AI-Hallucinated Precedents in Adjudication: Why the Supreme Court Set Aside a ₹425-Crore Customs Penalty

An adjudicatory order cannot acquire legal validity from authorities that do not exist or from judgments that say something materially different from the proposition attributed to them. In Vijay Ghanshyam Gadiya v. Union of India, 2026 INSC 947, the Supreme Court applied that principle to a customs penalty of ₹425,27,99,100 and delivered a direct warning about artificial intelligence in legal decision-making: AI may assist adjudication, but it cannot replace the adjudicator’s own verification, reasoning and responsibility.

The ruling is not an acquittal on the customs allegations. The Supreme Court did not decide whether the consignment had been misdeclared or whether a penalty was ultimately warranted. It set aside a legally tainted process, revived the proceedings and directed fresh adjudication by another officer of the same rank. That distinction is essential to understanding both the remedy and the wider precedent.

The customs dispute

The Additional Commissioner of Customs, Surat passed an Order-in-Original on 8 October 2025 imposing the penalty under Section 114 of the Customs Act, 1962. The allegation was that a consignment of natural diamonds had been misdeclared as lab-grown diamonds to obtain a lower tariff. The Gujarat High Court dismissed the trader’s challenge on 20 January 2026.

Before the Supreme Court, the appellant argued that several judgments and articles cited in the adjudication order had been generated through artificial intelligence. The Court considered it unnecessary to enter the merits of the customs dispute because the challenge to the integrity of the reasoning process was capable of disposing of the appeal.

The Court independently checked the cited material

Paragraphs 3 and 4 of the order record that the Supreme Court itself examined the authorities relied upon. It found three distinct defects: some supposed cases did not exist; some carried fake citations; and some judgments existed but did not lay down the legal propositions attributed to them. The last category is especially important. Citation checking is not satisfied merely by confirming that a case name appears in a database. The decision must be read sufficiently to confirm the court, factual setting, issue, ratio, qualifications and final disposition.

The Court described the incorrect attribution of propositions to genuine cases as an apparent AI hallucination. This meant that the adjudicating authority’s reasoning was not simply affected by a typographical mistake or an incomplete citation. It had invoked false legal authority as part of the decisional process.

Why the defect was fatal

The Supreme Court relied on Pooja Ramesh Singh v. Jammu & Kashmir Bank Ltd., 2026 SCC OnLine SC 12583. In that case, orders of the National Company Law Tribunal and National Company Law Appellate Tribunal were set aside after the adjudicatory chain was found to contain fabricated or misattributed precedents. The Court declared zero tolerance for producing, citing or using AI-generated precedents without verification and held that reliance on hallucinated authority destroys the integrity of adjudication.

Vijay Ghanshyam Gadiya applies that rule beyond conventional courts and insolvency tribunals to a statutory customs adjudicator. Paragraph 7 holds that reliance on dubious AI-produced material was fatal to the sustainability of the penalty order. The Additional Solicitor General did not contest remand on this ground.

The rule protects more than citation accuracy. A person affected by a penal or quasi-penal order is entitled to a decision based on real law, relevant evidence and reasons genuinely formed by the authorised decision-maker. If invented precedent enters the reasoning, the affected person cannot meaningfully test its authority, distinguish its facts or challenge its ratio. Appellate review is equally distorted because the apparent legal foundation is fictitious.

AI assistance is permitted; delegation of judgment is not

The Court did not prohibit artificial intelligence. Paragraph 6 acknowledges its potential as an assistive tool for accelerating work and refers to the draft Regulations for Use of Artificial Intelligence in Courts, 2026. The controlling distinction is between assistance and substitution.

AI may help locate possible authorities, organise issues, translate text, create a preliminary chronology or identify questions for human review. It cannot be entrusted with determining what the law is, whether a cited authority exists, what that authority actually decided or how law applies to disputed facts without responsible human verification. The legal decision must remain the work of the person or institution to whom the law entrusts adjudication.

This is why merely adding a generic statement that “AI was used” cannot cure a hallucinated citation. Transparency may reveal the method, but validity still depends on checking every authority and independently adopting the reasoning.

What the Supreme Court actually ordered

Paragraph 8 set aside both the Gujarat High Court order and the Order-in-Original. The customs proceedings were revived for a fresh decision by an officer holding the same rank, but not the officer who authored the defective order. Paragraph 11 left it to the appointing authority to determine whether action should be initiated against that officer in accordance with law.

The Court therefore supplied procedural restoration, not substantive exoneration. The department remains free to establish the alleged misdeclaration and the statutory basis for penalty through admissible evidence and valid legal reasoning. The appellant remains free to contest jurisdiction, classification, valuation, mens rea, evidence, proportionality and every other available defence. Neither side may treat the set-aside penalty as a final determination of the customs merits.

