Author(s): Medha Arora, Bhavya Sharma, Monalisa Chaudhari
Paper Details: Volume 4, Issue 4
Citation: IJLSSS 4(4) 09
Page No: 84 – 95
ABSTRACT
The integration of generative artificial intelligence into legal research, drafting, and adjudication has generated a novel category of epistemic risk: hallucinated legal authorities, fabricated precedents, invented statutory provisions, and non-existent citations presented with the same rhetorical confidence as genuine authority. The Supreme Court of India’s decision in Pooja Ramesh Singh v. Jammu & Kashmir Bank Ltd. marks the country’s first authoritative judicial response to this challenge, adopting a “zero-tolerance” approach and directing the Bar Council of India to formulate regulatory safeguards. This paper argues that the judgment’s significance extends far beyond the facts of a single insolvency dispute. Situating it within a dense pre-existing pattern of Indian tribunal decisions and a comparative body of American, English, Canadian, Singaporean, and Australian jurisprudence, the paper contends that fabricated legal authorities are not isolated technological errors but manifestations of a broader crisis of legal authenticity, professional responsibility, and institutional accountability. It argues that an exclusively punitive zero-tolerance standard is insufficient unless accompanied by systematic verification protocols, disclosure norms, judicial training, and institutional reform, and proposes a five-pillar Indian AI Citation Verification Framework (“IACVF”) designed to preserve the efficiency benefits of AI-assisted legal research while safeguarding accuracy, transparency, fairness, and judicial independence.
Keywords: Artificial Intelligence, Hallucinated Citations, Judicial Integrity, Rule of Law, Professional Ethics, Bar Council of India, Human-in-the-Loop, Comparative Procedure, Constitutional Adjudication
INTRODUCTION
Generative AI systems are now routinely used by advocates, judicial clerks, tribunal researchers, and academics to summarise precedent, draft pleadings, and conduct legal research. Unlike a search engine, which retrieves existing material, a large language model generates text probabilistically, predicting linguistic patterns rather than retrieving verified fact.[1] The consequence is a well-documented failure mode: such systems generate citations, quotations, and entire judgments that do not exist, presented with the same fluency and confidence as genuine authority. A 2024 empirical study found legal hallucination rates of between 69% and 88% across widely used models when tested on legal questions, a finding its authors described as cautioning against the unsupervised integration of large language models into legal tasks.[2]
Unlike routine research errors, the risks identified here represent a qualitatively different category of threat. A misremembered citation is usually identifiable through routine scrutiny; a fabricated one is often stylistically indistinguishable from the real thing, and correspondingly more likely to survive verification in high-volume litigation environments. Where such material enters a judicial record unchecked, the harm is not confined to professional embarrassment: it corrodes the evidentiary and doctrinal foundation on which the decision itself rests, and, because the doctrine of precedent depends on an authoritative, verifiable body of decisions, threatens the architecture of legal certainty that common law systems are built on.
India entered this discourse decisively in Pooja Ramesh Singh v. Jammu & Kashmir Bank Ltd., in which the Supreme Court set aside orders of the National Company Law Tribunal (“NCLT”) and National Company Law Appellate Tribunal (“NCLAT”) after discovering that the Tribunal itself, not merely counsel appearing before it, had relied on non-existent, AI generated precedents. Read in isolation, the case looks like a narrow insolvency correction. Read against the pattern this paper assembles four prior Indian tribunal and High Court Cases, and a materially identical set of problems already litigated in the United States, England, Canada, Singapore, and Australia, it is better understood as the moment an already developing crisis of legal authenticity crystallised into binding Indian doctrine. By positioning Pooja Ramesh Singh within this wider context, this study contends that mere judicial censure, regardless of its severity, is insufficient without the implementation of a comprehensive systemic verification framework.
