TL;DR: On 2 July 2026 the Supreme Court of India decided Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. & Anr., 2026 INSC 668, and set aside orders of the NCLT and the NCLAT because the NCLT had relied on six precedents that were fake, wrongly cited, or carried paragraphs that do not exist in the real judgments. The bench of Justice Pamidighantam Sri Narasimha and Justice Alok Aradhe held that courts must adopt a zero-tolerance mode for producing, citing or using AI-generated precedents without verification, that citing such judgments without verification is misconduct on the part of an advocate, and that a decision built on fake or hallucinated material is no decision in the eyes of the law even if only an iota of it entered the reasoning. The Court directed the Bar Council of India to constitute a committee and prescribe both a guiding principle and the disciplinary action that will follow a breach.

On this page

What the Supreme Court held on 2 July 2026

The Supreme Court of India decided Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. & Anr., Civil Appeal No. 11950 of 2025, on 2 July 2026. The neutral citation is 2026 INSC 668 and the judgment is marked reportable. The bench was Justice Pamidighantam Sri Narasimha and Justice Alok Aradhe. The full text is on the Supreme Court of India’s judgment portal and on Indian Kanoon.

Two holdings matter above everything else.

The first is the rule of conduct. At paragraph 7 the Court said it is necessary for courts to adopt a zero-tolerance mode for producing, citing or using AI-generated precedents without verification, and that it is misconduct on the part of an advocate to cite such judgments without verification. The same paragraph says it is a serious lapse if a judge relies on fake or hallucinated AI-generated material as a precedent.

The second is the consequence. Also at paragraph 7, the Court declared that such a decision is no decision in the eyes of the law, irrespective of whether the material had a direct or indirect bearing on the decision-making, and that such decisions are to be set aside even if an iota of fake or hallucinated material enters the decision-making process. Paragraph 17 puts it more strongly: a decision of a court or adjudicating authority based on material which is fake and hallucinated is no decision at all, and amounts to subversion of the rule of law.

The opening line tells you this was not a first occurrence. The Court began by saying this is yet again a case where the Tribunal relied on non-existent, fake and hallucinated material, generated through Artificial Intelligence, as if it were a precedent. By July 2026 the problem of AI hallucinated citations in India had already reached the Supreme Court more than once.

How fake precedent got into an insolvency order

The appellant, Pooja Ramesh Singh, is a suspended director of Essel Infraprojects Ltd., the corporate debtor and the corporate guarantor of the original borrower, Pan India Utilities Distribution Company Ltd. Pan India Utilities had taken loan facilities from Jammu and Kashmir Bank Ltd., the first respondent, secured by a corporate guarantee from Essel Infraprojects. It went into financial stress, failed to keep to its repayment schedule, and its loan accounts were classified as non-performing assets.

Jammu and Kashmir Bank filed an application under Section 7 of the Insolvency and Bankruptcy Code, 2016 before the National Company Law Tribunal at Mumbai. The NCLT found debt and default and, by an order dated 28 August 2024, admitted the application, appointed an interim resolution professional, and declared a moratorium under Section 14 of the Code. The wider procedural context of this jurisdiction is set out in the IBBI CIRP fourth amendment 2026 explainer.

The appellant went to the National Company Law Appellate Tribunal in Company Appeal (AT) (Insolvency) No. 1808 of 2024. Her argument was on the merits: the liabilities of the corporate guarantor had been transferred to another company through a demerger followed by an amalgamation, and a renewed sanction letter dated 18 November 2017 did not mention the guarantee. The NCLAT dismissed the appeal on 11 September 2025, holding that clause 8 of the guarantee deed kept the guarantee alive through any absorption or amalgamation of the corporate debtor. In doing so it reproduced in paragraph 12 of its own opinion the six Supreme Court decisions the NCLT had relied on. That list unravelled the case.

Before the Supreme Court, senior counsel Ms. Madhavi Divan appeared for the appellant and pointed out at the outset that the citations relied on by the NCLT were fake and non-existent, probably AI-generated, and that even where a citation was accurate, the paragraph excerpted from it could not be traced in the law reports. The Court then ran its own check.