The High Court dimension

The Gujarat High Court’s order was also set aside because it had confirmed the defective adjudication. The Supreme Court’s disposition underscores that judicial review cannot safely proceed on the assumption that every authority printed in an administrative order is genuine. Where a litigant identifies concrete signs of fabricated citations, the reviewing court should verify the authorities rather than assess the challenged reasoning on a false legal foundation.

This does not require courts to investigate every citation sua sponte in every case. It does require attention when specific discrepancies are demonstrated, particularly where the order imposes a massive civil penalty, confiscation, loss of licence or another serious consequence.

Duties of advocates and adjudicating authorities

Pooja Ramesh Singh states that an advocate’s citation of unverified AI-generated judgments can amount to professional misconduct. Vijay Ghanshyam Gadiya focuses on the adjudicating authority but reinforces the same institutional standard. Responsibility cannot be shifted to software, a junior, a researcher or a party once the authority appears in a signed submission or decision.

For advocates, a minimum verification protocol should include opening the complete judgment; confirming the cause title, court, bench, date, case number and citation; locating the precise paragraph; checking whether the passage is ratio, obiter, a party submission or a quotation from another case; reading the disposition; and checking subsequent treatment.

For adjudicators, the record should disclose the statutory provision applied, material evidence, submissions, findings and authentic authorities supporting each decisive proposition. Authorities discovered independently after hearing require particular caution. If a proposition materially affects the outcome, procedural fairness may require that parties be given an opportunity to address it.

A practical citation-audit protocol

Every legal office using generative AI should maintain a source ledger for each draft. The ledger should record the direct source link, court, case number, date, paragraph relied upon, proposition supported and the name of the human reviewer. Secondary summaries may be useful discovery tools, but should not be the final proof of a ratio when the judgment is publicly accessible.

Quoted words should be copied from the judgment, not reconstructed from an AI answer. Parallel citations should be checked for mismatches. A case that cannot be located should be removed rather than softened with phrases such as “reportedly held.” Where the primary record is unavailable, the limitation should be stated and the authority should not carry a dispositive proposition.

The principle extends beyond generative AI

Although the Court addressed AI hallucination, the governing defect is false authority within adjudication. The same concern can arise from an inaccurate headnote, a careless digest, a fabricated quotation, an unchecked intern’s note or a mistaken database result. Technology changes the speed and scale of the risk; it does not change the adjudicator’s duty to reason from authentic law.

The ruling also does not mean that any minor citation error automatically nullifies every order. Vijay Ghanshyam Gadiya concerned non-existent, fake and materially misdescribed authorities used in the decisional process, and it applied the stringent rule already declared in Pooja Ramesh Singh. Courts will still need to distinguish a harmless clerical defect from reliance on fictitious legal material. The safe institutional response is exact verification before signing, not litigation over whether an avoidable hallucination was harmless.

Conclusion

Vijay Ghanshyam Gadiya establishes that a high-value statutory penalty cannot stand when its legal foundation is polluted by fabricated or materially misrepresented precedents. The Supreme Court preserved the legitimate use of AI as an assistant while insisting on human primacy, source verification and accountable reasoning. Its remedy was equally disciplined: the defective orders were set aside, the merits were left open, and a fresh adjudication was assigned to a different officer.

The enduring lesson is simple but exacting. AI can propose a case; only a responsible human reader can establish that the case exists, understand what it decided and determine whether it lawfully governs the dispute.

Sources

Vijay Ghanshyam Gadiya v. Union of India, 2026 INSC 947, Civil Appeal arising from SLP (Civil) No. 15605 of 2026, Supreme Court order dated 2 September 2026, especially paragraphs 2–11. The judgment is searchable by neutral citation on the Supreme Court of India judgments portal: https://www.sci.gov.in/judgements-judgement-date/

Pooja Ramesh Singh v. Jammu & Kashmir Bank Ltd., 2026 SCC OnLine SC 12583, Supreme Court judgment dated 2 July 2026, especially paragraphs 7–17. The judgment is searchable on the Supreme Court of India judgments portal: https://www.sci.gov.in/judgements-judgement-date/

Customs Act, 1962, including Section 114: https://www.indiacode.nic.in/handle/123456789/2475

Supreme Court of India, draft Regulations for Use of Artificial Intelligence in Courts, 2026, referred to in paragraph 6 of Vijay Ghanshyam Gadiya: https://www.sci.gov.in/

This article provides general legal information and does not constitute advice concerning any customs adjudication, disciplinary proceeding or use of artificial intelligence.

#ArtificialIntelligence #SupremeCourt #CustomsLaw