AI HALLUCINATIONS AND THE EPISTEMIC CRISIS OF LEGAL AUTHORITY
Legal research has traditionally rested on an epistemic model grounded in identifiable primary sources and institutional mechanisms of verification: a citation could, in principle, always be checked against a reporter. Generative AI disrupts this model by replacing retrieval with prediction. A model’s architecture prioritises fluency over evidentiary certainty, so it may generate a judicial decision bearing a realistic case name, citation, and doctrinal proposition despite the complete non-existence of the underlying authority.[3] Rather than a transient software bug amenable to patching, this issue represents an inherent structural feature of next-token prediction paradigms. This vulnerability is further compounded by historical evaluation benchmarks that disproportionately reward confident fabrication over justified abstention..[4]
A parallel debate in the research-integrity literature is instructive here. Scholars have argued that hallucinated citations may constitute research misconduct under governing scientific integrity frameworks where the citations function as data supporting a scholarly finding and the author is indifferent to the risk of fabrication.[5] The analogy to adjudication is direct: a judgment, like a scholarly paper, is a document whose legitimacy depends on the authenticity of what it cites in support of its conclusions. If data fabrication vitiates scientific research, then pseudo-precedent must, a fortiori, vitiate legal adjudication. This is the precise logic Pooja Ramesh Singh later operationalised, even if it did not explicitly formalise it as such.
The legal implication is that hallucinations are not merely citation-accuracy problems. Because common law legitimacy depends on demonstrable continuity between decisions, a fabricated authority severs that chain, threatens the principle that like cases be decided alike, but also implicates due process and the rule of law, not simply professional competence.
COMPARATIVE JURISPRUDENCE: A GLOBAL PATTERN
A. United States. The first widely reported instance was Mata v. Avianca, Inc., in which counsel submitted a brief containing fictitious ChatGPT-generated opinions and continued to affirm their authenticity after opposing counsel questioned them; the Southern District of New York imposed monetary sanctions, holding that technological innovation does not dilute the duty of candour to the court.[6] The case is now the baseline precedent in a rapidly expanding cross-jurisdictional body of law: a continuously updated database recorded, by mid-2026, over thirteen hundred court proceedings across more than a hundred countries involving AI-hallucinated material, with several hundred attorneys sanctioned or referred for discipline.[7]
B. England and Wales. In R (Ayinde) v. London Borough of Haringey, heard with Al-Haroun v. Qatar National Bank QPSC, the Divisional Court dealt with two unrelated instances of fabricated authority; the President of the King’s Bench Division called the conduct “wholly improper” and referred the practitioners to their regulators, while declining to treat the underlying pleadings as automatically void, professional consequence was addressed through the costs jurisdiction and regulatory referral, not automatic invalidation.[8] This graduated, culpability-sensitive posture, distinguishing deliberate fabrication from careless but honest reliance, has since been applied consistently by the English courts and now anchors judicial AI guidance across England and Wales.[9]
C. Canada. Zhang v. Chen, the first reported Canadian decision on the point, involved counsel citing two ChatGPT-fabricated cases in a family law application; the British Columbia Supreme Court declined special costs after accepting the lawyer’s conduct was not intended to deceive, but held that citing fake cases is an abuse of process capable of causing a miscarriage of justice, and ordered costs against counsel personally.[10] Ontario has since gone further than judge-made doctrine: amendments to the Rules of Civil Procedure now require a factum to carry a signed certification of the authenticity of every cited authority, a requirement invoked in Ko v. Li to refer counsel to a show-cause hearing after AI-fabricated citations survived uncorrected through to judgment.[11] The Federal Court of Canada has separately required a formal declaration whenever content is AI-generated rather than merely AI-assisted.[12]
D. Singapore. Singapore’s approach is more clearly governance-oriented than case-driven. The Supreme Court’s Registrar’s Circular No. 1 of 2024, the “Guide on the Use of Generative Artificial Intelligence Tools by Court Users”, applies across the Supreme Court, State Courts, and Family Justice Courts, takes a neutral stance on GenAI use, but places the burden of verifying every AI-provided citation against an authoritative repository squarely on the court user, without imposing a general disclosure requirement.[13] This model, regulation through institutional design and verification protocol rather than reactive sanction, offers a template distinct from the American and Canadian sanctions-driven approach.