One further fact changes who is responsible. Paragraph 16 records that the first respondent filed an affidavit stating that the alleged judgments relied on by the NCLT were not cited by its counsel at the bar, and were obtained through the adjudicating authority’s own research. The fake authorities did not come from the parties. They came from the tribunal.

The six citations and the three ways they failed

Paragraph 15 sets out the Court’s independent examination of each of the six decisions the NCLT relied on. It is the most instructive part of the judgment, because fabrication turns out to have three distinct failure modes.

Citation relied on by the NCLTWhere citedReal case?Real paragraph?
State Bank of India v. M/s Shree Ram Urban Infrastructure Ltd., 2020 SCC OnLine SC 341Para 44 of the NCLT judgment✗ Wrong cause title for an existing reported judgment✗ Non-existent paragraph
Everest Kento Cylinders Ltd. v. Union of India, (2015) 2 SCC 1Para 45✓ Correct citation✗ Non-existent paragraph
ICICI Bank Ltd. v. Urban Infrastructure Real Estate Ltd., (2019) 16 SCC 528Para 47✗ Non-existent citation✗ Nothing to quote
V.S. Dempo & Co. Ltd. v. Reliance Communications Ltd., (2021) 10 SCC 176Para 49✗ Non-existent citation✗ Nothing to quote
Canara Bank v. N.G. Subbaraya Setty & Anr., (2018) 16 SCC 228Para 51✓ Correct citation✗ Non-existent paragraph
Sarbjit Singh v. Union Bank of India, (2022) 7 SCC 464Para 53✗ Non-existent citation✗ Nothing to quote

Three of the six citations do not exist at all. Those are the easy catches. Search the reporter, find nothing, and you know.

Two of the six are real, correctly cited judgments with paragraphs that are not in them. Everest Kento Cylinders Ltd. v. Union of India, (2015) 2 SCC 1 and Canara Bank v. N.G. Subbaraya Setty & Anr., (2018) 16 SCC 228 are genuine Supreme Court decisions. A verification method that stops at confirming the case exists passes both and misses the fabrication, because the invented text sat inside a real authority.

One of the six is a hybrid. The NCLT cited State Bank of India v. M/s Shree Ram Urban Infrastructure Ltd. at 2020 SCC OnLine SC 341. The Court’s footnote records that the correct cause title for that citation is M. Subramaniam v. S. Janaki, (2020) 16 SCC 728, also reported at 2020 SCC OnLine SC 341. The reporter reference points at a real judgment between entirely different parties, and the paragraph attributed to it does not exist either. Searching by case name returns nothing and looks like a typing error. Searching by citation returns a real judgment and looks like garbled transcription. Neither catches it without opening the report.

So “does this case exist” is the first question and not the last. The question that would have caught five of these six is whether the paragraph exists in that judgment, in those words. The mechanics are in how to cite Indian judgments and how to read a judgment.

Zero tolerance: the rule the Court laid down

Paragraph 6 contains the image that will follow this case around. The Court said that for those in the province of adjudication, the production of fake, non-existent and hallucinated material and its use as precedent in law is like the release of methyl isocyanate, the gas released at Bhopal in December 1984, in the province of law and justice. Such material, the Court added, not only contaminates but takes away the very lifeblood of judicial determination.

Paragraph 7 converts that into a rule with four limbs.

Courts must adopt a zero-tolerance mode. The rule covers producing, citing and using AI-generated precedents without verification. Note the three verbs. Producing is the person who generated the material. Citing is the person who put it before the court. Using is the adjudicator who relied on it. All three are inside the rule.

Citing without verification is advocate misconduct. The judgment says it plainly. This is not a warning about a future rule. It is a characterisation of conduct under the existing professional framework, which for advocates flows from the Advocates Act, 1961 and the Bar Council rules made under it.

Judicial reliance is a serious lapse. The Court used a different word for judges than for advocates but did not treat it as a lesser problem. The decision that results does not survive.

The zero tolerance binds both sides. The Court declared zero tolerance for the Bar as well as the Bench to cite, refer to, or rely on such material. The tribunal here was the source of the fabrications, not counsel.

