Every so often a judgment lands that is less about the parties before the court and more about the health of the system that produced it. Vijay Ghanshyam Gadiya v. Union of India (2026 INSC 947), decided by the Supreme Court on 2 September 2026, is one of them. On the face of it, this is a customs valuation dispute over diamonds. Read a little closer, and it is a warning shot to every officer, advocate and consultant who has quietly started leaning on generative AI to do their thinking for them.
The case, in brief
The Additional Commissioner of Customs, Surat, passed an Order-in-Original on 8 October 2025 imposing a penalty of ₹425,27,99,100, a shade over ₹425 crore, on the appellant under Section 114 of the Customs Act, 1962. The allegation was familiar enough: a consignment of natural diamonds had allegedly been mis-declared as lab-grown diamonds to attract a lower rate of duty. The appellant challenged the penalty before the High Court of Gujarat and lost, his appeal dismissed on 20 January 2026. He then carried the matter to the Supreme Court.
What is remarkable is that the Supreme Court never reached the merits of the valuation dispute at all. It did not need to. Counsel for the appellant contended that several of the judgments and articles the department had relied upon in the Order-in-Original were generated by Artificial Intelligence. The Bench, Justices Dipankar Datta and Sheel Nagu, did something that ought to be unremarkable but is becoming depressingly necessary: it checked. Judge by judge, citation by citation.
The findings were damning. Some of the case laws relied upon by the department were non-existent or carried fake citations. Others were real decisions that simply did not lay down the proposition attributed to them: the classic signature of an AI “hallucination,” where a model confidently manufactures authority that reads plausibly but says nothing of the sort. An adjudicating authority of the Union of India had, in other words, built a ₹425-crore penalty partly on foundations that dissolved the moment anyone looked.
What the Court held
The Supreme Court anchored itself in its own recent precedent, Pooja Ramesh Singh v. Jammu & Kashmir Bank Ltd. (2026 SCC OnLine SC 1258), which had already declared a policy of zero tolerance for producing, citing or relying on AI-generated precedents without verification. That decision framed the point in stark terms: reliance on fake or hallucinated material renders a decision “no decision in the eyes of the law,” regardless of whether the tainted material actually influenced the outcome. Even an iota of such material, the Court had said, is enough to vitiate the sanctity of adjudication, and the standard binds the Bench and the Bar alike.
Applying that to the Surat order, the consequence was almost mechanical. The reliance on dubious material, the Court held, was fatal to the sustainability of the penalty. Tellingly, the learned Additional Solicitor General did not join issue. The Union did not even try to defend the order on this point.
The Bench then added a passage that deserves to be pinned above every desk in every adjudicating office in the country. Noting that the Supreme Court has itself released draft Regulations for Use of Artificial Intelligence in Courts, 2026, it accepted that AI as an assistive tool to speed up decision-making cannot be denied, but sounded a caution that is hard to improve upon: 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.
The relief followed the logic. The Court set aside both the High Court’s order and the Order-in-Original, revived the proceedings, and directed that the matter be decided afresh by an officer of the same rank other than the one who passed the original order. And in a line that should focus minds, it left it open to the appointing authority to initiate such action as may be warranted against the author of the order, a reminder that hallucinated citations are not a victimless drafting shortcut.
This is not a one-off
It would be comforting to treat Gadiya as an aberration. It is not. It is the latest entry in a lengthening list.
The problem announced itself in India in December 2024, when the Bengaluru bench of the Income Tax Appellate Tribunal decided the Buckeye Trust matter, a trust-taxation dispute running into hundreds of crores, on the strength of Supreme Court and High Court decisions that turned out not to exist. The Tribunal had to recall its own order once the fabrications came to light. In October 2025, the Bombay High Court quashed a faceless income-tax assessment of roughly ₹27.91 crore after finding it rested on three non-existent precedents. A trial court in Andhra Pradesh was found in early 2026 to have relied on four fabricated judgments, drawing initially only a “word of caution” from the High Court, a leniency the Supreme Court has since firmly disowned in favour of the zero-tolerance line now applied in Gadiya.
Two features of the Indian experience stand out. First, this is not merely a problem of errant lawyers; the fabrications have crept into orders written by tax officers, tribunals and even judges. Second, and this is the part that matters most for the trade and tax community, the exposure runs right through the quasi-judicial machinery that businesses deal with every day: Assessing Officers, GST Enforcement Officers, and Customs Adjudicating Authorities. The Supreme Court’s November 2025 White Paper on AI and the judiciary flagged exactly this risk, and the draft 2026 Regulations respond to it by insisting on human primacy, treating AI output as merely advisory, and requiring that its accuracy be verified before it is relied upon.
Our view
We read Gadiya less as a story about diamonds and more as a structural cure for a structural rot.
First, the principle is exactly right, and its breadth is deliberate. By holding that even a trace of hallucinated material voids a decision irrespective of whether it influenced the outcome, the Court has refused to get drawn into the impossible exercise of measuring how much a phantom citation “really mattered.” That is sound. The integrity of an order is not a question of degree. An adjudicating authority that cannot vouch for its own authorities has not adjudicated; it has performed adjudication. The distinction is everything.
Second, this cuts both ways, and practitioners should sit up. The same standard that protected Mr Gadiya is now a live ground of challenge for any importer, exporter or taxpayer facing an order-in-original, a show-cause notice or an assessment that leans on suspect authority. In our own practice, across customs appeals, GST adjudications and SCN replies, verifying the citations in the department’s own order has quietly become part of the drill. Where an OIO or SCN cites a decision, the question is no longer only “is the ratio correctly applied?” but “does this decision exist, and does it say what the officer claims?” Gadiya elevates that check from good hygiene to a potentially dispositive argument. We would go further: a well-drafted reply should now, where warranted, put the citation on record and invite the authority to verify it, turning the department’s own carelessness into a documented infirmity.
Third, and this is the caution, the shoe is on our foot too. The Court was scrupulously even-handed. Its zero-tolerance rule binds the Bar as much as the Bench, and the draft Regulations will require disclosure of AI use in filings. A consultant or advocate who fires an AI-generated draft into a reply without checking every authority is not saving time; they are manufacturing a liability, and now a possible finding of professional misconduct. The tools are extraordinary. We use them, and this very article was checked against primary sources precisely because the subject demands it. But the responsibility for what leaves the office is, and must remain, human.
The real lesson
It is fashionable, after a judgment like this, to reach for the conclusion that AI has no place in law. That is the wrong lesson, and the Court took care not to draw it. The problem in Gadiya was never the tool; it was the abdication: the moment an officer let a machine’s confident output stand in for the discipline of actually reading the law. AI in the drafting seat is a research assistant that never says “I’m not sure.” Treated as a starting point that a trained mind then verifies, it is a genuine multiplier.
The Supreme Court’s metaphor is the one to remember. “Training wheels, not the pilot’s seat”. For everyone who drafts, files or decides in the world of trade and tax, that is now less a suggestion than a rule of survival.
This article is for general information and does not constitute legal advice. The views expressed are personal.
