A ₹425-crore order built on cases that don't exist
On September 2, 2026, the Supreme Court set aside a ₹425.27-crore penalty imposed on diamond trader Vijay Ghanshyam Gadiya, ruling that the Additional Commissioner of Customs in Surat had relied on case law that was either non-existent or fabricated when passing the original October 8, 2025 order. The Gujarat High Court's January 20, 2026 order upholding that penalty was struck down too. A bench of Justices Dipankar Datta and Sheel Nagu found some citations carried fake references altogether, while others were real judgments stripped of their actual reasoning and assigned propositions they never made — a hallmark of generative AI hallucination (Bar and Bench; 2026 INSC 947).
The Court's line will likely be the one quoted for years: "AI may well serve as training wheels but entrusting it with the pilot's seat would be both imprudent and dangerous." The case was remanded for fresh adjudication by a different officer of equal rank, and the appointing authority was told to weigh disciplinary action against the original decision-maker (LiveLaw).
Not a one-off — the third strike in eight months
What makes this ruling significant isn't novelty; it's pattern. Two months earlier, on July 2, 2026, the same Court set aside an NCLT insolvency order and its NCLAT affirmation in Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. (2026 INSC 668) after finding the tribunal had propped up its decision with six "authorities," most either wholly invented or real citations with fabricated paragraphs attached. That bench went further, declaring a formal "zero-tolerance" standard: a decision resting on even a fragment of hallucinated material "is no decision at all," regardless of whether the fake material changed the outcome, and citing such material without verification is professional "misconduct" (2026 INSC 668). Before that, in December 2024, Bengaluru's Income Tax Appellate Tribunal recalled a roughly ₹669-crore order within a week after discovering four fabricated citations had been fed into the ruling — reportedly the product of an unverified ChatGPT query by a tax department representative (LiveLaw, "Phantom Precedents").
Across customs, insolvency, and tax adjudication, the failure mode is identical: an overworked decision-maker under caseload pressure treats a generative model's confident output as a finished citation rather than a lead requiring verification.
The steelman: why courts are right to be alarmed
The case for a hard line here is genuinely strong, and it would be dishonest to wave it away. Adjudication — whether by a customs officer, a tribunal, or a judge — derives its legitimacy from reasoned engagement with real precedent. When a ₹425-crore penalty rests on a citation that was never decided, the losing party isn't just wronged on the merits; they're denied the basic due-process guarantee that someone actually reasoned through their case. Left unchecked, this erodes public trust in every subsequent ruling those authorities issue, verified or not. The Court's zero-tolerance framing — voiding a decision even if the fake material was peripheral — is a defensible bright-line rule precisely because ex-post, case-by-case judgments about "how much the hallucination mattered" invite exactly the kind of hair-splitting that let this problem fester across at least three tribunals over two years.
Why the ruling is still the right calibration
That said, the Supreme Court did not ban AI-assisted research, and it shouldn't have. Both rulings are explicit that AI as an assistive tool — speeding up research, drafting summaries, surfacing candidate precedents for a human to check — remains legitimate. The Court's own AI Committee has been piloting SUPACE, an internal research-assistance tool, since 2021, precisely on the theory that automation can compress research time without displacing judgment. The failure in both Gadiya and Pooja Ramesh Singh wasn't that AI touched the file; it was that nobody checked its output against a citator or the actual reported judgment before it became the basis for a nine-figure penalty. That is a verification failure, not an automation failure, and the remedy the Court has now applied twice — void the tainted order, remand for fresh adjudication by someone else, refer the individual for professional consequences — targets the right layer of the problem without foreclosing the tool.
What should follow
The risk is that overworked tribunals and revenue authorities, facing personal liability for an unverified citation, over-correct into blanket bans on AI-assisted drafting — trading a hallucination problem for a productivity collapse in a court system already carrying enormous pendency. The better fix, consistent with the Court's own reasoning, is mandatory citation verification against an authoritative database (SCC Online, Indian Kanoon, or the e-Courts system itself) before any AI-assisted draft is signed — a compliance step, not a technology ban. India's judiciary doesn't need to fear AI. It needs the verification discipline that turns a training-wheel tool into a safe one.