The reasoning runs through paragraphs 3 to 5. The Court accepted that courts have absorbed technology before without difficulty, but distinguished AI on the ground that it is not just an aid but an alternative to our own thinking, reasoning and even decision making. It said that if thinking is delegated and it forms a habit, the consequences reach the core of human existence, which lies in the capacity to think. Paragraph 5 then conceded the limits of judicial power: this is not an issue that can be resolved through judicial orders and declaratory judgments, but only through public policy and enforceable rules and regulations.

No decision in the eyes of the law

The remedial holding has the widest reach, because it applies to orders already passed. The Court declared that a decision resting on fake or hallucinated material is no decision in the eyes of the law, irrespective of whether such material had a direct or indirect bearing on the decision-making. It then set the threshold at its lowest possible point: such decisions are to be set aside even if an iota of fake or hallucinated material enters the decision-making process, as it would violate the sanctity of adjudication.

That formulation removes the argument that would otherwise be made in every one of these cases. Ordinarily, a party defending an order argues that the impugned reasoning is severable, that the conclusion rests on other material, and that a single bad citation is harmless error. Paragraph 7 forecloses it. The test is not whether the fabricated authority changed the result. The test is whether it entered the process. Paragraph 17 restates that as a rule about the rule of law rather than about error correction: a decision resting on materials which are unreal, fake, and do not exist at all is unsustainable and has to be set aside at the earliest.

The operative order followed. At paragraph 18 the Court set aside the NCLT order of 28 August 2024 and the NCLAT judgment of 11 September 2025, and restored the Section 7 application, RCP (IB) 6/MB/2023, to its original number. Paragraph 19 makes clear that the Court expressed no opinion on the merits. Paragraph 20 directs the NCLT to dispose of the restored application expeditiously, preferably within two weeks, with the parties maintaining status quo until then.

Notice the cost. An admission order from August 2024 and an appellate judgment from September 2025 were both undone, and a Section 7 application filed in 2023 went back to the start, because nobody opened six citations. The interim resolution professional’s appointment and the moratorium fell with it. For the wider setting, see AI in Indian courts and the Supreme Court digest for June 2026.

What the Court asked the Bar Council of India to do

Paragraph 8 is one sentence and it is the hinge of the judgment: mere declaration of prohibitory action is not sufficient, and there must be a consequential action following accountability.

Paragraph 9 supplies the machinery. The Court directed the Bar Council of India, being the apex statutory body, to constitute a committee and deliberate on the issue of members of the bar submitting fake and hallucinated material before the court as if they are precedents of law. The Bar Council was asked to take up the issue with utmost seriousness, deliberate earnestly, and prescribe two things: a guiding principle to prevent such occurrences, and the disciplinary action that will follow a violation of the norms.

Three points about the scope of that direction.

It is addressed to the Bar Council of India specifically, not to State Bar Councils, and it names the Bar Council’s statutory character as the reason. Disciplinary jurisdiction over advocates sits within the Bar Council structure under the Advocates Act, 1961, which is why the direction goes there rather than to the courts’ own contempt jurisdiction.

It is a direction to deliberate and prescribe, not to punish anyone in this case. No advocate was proceeded against, because the fabrications came from the tribunal’s own research.

It leaves the shape of the norms to the Bar Council. As at 2 August 2026, no committee report or Bar Council norm has been published. Until one is, the operative standard for advocates is the one stated in paragraph 7: citing an AI-generated judgment without verification is misconduct. That standard does not wait for a rule to be framed, which is the same point made in the lawyer’s duty to verify AI output.

The February 2026 order that set this up

The same two judges had taken the issue up months earlier, in a matter that is still pending.

In Gummadi Usha Rani & Anr. v. Sure Mallikarjuna Rao & Anr., the bench of Justice Narasimha and Justice Aradhe took cognisance of a trial court order that relied on decisions which did not exist. The underlying dispute was a property matter in Andhra Pradesh in which an advocate commissioner had been appointed to note the physical features of the property, and the trial court dismissed objections to the commissioner’s report by an order dated 19 August 2025. As reported by SCC OnLine, the Supreme Court’s order of 27 February 2026 recorded that a decision based on such non-existent and fake alleged judgments is not an error in the decision making, and that it may amount to misconduct with legal consequences to follow.

