Pooja Ramesh Singh: Supreme Court on Fake AI Citations

Independently written and published by Shahbaz Shah Legal Journal.

Why this judgment matters

*Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. & Anr.*, 2026 INSC 668, is an important Supreme Court of India ruling on the use of artificial intelligence in legal research and adjudication. The case was not decided against AI as a technology. It was decided against unverified legal authority entering a judicial order and then being repeated on appeal.

The Court held that an adjudicatory order founded, even partly, on fabricated or incorrectly attributed precedents cannot be treated as a lawful judicial decision. It placed responsibility at every point in the legal process: advocates must verify what they cite, judges and tribunals must verify what they rely upon, and appellate bodies must detect rather than reproduce errors in the order under challenge.

That principle is larger than one insolvency dispute. Courts exercise public power through reasons. If the stated authorities do not exist, the reasoning no longer permits the parties, the appellate court, or the public to test how the conclusion was reached. The defect therefore affects the integrity of adjudication itself.

Case at a glance

• **Case:** *Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. & Anr.*
• **Case number:** Civil Appeal No. 11950 of 2025
• **Neutral citation:** 2026 INSC 668
• **Court:** Supreme Court of India
• **Decision date:** 2 July 2026
• **Bench:** Justice Pamidighantam Sri Narasimha and Justice Alok Aradhe
• **Proceedings below:** National Company Law Tribunal, Mumbai Bench, and National Company Law Appellate Tribunal
• **Central issue:** Whether orders relying on fabricated, miscited, or falsely attributed authorities can survive judicial review
• **Result:** Both orders were set aside and the insolvency application was restored for a fresh decision on the merits

Background: the insolvency proceedings

The dispute arose from an application by Jammu and Kashmir Bank under section 7 of the Insolvency and Bankruptcy Code, 2016. The Bank sought initiation of the corporate insolvency resolution process against the corporate debtor. The National Company Law Tribunal admitted the application on 28 August 2024, appointed an interim resolution professional, and declared the statutory moratorium under section 14 of the Code.

The appellant challenged that order before the National Company Law Appellate Tribunal. On 11 September 2025, the NCLAT dismissed the appeal. The matter then reached the Supreme Court.

The Supreme Court did not finally decide the underlying insolvency merits. Its concern was more fundamental: the NCLT order had referred to authorities that either did not exist, did not contain the attributed passages, or had been given the wrong citation. The appellate order repeated those authorities without detecting the problem.

What was wrong with the cited authorities?

The Court examined six authorities used in the NCLT order. The errors were not of one kind.

• One citation belonged to another existing case, while the paragraph attributed to it was not found in that authority.
• Two citations referred to real decisions, but the passages said to come from them did not exist in those judgments.
• Three citations were entirely nonexistent.

An affidavit filed on behalf of the Bank stated that its counsel had not cited these authorities during arguments. The NCLT had obtained them in the course of its own research. That fact did not reduce the problem. It made clear that verification is an institutional duty and not merely a dispute between opposing lawyers.

The NCLAT then relied on the same chain of authority. An appellate forum is expected to examine whether the order under challenge rests on valid law. Repeating the original errors converted a failure at first instance into a failure of appellate control.

The Supreme Court's holding

The Supreme Court treated fabricated precedent as a structural defect in adjudication. A judgment is not merely an announcement of the winner. It is an application of valid legal rules to established facts through reasons that can be inspected and challenged. A nonexistent precedent cannot perform that function.

The Court's central conclusions may be stated as follows:

1. **Human verification is mandatory.** AI-assisted research may be useful, but its output cannot be relied upon without checking the original reported or official source.
2. **Fabricated authority taints the decision-making process.** An order relying on invented cases or nonexistent passages is not a legally sustainable adjudication.
3. **Advocates remain professionally responsible.** A lawyer who places an unverified fake authority before a court may face professional-misconduct consequences.
4. **Judges and tribunals carry an independent duty.** Judicial research tools do not transfer decisional responsibility from the human adjudicator to software.
5. **Appellate review must catch the defect.** An appellate body cannot safely reproduce authorities from the impugned order without verifying them.
6. **Legitimate AI use is not prohibited.** The ruling targets unverified reliance, not responsible use of technology for assistance.

