A fabricated precedent, a real penalty
On September 2, 2026, the Supreme Court set aside a ₹425.27-crore customs penalty imposed on diamond trader Vijay Ghanshyam Gadiya, after finding that the order underpinning it was built on legal ground that did not exist. In Vijay Ghanshyam Gadiya v. Union of India (2026 INSC 947), Justices Dipankar Datta and Sheel Nagu found that the Additional Commissioner of Customs in Surat had relied on case citations that were either fabricated outright or attributed legal propositions the cited judgments never actually held — the signature fingerprint of an AI hallucination. The original order, issued October 8, 2025 under Section 114 of the Customs Act, 1962, accused Gadiya of mis-declaring natural diamonds as lab-grown to reduce tariff liability. The Gujarat High Court had upheld it. The Supreme Court has now remanded the matter for fresh adjudication by a different officer, invited the appointing authority to consider disciplinary action, and delivered a line that will likely outlive the case itself: AI "may well serve as training wheels but entrusting it with the pilot's seat would be both imprudent and dangerous."
Not an aberration — a pattern
What makes this ruling significant isn't its novelty but its recurrence. Barely two months earlier, on July 2, 2026, the same court — a differently constituted bench of Justices P.S. Narasimha and Alok Aradhe — set aside NCLT and NCLAT insolvency orders in Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. (2026 LiveLaw (SC) 653) after discovering that a tribunal had propped up its ruling with six "precedents," several of which were AI-fabricated. That ruling went further than Gadiya's case procedurally: it declared a zero-tolerance standard for the Bar and Bench citing unverified AI material, held that a decision resting even partly on hallucinated authority is "no decision in the eyes of the law," and directed the Bar Council of India to examine the problem. Lower courts and tribunals have supplied a steady drumbeat of smaller episodes since — a Bengaluru tax tribunal recalling a roughly ₹669-crore order after citing four non-existent judgments, a Bombay High Court quashing a ₹27.91-crore assessment on similar grounds, an Andhra Pradesh trial court order taken up suo motu by the Supreme Court in February 2026. What connects Gadiya's case to the rest is that the fabrication didn't come from a litigant cutting corners — it came from the state's own adjudicating machinery, wielding coercive power over a private party on the strength of citations nobody had checked.
The case for tighter guardrails
There is a real argument for treating this as more than a training issue. A customs officer's order isn't a law-school memo — it authorizes the state to extract ₹425 crore from a citizen, and the citizen has no practical way to detect that the legal reasoning against them is synthetic until it reaches appellate scrutiny, by which point months and legal fees have been spent. Quasi-judicial and administrative bodies operate under time and staffing pressure that makes AI-assisted drafting attractive, and unlike a High Court bench, an Additional Commissioner of Customs has no law clerks cross-checking citations. A reasonable regulator could conclude that any AI-assisted order affecting a citizen's property or liberty needs a mandatory verification step — a human sign-off confirming every cited authority actually exists and says what it's quoted as saying — before it goes out. MeitY's own India AI Governance Guidelines, unveiled in November 2025, already gesture at this with a "do no harm" principle built around human-centric oversight rather than blanket AI bans.
Why the Court got the remedy right
But the Supreme Court, notably, didn't reach for a ban. Both rulings are careful to say the problem is unverified reliance on hallucinated material presented as precedent, not the underlying use of AI as a drafting or research aid. That distinction matters, because it points to the actual failure mode: a customs officer (or a judicial clerk, in Pooja Ramesh Singh) treated AI output as citation-ready without opening a single one of the cases it cited. That is a verification failure, not an AI failure, and it long predates large language models — Indian courts have quashed orders over misquoted or non-existent precedents pulled from bad research long before ChatGPT existed. The fix the Court has now supplied twice in two months — void the order, remand to a fresh officer, put professional consequences on the table — is proportionate and self-enforcing: it makes the cost of unverified AI output fall on the official who signed the order, not on AI tools generally. A blanket prohibition on AI-assisted drafting in government offices would forfeit real productivity gains — case-law search and first-draft synthesis are exactly the kind of grunt work AI shortens — to solve a problem that a citation-verification checklist solves more cheaply. India's regulators would do better to formalize what the Court has effectively already mandated through case law: AI can draft, but a human must verify every citation against the primary source before an order with legal force goes out. That's a training-wheels policy, not a pilot's-seat ban — and it is the one the Court itself just prescribed.