The Court did not stop at an observation. It issued notice to the Attorney General for India, the Solicitor General and the Bar Council of India, and appointed Senior Advocate Shyam Divan as amicus curiae, as Verdictum reported. That is the structure a court builds when it intends to lay down a general framework rather than dispose of one appeal.

Read together, the two matters explain the shape of the July judgment. Gummadi Usha Rani is the vehicle for the general framework. Pooja Ramesh Singh is the appeal that had to be decided in the meantime, and the Court used it to state the operative rule and to put the Bar Council to work.

The Indian record before Pooja Ramesh Singh

The Supreme Court’s word “again” is carrying a real list. These are decided or reported matters, and they run through every level of the system.

A tribunal recalled its own order. The Bengaluru bench of the Income Tax Appellate Tribunal passed an order on 30 December 2024 in Buckeye Trust v. PCIT, ITA No. 1051/Bang/2024, relying on authorities that did not exist. The tribunal recalled the order within about a week. The original order remains on the ITAT’s public record.

A High Court quashed a tax assessment. On 6 October 2025 the Bombay High Court quashed a faceless income tax assessment that had raised a company’s assessed income to Rs 27.91 crore on the strength of non-existent, AI-generated case law.

A High Court threw out a petition drafted with a chatbot. A single judge of the Delhi High Court, hearing a homebuyers’ petition, was told by senior counsel that the petition leaned on ChatGPT-generated material. It quoted paragraphs 73 and 74 of Raj Narain v. Indira Nehru Gandhi. The real judgment has 27 paragraphs. The court treated the use of AI in that manner as impermissible and allowed the plea to be withdrawn.

A High Court imposed costs. In Deepak s/o Shivkumar Bahry v. Heart & Soul Entertainment Ltd., 2026:BHC-AS:828, Justice M.M. Sathaye of the Bombay High Court dealt with written submissions that cited a judgment titled Jyoti w/o Dinesh Tulsiani v. Elegant Associates, which neither the judge nor the court’s law clerks could trace. He imposed costs of Rs 50,000 payable to the High Court Employees Medical Fund, and said that if an AI tool is used in aid of research it is welcome, but there is a great responsibility on the party to cross verify the references. The order text is hosted by Bar and Bench.

LiveLaw’s survey of phantom precedents in Indian courts collects further instances, including a property dispute in which the Andhra Pradesh High Court flagged four non-existent Supreme Court judgments as AI-generated.

The overseas record was in front of the Court too. Paragraph 2 notes that the Solicitors Regulation Authority approved in 2025 the first purely AI-driven law firm, Garfield Law Limited. Footnote 7 records the other side: in Anthony Malcolm Cork & Anor v. Mark Smith, [2026] EWHC 1199 (Ch), a junior associate at Pinsent Masons LLP accepted a fabricated statutory quote produced by the firm’s internal AI program without checking it, the error passed supervising partners, and the firm then sent the court a misleading explanation. The judge declined to start contempt proceedings for want of deliberate dishonesty but admonished the firm publicly. The originating incident is older still: in June 2023 the United States District Court for the Southern District of New York sanctioned two attorneys in Mata v. Avianca, Inc. after they filed a brief citing six cases produced by ChatGPT, none of which existed.

What the judgment deliberately does not do

It is easy to read Pooja Ramesh Singh as a judgment against AI in legal work. It is not, and the Court said so twice. Paragraph 1 states the Court’s resolve to adopt AI technology in aid of adjudication, while at the same time asserting and declaring total and absolute control over adjudication, with a human in the loop at every stage. Paragraph 7 closes with a clarification that most of the coverage has missed: the judgment shall have no bearing on the rightful use of AI, but on the presentation of or reliance on fake or hallucinated material as if it were a court precedent. Using an AI research tool is not the wrong. Putting unverified output in front of a court as though it were law is.

Four other limits are worth stating precisely, because overstating this judgment will cost you credibility with a bench.

