decision.
The broader institutional question is how to combine technological assistance with human oversight, professional accountability and reliable verification. As AI becomes more accessible, training, internal review procedures and clear standards for the use of generated material will become increasingly important across the legal ecosystem.
Conclusion: Technology May Assist the Law, but It Cannot Authenticate Itself
The Supreme Court’s decisions in Pooja Ramesh Singh and Vijay Ghanshyam Gadiya mark an important stage in the legal system’s response to AI-generated hallucinations. The lesson is not that artificial intelligence has no place in legal practice. It is that the efficiency of a tool cannot replace the discipline of legal verification. A fabricated judgment does not become law because it appears in polished language, carries a plausible citation or is reproduced in an official-looking document.
Ultimately, AI can help locate, organise and explain legal material. It cannot relieve a legal professional or adjudicator of responsibility for what is placed before a court or relied upon in a decision. When fabricated authorities enter the courtroom, the issue is no longer simply whether the technology has made a mistake. It is whether the human institutions entrusted with administering justice have preserved the safeguards necessary to detect that mistake before it affects a party’s rights.
The future of AI-assisted law will depend not on how convincingly a machine can produce a citation, but on how rigorously the legal profession verifies it.
