A ₹425-Crore Order Built on Cases That Don't Exist
On September 2, 2026, the Supreme Court of India set aside a ₹425.28 crore ($51 million) customs penalty in Vijay Ghanshyam Gadiya v. Union of India and Another (2026 LiveLaw (SC) 884), after finding that the adjudicating Additional Commissioner of Customs in Surat had relied on judicial precedents that were either wholly fabricated or misattributed — what the bench of Justices Dipankar Datta and Sheel Nagu called "a hallucination of AI."
The underlying dispute was mundane: the Commissioner's October 8, 2025 order had penalized diamond trader Vijay Ghanshyam Gadiya under Section 114 of the Customs Act, 1962, for allegedly mis-declaring natural diamonds as lab-grown stones to secure a lower tariff. The Gujarat High Court upheld the penalty on January 20, 2026. What made the case extraordinary was what the Supreme Court found when it checked the citations underpinning the order — case law that, on independent verification, simply did not say what it was cited for, or did not exist at all.
The Court's language was unambiguous: "Assistance can never be substituted for adjudication. AI may well serve as training wheels but entrusting it with the pilot's seat would be both imprudent and dangerous." It set aside both the High Court's confirmation and the original penalty order, remanded the matter to a different officer of the same rank, and directed the appointing authority to consider disciplinary action against the officer who wrote the flawed order.
Part of a Pattern, Not a One-Off
This is not India's judiciary encountering AI hallucination for the first time. On July 2, 2026, in Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. (2026 INSC 668), the Supreme Court set aside NCLT and NCLAT orders after finding that six of the tribunal's cited precedents were fabricated or contained invented paragraphs — and declared a "zero-tolerance" rule that citing unverified AI-generated case law is professional misconduct by an advocate, and a "serious lapse" if a judge relies on it. That ruling was aimed at lawyers and tribunals. Gadiya extends the same logic to a customs officer performing quasi-judicial adjudication — arguably a more consequential extension, because tax and customs authorities issue vastly more adjudicatory orders than courts do, with far less scrutiny of each one's reasoning.
The pattern is broader still: Bombay High Court imposed a ₹50,000 cost in January 2026 for a party's submission of fake case law, and Delhi High Court saw a similar petition withdrawn in September 2025 after opposing counsel exposed fabricated citations. The Supreme Court's own AI Committee, chaired by Justice P.S. Narasimha, published draft "Regulations for Use of Artificial Intelligence (AI) in Courts, 2026" on June 3, 2026, which bar AI from determining judicial outcomes, assessing bail risk, or operating as an unexplainable system in any rights-affecting proceeding, while permitting it for case management, transcription, and legal research support.
The Case for Caution — and Its Limits
The strongest argument for a hard line here is not abstract: an adjudicator who cannot tell a real citation from a fabricated one has, by definition, failed to actually adjudicate. A ₹425-crore penalty is not a clerical matter, and if the reasoning supporting it rests on invented law, the order has no legitimate basis regardless of whether the underlying facts might have supported a penalty anyway. Courts also have a narrower institutional worry: citation integrity is the substrate of common-law reasoning. If tribunals or officers start treating AI output as settled precedent without verification, the entire apparatus of stare decisis degrades quietly and invisibly, one uncontested order at a time. A bright-line rule that any hallucinated material voids the decision — regardless of whether it was outcome-determinative — is a reasonable response to a genuinely hard-to-detect failure mode.
But the remedy the Court chose is also a proportionate one worth noting, precisely because it did not reach for a blanket ban on AI-assisted drafting in adjudication. The bench's own framing — AI as "training wheels," not the "pilot" — endorses continued use of AI tools for research and drafting support, so long as a human adjudicator actually verifies the output before signing an order. That is the correct calibration: the failure in Gadiya was not that a customs officer used a generative AI tool, but that the officer (or whoever prepared the order) never checked whether the tool's citations were real. Punishing the tool's use categorically would slow down an already overburdened adjudicatory system — India's customs and tax authorities face enormous case backlogs — without addressing the actual defect, which is a verification gap, not a technology-use problem.
What Should Follow
The draft Supreme Court AI regulations, still pending finalization after a public comment period that closed July 15, 2026, are the right vehicle to formalize what Gadiya and Pooja Ramesh Singh have established through case law: AI-assisted drafting is permissible, verification of every citation before signing is mandatory, and the penalty for skipping that step is that the order carries no legal weight, however sound its ultimate conclusion might otherwise have been. Extending an explicit verification protocol — not a use-ban — to tax, customs, and other quasi-judicial administrative bodies, which output far more adjudications annually than the appellate courts do, would close the gap this case exposed before ₹425-crore orders are the norm rather than the outlier.