It imposed no costs and initiated no proceedings against any advocate. The fabrications came from the tribunal’s own research, established by the respondent bank’s affidavit, so there was no advocate to proceed against.

It did not name or censure the individual members of the NCLT or NCLAT. The Court characterised judicial reliance as a serious lapse, set the orders aside, and left it there.

It did not lay down a disclosure requirement. Nothing in the judgment requires a party to declare that AI was used. That duty appears in the draft regulations, not in this case, and the difference is set out in AI disclosure in pleadings.

It did not prescribe a verification method. The Court required verification and left the mechanics to the profession and to the Bar Council committee. What a workable method looks like is set out next and in how to vet legal AI for citation accuracy.

The verification duty in practice

The rule from paragraph 7 is short. Applying it to a brief with thirty authorities on a Sunday night is where it gets real. Start by separating two questions that most people collapse into one. Does this case exist, and does this passage exist in it. Three of the six citations in this case failed the first question and would be caught by a name search. Two failed only the second, and would pass any check that stopped at confirming the case is real.

The workflow below is the minimum that catches every failure mode present in Pooja Ramesh Singh.

flowchart TD
    A[Authority appears in your draft] --> B{Found in a law report or on the court portal?}
    B -->|No| C[Delete the citation and the argument resting on it]
    B -->|Yes| D{Party names match the citation you were given?}
    D -->|No| E[Recheck the neutral citation, correct the cause title]
    D -->|Yes| F[Open the full text of the judgment]
    E --> F
    F --> G{Quoted paragraph exists in those words?}
    G -->|No| C
    G -->|Yes| H{Paragraph says what your draft claims?}
    H -->|No| C
    H -->|Yes| I[Run a good law check on later treatment]
    I --> J{Overruled or doubted?}
    J -->|Yes| K[Cite the later authority instead]
    J -->|No| L[Safe to file, record the check]

Some practical notes on running it.

Search by citation as well as by name. The State Bank of India entry shows why. The reporter reference was real and pointed at a different judgment, so a name-only search returned nothing and looked like a typing error.

Open the report, do not trust a summary. Everest Kento and Canara Bank would survive any check that read a headnote or a database abstract, because the fabricated text was at paragraph level. Machine summaries carry their own risks, covered in AI judgment summarisation.

Check the paragraph count. The Delhi High Court petition quoted paragraphs 73 and 74 of a judgment with 27 paragraphs. A paragraph number higher than the judgment has is a fast and reliable signal.

Prefer the neutral citation. Neutral citations issued by the court itself tie an authority to a document you can retrieve rather than to a reporter volume you have to trust, as explained in e-SCR and neutral citations. Where the matter is important enough, a certified copy of the judgment removes any remaining doubt.

Check later treatment. A real, correctly quoted case that has been overruled is as useless in argument as a fabricated one, and more embarrassing, because opposing counsel will have checked. The method is in good law checking.

Do not ask the tool to confirm its own output. A model that fabricated a citation will confirm it when asked. That was the second error in Mata v. Avianca, and it is structural rather than a bug in one product, as set out in native legal AI versus generic chatbots and ChatGPT for lawyers in India.

Record what you checked. A contemporaneous note of the authorities you opened and when is the best answer available if a bench later asks how a citation reached your written submissions.

How much of that workload lands on you by hand depends on where the research came from.

Verification stepRetrieval-grounded legal research toolGeneral-purpose chatbot
Citation traceable to a source document✓ Linked to the judgment it came from✗ Generated from training patterns
Quoted paragraph checkable in one click✓ Yes, open the judgment✗ No source to open
Later treatment of the authority✓ Available through a citator✗ Not available
Behaviour when asked to confirm its own citation✓ Points you at the document✗ Confirms the fabrication
Coverage of Indian High Court and Supreme Court law✓ Built for it✗ Uneven and unstated

A general chatbot hands you a citation and nothing to open, so every line has to be reconstructed from scratch. A system that answers from real Indian judgments and links each answer to the case it came from, which is what Niyam is built to do, turns the check into a click. It does not remove the duty the Supreme Court described in paragraph 7. It makes discharging that duty a matter of minutes, which is the difference between a rule people follow and a rule people skip. A wider survey sits in the best AI legal research tools in India and choosing an Indian case law search engine.