The Court directed the Bar Council of India to constitute a committee and frame appropriate guidelines concerning the use of AI and the citation of authorities. It also indicated that disciplinary consequences should follow where fake cases are cited without verification.

Final outcome

The Supreme Court set aside the NCLT admission order and the NCLAT appellate order. It restored the Bank's section 7 application to the NCLT for a fresh determination on the merits, preferably within two weeks. It directed the parties to maintain status quo in the meantime.

This procedural result is important. The Court did not hold that the insolvency application must fail, nor did it accept the appellant's substantive defence. It removed decisions produced through a defective reasoning process and returned the controversy to the competent tribunal for lawful adjudication.

Why a fake citation is not a minor technical error

A typographical mistake in a citation may sometimes be corrected where the judgment, proposition, and reasoning remain identifiable. Fabricated authority is different. It introduces a source that cannot be examined because it does not exist, or assigns a proposition to a real court that the court never expressed.

Four legal safeguards are lost when that happens.

### 1. Notice to the parties

Parties must know the legal material on which their rights will be decided. They cannot distinguish, answer, or explain a nonexistent case.

### 2. Discipline of precedent

Indian law depends on a hierarchy of courts and identifiable legal holdings. A fabricated case falsely claims the authority of that hierarchy.

### 3. Appellate review

An appellate court must be able to inspect the source, context, ratio, and treatment of the authority. A hallucinated citation defeats that process.

### 4. Public confidence

Judicial legitimacy depends on the belief that reasons are genuine, verifiable, and honestly connected to the law. Invented authorities make a reasoned order resemble an answer whose evidence has been manufactured.

Human in the loop is a legal duty, not a slogan

The judgment uses the idea of keeping a human in the loop. In legal practice, that requires more than a person pressing a final approval button. The reviewer must perform the checks that the tool itself cannot be trusted to perform.

For every authority produced by an AI system, lawyer, clerk, researcher, or search platform, the responsible person should:

1. open the judgment from the official court website or a reliable law report;
2. match the case title, court, date, case number, and citation;
3. open the exact paragraph said to support the proposition;
4. read enough surrounding text to understand the context and actual holding;
5. confirm that the proposition is part of the ratio rather than an argument, quotation, or incidental observation;
6. check whether the decision has been reversed, distinguished, doubted, or superseded;
7. verify the current statutory text independently; and
8. preserve the source used so that the court and opposing party can inspect it.

If any one of those checks fails, the authority should not be cited until the uncertainty is resolved. Confidence generated by fluent language is not legal verification.

Duties of judges, tribunals, and court staff

The decision correctly recognises that the risk is not confined to the bar. Courts increasingly use digital databases, search tools, summaries, and drafting assistance. Those tools can improve speed, but they also make it easy to import polished errors into formal orders.

A sound judicial protocol should include:

• direct access to authenticated court databases and legislation;
• a citation check before an order is signed;
• preservation of the authorities independently found by the court;
• an opportunity for parties to address a decisive authority found after the hearing;
• special scrutiny of quotations and paragraph references; and
• training that distinguishes search assistance from adjudicatory judgment.

Where an authority is material and was not argued by either party, procedural fairness may require the court to disclose it before relying on it. Technology does not remove the audi alteram partem principle.

Critical analysis: the judgment's major strengths

### It protects the truth conditions of judicial reasoning

The strongest part of the ruling is its refusal to treat invented precedent as an ordinary research mistake. A court order obtains authority from law that exists. The Supreme Court therefore placed the integrity of reasons before the convenience of preserving an otherwise plausible result.

### It assigns responsibility to humans, not software

The judgment avoids the false defence that the machine produced the error. An AI tool is not counsel, judge, tribunal, or disciplinary subject. The person who submits, adopts, or signs the material remains responsible.

### It does not impose a technology ban

A blanket prohibition would be both unrealistic and counterproductive. Legal professionals may properly use AI to organise issues, identify possible search terms, compare documents, or assist with first drafts. The Court instead drew the line at reliance without verification.