E. Australia. Re Dayal saw a Victorian solicitor become the first Australian lawyer professionally sanctioned for tendering AI-fabricated citations, after admitting he had not verified the content; the court articulated duties not to mislead the court and to deliver services competently regardless of the tool used.[14] In Luck v. Secretary, Services Australia, a self-represented litigant cited a fabricated case to support a recusal application, illustrating that the risk extends beyond the represented Bar.[15]
Read together, these five jurisdictions reveal broad convergence on one point: verification is a non-delegable, human duty irrespective of the tool used, but real divergence on mechanism: sanctions-led in the United States and Australia, certification-led in Canada, guidance-led in Singapore, and graduated-culpability-led in England. This divergence is the comparative backdrop against which Pooja Ramesh Singh‘s uncompromising standard should be read.
THE INDIAN TRAJECTORY: POOJA RAMESH SINGH IN CONTEXT
Pooja Ramesh Singh did not arise on a blank slate. In March 2025, the Karnataka High Court directed an inquiry into a Bengaluru City Civil Court judge who had relied on two non-existent Supreme Court decisions to reject a return-of-plaint application, the first Indian signal that judicial reliance on fabricated precedent, not merely counsel’s citation of it, could attract institutional scrutiny.[16] In December 2024, the Bengaluru bench of the Income Tax Appellate Tribunal had already been compelled to recall its own order in the Buckeye Trust matter, a ₹669 crore trust-taxation dispute, after discovering the order rested on three Supreme Court judgments and one Madras High Court ruling that did not exist.[17] In September 2025, a petition before the Delhi High Court in Greenopolis Welfare Association v. Narender Singh was withdrawn after opposing counsel showed it had quoted paragraphs purportedly drawn from Raj Narain v. Indira Nehru Gandhi, a judgment that in fact runs to only twenty-seven paragraphs.[18] Most consequentially, the Bombay High Court in KMG Wires Pvt. Ltd. v. National Faceless Assessment Centre quashed a ₹28 crore faceless tax assessment in October 2025 after finding it rested on three non-existent precedents, holding that a quasi-judicial authority cannot blindly rely on AI generated output and that its failure to independently verify, combined with denial of an opportunity to respond, breached natural justice.[19] Set against this, the Andhra Pradesh High Court’s January 2026 decision in Gummadi Usha Rani v. Sure Mallikarjuna Rao took the opposite, harmless error view, declining to set aside a trial court order resting on four fabricated Supreme Court decisions on the ground that the legal principle applied was otherwise correct.[20]
Against this inconsistent record, Pooja Ramesh Singh arose from insolvency proceedings under section 7 of the Insolvency and Bankruptcy Code, 2016, concerning Jammu and Kashmir Bank’s invocation of a corporate guarantee from Essel Infraprojects Ltd.[21] The NCLT admitted the application in August 2024; the NCLAT upheld it in September 2025, reproducing without scrutiny six authorities the NCLT had relied on, several of which, senior counsel demonstrated before the Supreme Court, corresponded to no real decision, while others attached fabricated passages to genuine citations.[22] Notably, the Bank’s own affidavit disclaimed responsibility, indicating the Tribunal had sourced the fabricated material through its own research process.[23]
The Bench held that reliance on such material, irrespective of source, vitiates the entire adjudicatory exercise: a decision resting on even an “iota” of fabricated material is “no decision in the eyes of the law.”It declined to distinguish fabrication introduced by counsel from fabrication generated by the Tribunal’s own process, treating both as equally serious, and directed the Bar Council of India to constitute a committee to design preventive guidelines and disciplinary consequences for advocates, while clarifying that the ruling casts no doubt on AI’s legitimate use as a research aid, provided human oversight is retained throughout.[24] The judgment is, on this account, best understood not as a founding moment but as the point at which an already developing body of inconsistent Indian doctrine was resolved in favour of the most exacting standard yet articulated in any of the jurisdictions surveyed above a standard closer to a per se rule than to the harmless error or graduated culpability approaches that prevail abroad.