What tribunals, judges and law clerks have to change

This judgment is unusual in Indian AI case law because the fabrication did not come from a litigant. Paragraph 16 records the respondent bank’s affidavit that the alleged judgments were not cited by its counsel at the bar and were obtained through the adjudicating authority’s own research. The tribunal researched, the tribunal cited, and the appellate tribunal did not catch it.

The Court’s observation on that point deserves attention from anyone who supervises research. It said today’s courts and tribunals implicitly trust lawyers when referring to precedents cited before them, and asked the reader to imagine the hardship of a situation in which the court must verify the authenticity of each judgment cited by an advocate. That trust is the operating assumption of the adversarial system. Once a bench cannot assume a cited case is real, the cost of every hearing rises.

Three consequences follow for adjudicating authorities and their staff.

Research done in chambers needs the same verification as research filed by counsel. In this case the appellate tribunal treated the NCLT’s citations as reliable because they appeared in a tribunal order. Nothing checked them. The paragraph 7 rule covers using such material, not merely citing it, so chamber research has to arrive in a form a supervising officer can audit. Research run through a tool such as Niyam, which ties each proposition back to the judgment it was drawn from, leaves that trail without anyone repeating the search.

Appellate scrutiny is not a safety net. The Court pointed out that the fake, non-existent judgments escaped scrutiny by the first statutory appellate tribunal, which reproduced the citations in paragraph 12 of its own opinion. Reproduction is not verification.

Law clerks and research assistants need explicit instructions. The Kerala High Court’s policy on AI tools in the district judiciary, issued in July 2025, applies expressly to judicial officers, the staff assisting them, and interns and law clerks, and provides for disciplinary consequences on breach. It remains the clearest institutional model in the country for who is covered and what happens when the rule is broken.

How this sits against the draft AI regulations for courts

Two things are now in play and they do different work. Mixing them up is the fastest way to misstate the position to a court.

The Supreme Court’s AI Committee released the Draft Regulations for Use of Artificial Intelligence in Courts, 2026 on 3 June 2026 and invited public comments. A clause-by-clause reading is available from The Leaflet. The draft contains a disclosure duty on parties and counsel who used AI in preparing or summarising a document, a power for the court to ask which system was used and what verification was done, and a responsibility provision that refuses to let the AI character of the output work as a defence when the content turns out to be fabricated. Paragraph 5 of Pooja Ramesh Singh notes that the regulations will be notified in due course.

The judgment is different in kind. It is a binding declaration of law from a two-judge bench, delivered in a decided appeal, and it operates now.

QuestionPooja Ramesh Singh, 2 July 2026Draft AI Regulations, 3 June 2026
Binding today✓ Yes, decided and reportable✗ No, draft released for comment
Requires verification of every authority✓ Yes, express✓ Yes, through the responsibility provision
Requires disclosure that AI was used✗ No✓ Yes, at the time of submission
Names citing without verification as misconduct✓ Yes, for advocates✗ Not in those terms
Voids the resulting order✓ Yes, even on an iota of fake material✗ Not addressed
Directs a disciplinary framework✓ Yes, to the Bar Council of India✗ Not a Bar Council instrument
Applies to tribunals✓ Yes, on these very facts✓ Yes, by its stated scope

The practical position for an advocate filing next week is the left column plus prudence about the right. Verify, because the Supreme Court has said failing to verify is misconduct. Consider disclosing, because the direction of travel is settled even if the instrument is not notified. The institutional framework is set out in the Supreme Court’s AI rules for India, and what those rules demand of the software is covered in what AI tools need to meet the Supreme Court rules.

Frequently asked questions

What is the citation for the Supreme Court’s 2026 judgment on AI-generated case law?

The case is Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. & Anr., Civil Appeal No. 11950 of 2025, neutral citation 2026 INSC 668, decided on 2 July 2026 by Justice Pamidighantam Sri Narasimha and Justice Alok Aradhe. It is reported at 2026 LiveLaw (SC) 653. The judgment is marked reportable and the full text is published on the Supreme Court of India’s judgment portal.

