The Supreme Court Threw Out a Rs 425 Crore Penalty Because the Citations Were Fake

On 2 September 2026 the Supreme Court set aside a Rs 425.27 crore customs penalty against a diamond trader. The reason had nothing to do with diamonds. The order imposing the penalty cited judgments that do not exist.

What happened

The Additional Commissioner of Customs, Surat, imposed the penalty on 8 October 2025 under Section 114 of the Customs Act, 1962. The allegation was that the trader, Vijay Ghanshyam Gadiya, had mis-declared a consignment of natural diamonds as lab-grown diamonds to attract a lower tariff. The Gujarat High Court dismissed a challenge to that order on 20 January 2026.

When the appeal reached the Supreme Court, the bench of Justices Dipankar Datta and Sheel Nagu did something that turned out to decide the case. Before looking at the merits, they checked every judgment and article the original order relied on.

Some of the cases did not exist. Others carried fake citations. Others were real judgments that did not say what the order claimed they said. The court called this an apparent "hallucination of AI."

Why that was fatal to the penalty

The bench did not send the matter back for a correction. It quashed the High Court order and the original adjudication order outright, holding that reliance on dubious material went to whether the penalty could stand at all.

It leaned on its recent decision in Pooja Ramesh Singh vs Jammu & Kashmir Bank Ltd, which set out a zero-tolerance position on unverified AI-generated precedent. Advocates who cite such material without checking it may be committing misconduct. An adjudicating officer who relies on it commits a lapse that goes to the integrity of the decision.

The court's own line on where assistance ends: "assistance can never be substituted for adjudication."

What happens next

The proceedings have been revived for fresh adjudication by an officer of the same rank, though not the officer who passed the original order. Whether any action follows against the author of that order was left to the appointing authority.

Why this is not only a story about courts

The failure here is the one every organisation putting generative AI into a decision workflow is exposed to.

• The output was fluent, structured, and formatted exactly like the real thing.
• Nobody downstream checked the underlying references.
• The error survived a full tier of review, through a High Court.
• It was caught because someone at the top verified the citations by hand.

That last point is the one I keep coming back to. The control that worked was not a policy, a model guardrail, or a procurement clause. It was a person checking sources.

If you are putting AI into adjudication, underwriting, credit decisions, audit or compliance, this judgment is a fairly precise specification of what verification has to look like when the output is confident and wrong. The Supreme Court has also released draft guidelines on AI in Indian courts that bar its use for core judicial decision-making.

The penalty was Rs 425.27 crore. The thing that unravelled it was a citation nobody checked.

Source: Deccan Herald