PROFESSIONAL RESPONSIBILITY IN THE AGE OF ARTIFICIAL INTELLIGENCE
Generative AI does not merely facilitate access to legal information; it actively produces legal content, transforming the lawyer’s role from researcher to verifier. The duty of competence, long understood to require adequate knowledge of applicable law, now necessarily includes technological competence, an understanding of what a given AI system can and cannot reliably do.[25] The duty of candour is equally implicated: courts rely on advocates to represent the state of the law accurately, and a hallucinated authority introduces a fictitious proposition under the appearance of genuine research, regardless of whether the advocate acted in bad faith.[26]
The comparative material canvassed above suggests professional responsibility is evolving in a consistent direction: from an assumption that AI generated output may be trusted, toward an affirmative, non-delegable duty to verify every AI generated authority against a recognised legal database before it is relied upon in pleadings, submissions, or academic work.[27] Disclosure is emerging as a complementary rather than a substitute obligation; Ontario’s certification requirement and Singapore’s court user framework both impose verification duties without necessarily requiring pre-emptive disclosure of AI use, suggesting that the two obligations (verify, and disclose) can be calibrated independently rather than treated as a single package.[28]
JUDICIAL ACCOUNTABILITY AND THE HUMAN-IN-THE-LOOP PRINCIPLE
The most doctrinally significant fact in Pooja Ramesh Singh is that the fabricated authorities entered judicial reasoning through the Tribunal’s own research process, not through counsel’s submissions.[29] This shifts the conversation from lawyer misconduct to judicial accountability. Judicial legitimacy rests on public confidence that outcomes emerge from independent, reasoned, and verifiable analysis; an AI system, which predicts text from statistical patterns rather than exercising legal reasoning, cannot itself satisfy that requirement. The “human-in-the-loop” principle, under which AI functions solely as a decision support tool while ultimate responsibility remains with a human decision maker who must independently assess, verify, and consciously adopt any AI generated proposition, has accordingly emerged as the dominant international standard, expressed in comparable terms across the Singaporean guide, the Canadian Federal Court notice, and the English judicial guidance surveyed above. India’s own Draft Regulations for Use of Artificial Intelligence in Courts, 2026, released a month before Pooja Ramesh Singh by the Supreme Court’s AI Committee, codify this principle domestically: it rests on five stated principle, primacy of human decision making, transparency, accountability, data protection, and judicial independence, and absolutely prohibits AI from deciding a case, determining bail, or evaluating witness credibility.[30] Notably, its mandatory disclosure regime for AI assisted filings applies to parties and representatives, but the Draft Regulations, unlike Pooja Ramesh Singh‘s own reasoning, are the instrument best placed to reach a tribunal’s internal research process; the gap the judgment itself does not close.[31]
CONSTITUTIONAL DIMENSIONS: RULE OF LAW, DUE PROCESS, AND NATURAL JUSTICE
Fabricated authorities implicate more than software reliability. The rule of law requires that power be exercised according to publicly ascertainable norms; precedent performs this function by ensuring consistency and predictability across cases, and a fictional authority is incapable of independent verification by definition, rendering the constitutional promise of equal treatment illusory wherever it influences an outcome.[32] Natural justice is equally engaged: meaningful participation in adjudication presupposes that a party can identify, challenge, and distinguish the authorities relied upon against it, which is impossible where those authorities are fictional, precisely the reasoning the Bombay High Court applied in KMG Wires in finding a breach of natural justice independent of the merits of the tax addition itself.[33] Finally, the requirement of a reasoned decision, a recurring theme in Indian constitutional adjudication, is undermined at its root where the reasoning offered rests on doctrinal analysis that does not, in fact, exist. Viewed through this lens, Pooja Ramesh Singh is not merely a procedural correction but a reaffirmation that judicial legitimacy depends on demonstrable fidelity to authentic authority rather than technological convenience.