Did the Supreme Court ban the use of AI by lawyers or judges?

No. Paragraph 7 states expressly that the judgment has no bearing on the rightful use of AI, only on presenting or relying on fake or hallucinated material as if it were a court precedent. Paragraph 1 records the Court’s resolve to adopt AI technology in aid of adjudication while asserting absolute control over adjudication with a human in the loop at every stage.

Is citing an AI-generated fake judgment professional misconduct in India?

The Supreme Court said in paragraph 7 that it is misconduct on the part of an advocate to cite such judgments without verification. That characterisation applies now, under the existing professional framework. The Court separately directed the Bar Council of India to constitute a committee and prescribe both a guiding principle and the disciplinary action that will follow a violation, so the procedural consequences are still being worked out.

What happens to an order that relied on a fake citation?

It is liable to be set aside. The Court held that such a decision is no decision in the eyes of the law, irrespective of whether the fake material had a direct or indirect bearing on the outcome, and that the order goes even if only an iota of fake or hallucinated material entered the decision-making process. Here, both the NCLT order of 28 August 2024 and the NCLAT judgment of 11 September 2025 were set aside.

Who created the fake citations in Pooja Ramesh Singh?

Not counsel. Paragraph 16 records that the first respondent filed an affidavit stating the alleged judgments were not cited by its counsel at the bar and were obtained through the adjudicating authority’s own research. The fabricated authorities originated in the NCLT and were then reproduced by the NCLAT in paragraph 12 of its opinion without being checked.

Can a citation be real and still be fabricated?

Yes, and this is the harder failure to catch. Two of the six authorities in this case, Everest Kento Cylinders and Canara Bank v. N.G. Subbaraya Setty, are genuine Supreme Court judgments correctly cited. The paragraphs attributed to them do not exist. Confirming that a case is real does not confirm that the quoted passage is real, so verification has to reach paragraph level.

Does this judgment require me to disclose that I used AI?

No. Pooja Ramesh Singh imposes a verification duty, not a disclosure duty. The disclosure obligation appears in the Draft Regulations for Use of Artificial Intelligence in Courts, 2026, released by the Supreme Court’s AI Committee on 3 June 2026 for public comment. Those regulations are not notified and are therefore not binding. Disclosing anyway is prudent, but it is not what this judgment requires.

Were costs imposed in this case?

No. The Supreme Court set aside both tribunal orders, restored the Section 7 application to its original number, and directed the NCLT to dispose of it expeditiously, preferably within two weeks, with status quo maintained in the meantime. No costs were imposed and no proceedings were initiated against any advocate, because the fabrications came from the tribunal’s own research rather than from a party.

Has any Indian court imposed costs for AI-fabricated citations?

Yes. In Deepak s/o Shivkumar Bahry v. Heart & Soul Entertainment Ltd., 2026:BHC-AS:828, Justice M.M. Sathaye of the Bombay High Court imposed costs of Rs 50,000 payable to the High Court Employees Medical Fund after written submissions cited a judgment titled Jyoti w/o Dinesh Tulsiani v. Elegant Associates that neither the judge nor the court’s law clerks could trace.

Does this apply to tribunals as well as courts?

It applies directly to tribunals, because the orders set aside were those of the NCLT and the NCLAT. The Court’s language covers courts and adjudicating authorities alike, and paragraph 17 refers to a decision of a court or an adjudicating authority based on fake and hallucinated material. Quasi-judicial bodies deciding rights on evidence and precedent are inside the rule.

What should I do if I discover a fake citation in something already filed?

Act immediately and disclose to the court rather than waiting to be caught. Identify how the citation entered the document and check every other authority in the same filing, because a single fabrication suggests the research was not verified. Seek guidance from a senior colleague on the correct procedural step, whether that is an application to amend, a corrigendum, or a direct submission at the next hearing.

No tool discharges the duty. The Supreme Court’s rule is that the person citing an authority must have verified it, and that responsibility does not move to software. What a retrieval-grounded tool changes is the cost of compliance: if every proposition comes with the judgment it was drawn from, checking the citation and the paragraph takes a click rather than a search. The obligation to open the judgment remains yours.