### It orders a fresh merits decision

By restoring the section 7 application, the Court separated defective reasoning from the substantive insolvency dispute. This protects both sides: the appellant is not bound by an order resting on false authorities, and the Bank does not lose its claim merely because the tribunal's research was flawed.

Questions that future cases and guidelines must answer

The judgment establishes a necessary bright line, but implementation will require further precision.

### 1. Materiality and curable error

The Court's strict language is understandable where fabricated authorities form part of the legal foundation of an order. Future cases may need to distinguish that situation from a harmless citation typo or an unnecessary reference where the same conclusion independently follows from verified binding law.

The distinction should not excuse invented cases. It should help courts identify the correct remedy: correction of an accidental citation, rehearing of a material issue, or setting aside the entire order where the defect infects the reasoning process.

### 2. Professional misconduct and culpability

Deliberately manufacturing a case is very different from negligently failing to detect a false result generated by a research tool. Both can be serious, but discipline should still observe notice, proof, proportionality, and an opportunity to be heard under the Advocates Act and applicable Bar Council rules.

Guidelines should identify relevant factors: whether the lawyer opened the source, whether the authority was material, whether the error was promptly disclosed and corrected, whether it was repeated, and whether fabrication was knowing or reckless.

### 3. Institutional verification systems

Individual diligence is essential, but it is not enough. Courts and bar councils should improve access to authentic, searchable, machine-readable judgments. A verification protocol will work best when a neutral citation or case number reliably leads to the official text.

The judgment should therefore encourage infrastructure as well as punishment: complete official databases, stable links, clear correction histories, and citation-validation tools whose output remains subject to human review.

### 4. Confidentiality and data protection

The ruling was not a decision on client confidentiality or data protection. Those issues nevertheless belong in the Bar Council guidelines. Lawyers should not upload privileged pleadings, personal records, commercially sensitive documents, or sealed material to external AI services without lawful authority, adequate safeguards, and informed professional judgment.

Implications for Indian legal practice

For advocates, the decision makes source verification part of competent professional preparation. A citation list generated in seconds may require much longer to validate. That time is not administrative waste; it is part of the duty owed to the court and client.

For law firms, chambers, and in-house teams, a written AI-use policy is now prudent. It should specify permitted tools, restricted data, mandatory source checks, supervisory responsibility, and correction procedures. Junior work should be reviewed on the original authorities, not only on a fluent memorandum.

For tribunals, the case demonstrates that procedural speed cannot justify unverifiable reasoning. Specialist jurisdiction does not reduce the obligation to cite valid law.

For appellate courts, the decision is a warning against copying the authority section of the impugned order. The validity and relevance of cited cases form part of appellate scrutiny.

For legal education, citation verification, primary-source research, and responsible AI use should become practical skills rather than optional technology topics.

A wider lesson for South Asian courts

Although this is a Supreme Court of India decision, its reasoning has persuasive importance across jurisdictions that share common-law methods, hierarchical precedent, and rapidly expanding use of generative AI. Pakistan, India, Bangladesh, Sri Lanka, and other systems face the same basic risk: a confident-looking passage can enter a pleading or judgment before anyone opens the authority.

The durable rule is simple. AI may help locate a possible answer; it cannot authenticate the answer. The original judgment, statute, notification, or record remains the legal source.

Conclusion

*Pooja Ramesh Singh* is best understood as a rule-of-law judgment rather than an anti-technology judgment. The Supreme Court protected the minimum conditions of legitimate adjudication: real legal sources, reasons capable of verification, fair opportunity to respond, and identifiable human responsibility.

Its strict response is justified because fabricated precedent does more than weaken an argument. It attributes law to a court that the court did not make. The next step should be equally serious but constructive: Bar Council guidelines, reliable official databases, proportionate disciplinary procedures, judicial citation checks, and practical training for every professional who uses AI-assisted research.

The decisive professional habit remains unchanged by technology: never cite what you have not opened, read, and verified.

This commentary is independent legal analysis for research and general information. It does not determine the merits of the restored insolvency proceedings or provide case-specific advice. Readers should verify the official judgment and current statutory text before relying on any proposition in proceedings.