WHY HALLUCINATIONS CANNOT BE ELIMINATED: THE TECHNICAL EVIDENCE
A common assumption is that hallucination is a temporary limitation that improving models will resolve. The evidence does not support this. Hallucination arises from the probabilistic architecture of large language models themselves: such systems are optimised to produce fluent, contextually plausible responses, not to verify that a cited decision exists, and empirical testing has found legal hallucination rates as high as 88% for some widely used models, with the added finding that these systems frequently cannot themselves predict when they are hallucinating. Fabrications tend to increase, moreover, precisely in the contexts where legal researchers most need reliable output, specialised, recent, or niche authorities.[34] The implication for policy is direct: because the failure mode is structural rather than incidental, the appropriate response is institutional governance and mandatory verification, not confidence that the next model generation will resolve the problem on its own.[35]
TOWARDS AN INDIAN AI CITATION VERIFICATION FRAMEWORK (IACVF)
Pooja Ramesh Singh‘s zero tolerance standard is a necessary but insufficient response, since judicial disapproval alone does not build the verification infrastructure the problem requires. This paper proposes a five pillar Indian AI Citation Verification Framework, distributing responsibility across every participant in the justice-delivery system rather than concentrating it solely on the Bar.
Pillar I: Mandatory Verification. Every AI generated authority should be cross checked against an authenticated repository (Supreme Court and High Court reports, SCC Online, Manupatra, or the official e-Courts database) before inclusion in a pleading, submission, or order. This obligation should extend explicitly to judicial clerks and tribunal researchers, not only to advocates, a lesson Pooja Ramesh Singh itself teaches, given that the fabricated material there originated with the Tribunal’s own process.[36]
Pillar II: Professional Disclosure. Rather than prohibiting AI use, advocates should certify, following the Ontario model, that every cited authority has been independently verified regardless of whether AI assisted its identification.[37] Disclosure should not, of itself, carry an adverse inference; the Singaporean model’s neutral stance is instructive here.[38]
Pillar III: Judicial AI Governance. Courts and tribunals require an internal AI policy distinct from advocate facing rules, covering permissible and prohibited uses, verification procedures, and record keeping. India’s Draft Regulations, 2026 already supply much of this architecture: an apex AI body, AI Committees and Secretariats in every High Court, a proposed Centre of Research and Excellence on Artificial Intelligence, and mandatory annual audits, and should, in final form, be extended explicitly to cover the internal research workflows of tribunals such as the NCLT, NCLAT, and ITAT.[39]
Pillar IV: Institutional Capacity Building. The National Judicial Academy, State Judicial Academies, and the Bar Council of India should mandate training on how large language models function, their hallucination risks, and verification techniques, mirroring the professional education response already underway in Canada and Singapore.[40]
Pillar V: Legal Education Reform. AI literacy, including the discipline of verifying rather than merely using AI output, should be incorporated into the core law school curriculum, so that the duty of verification becomes a professional reflex rather than a post-qualification compliance burden.
BEYOND ZERO TOLERANCE: FROM PUNISHMENT TO GOVERNANCE
An exclusively punitive standard risks two costs: it may discourage the disclosure of AI use, reducing transparency precisely where it is most needed, and excessive fear of disciplinary consequence may deter the legitimate efficiency gains AI can offer, particularly for under-resourced litigants and courts.[41]The comparative record bears this out, Singapore’s governance first model and England’s graduated, culpability-sensitive sanctions regime both aim at the same four objectives Pooja Ramesh Singh implicitly serves, encouraging responsible innovation, preserving judicial integrity, protecting litigants, and ensuring human accountability, without collapsing every unverified citation into presumptively equivalent misconduct.[42] Read as a floor rather than a template, Pooja Ramesh Singh‘s per se rule and the IACVF’s governance architecture are complementary rather than competing responses to the same underlying problem.
FUTURE CHALLENGES
The issues Pooja Ramesh Singh raises are unlikely to remain confined to citation fabrication. Emerging risks already visible abroad AI generated witness statements, fabricated documentary evidence, deepfake evidence, and increasingly autonomous drafting tools will test Indian courts well before comprehensive doctrine develops. The Draft Regulations, 2026 anticipate several of these risks by prohibiting AI from evaluating witness credibility or generating factual findings, but their final form, and their reach into tribunal level adjudication, remain open questions for further research.[43]
CONCLUSION
Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. is, on its facts, a narrow insolvency appeal remanded for fresh adjudication. Read against the comparative and Indian record assembled here, the Karnataka High Court’s disciplinary probe, the Buckeye Trust recall, the Delhi High Court’s withdrawn petition, the Bombay High Court’s quashing of a ₹28 crore tax assessment, the Andhra Pradesh High Court’s contrary and now effectively overruled leniency, and the parallel American, English, Canadian, Singaporean, and Australian jurisprudence, the judgment is better understood as the point at which an already developing crisis of legal authenticity was resolved, in India, in favour of the most exacting standard yet articulated anywhere. But zero tolerance is a floor, not a solution. Its durability will depend on whether the institutional architecture now under construction, the Bar Council’s disciplinary committee and the Supreme Court’s own Draft AI Regulations, extended, as this paper has argued, to reach the tribunal’s own research process, succeeds in converting judicial condemnation into the verification infrastructure the problem actually requires.
[1] Robin Emsley, ChatGPT: These Are Not Hallucinations – They’re Fabrications and Falsifications, 9 NPJ Schizophrenia 1 (2023).
[2] Matthew Dahl et al., Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models, arXiv:2401.01301 (2024).
[3] David B. Resnik & Mohammad Hosseini, Hallucinated Citations Produced by Generative Artificial Intelligence May Constitute Research Misconduct When Citations Function as Data in Scholarly Papers, Accountability in Research (2026), https://doi.org/10.1080/08989621.2026.2645390.
[4] Adam Tauman Kalai & Ofir Nachum, Why Language Models Hallucinate (OpenAI, 2025), https://cdn.openai.com/pdf/d04913be-3f6f-4d2b-b283-ff432ef4aaa5/why-language-models-hallucinate.pdf.
[5] David B. Resnik & Mohammad Hosseini, Hallucinated Citations Produced by Generative Artificial Intelligence May Constitute Research Misconduct When Citations Function as Data in Scholarly Papers, Accountability in Research (2026), https://doi.org/10.1080/08989621.2026.2645390.
[6] Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), 2023 WL 4114965 (S.D.N.Y. June 22, 2023).
[7] Damien Charlotin, AI Hallucination Cases Database (2026), https://www.damiencharlotin.com/hallucinations/; HAQQ, AI Hallucinations in Law: 1,313 Court Cases and Counting, https://haqq.ai/blog/when-ai-lies-to-the-court (2026).
[8] R (Ayinde) v. London Borough of Haringey; Al-Haroun v. Qatar National Bank QPSC [2025] EWHC 1383 (Admin); R (Ayinde) v. London Borough of Haringey [2025] EWHC 1040 (Admin).
[9] R (Ayinde) v. London Borough of Haringey; Al-Haroun v. Qatar National Bank QPSC [2025] EWHC 1383 (Admin); R (Ayinde) v. London Borough of Haringey [2025] EWHC 1040 (Admin).
[10] Zhang v. Chen, 2024 BCSC 285 (Can.).
[12] Federal Court of Canada, Notice to the Parties and the Profession: The Use of Artificial Intelligence in Court Proceedings (May 7, 2024, updated May 2024).
[13] Supreme Court of the Republic of Singapore, Registrar’s Circular No. 1 of 2024, Guide on the Use of Generative Artificial Intelligence Tools by Court Users (Sept. 23, 2024).
[14] Re Dayal (2024) 386 FLR 359 (Austl.).
[15] Luck v. Secretary, Services Australia [2025] FCAFC 26 (Austl.).
[16] Pooja Ramesh Singh v. Jammu & Kashmir Bank Ltd., 2026 INSC 668 (India) (Civil Appeal No. 11950 of 2025, decided July 2, 2026).
[17] Medianama, 10 Cases That Show Indian Courts Have an AI Hallucination Problem (July 2026), https://www.medianama.com/2026/07/223-10-cases-ai-hallucination-cases-in-indian-courts/; In re Buckeye Trust, I.T.A. (ITAT Bengaluru Bench, order dated Dec. 30, 2024, recalled under § 254(2), Income-tax Act, 1961).
[18] The Wire, Fake AI Cases ‘Entered’ an NCLT Insolvency Order. The Supreme Court Quashed It. But Who Is Accountable? (July 2026), https://m.thewire.in/article/law/fake-ai-cases-entered-an-nclt-insolvency-order-the-supreme-court-quashed-it-but-who-is-accountable.
[19] KMG Wires Pvt. Ltd. v. National Faceless Assessment Centre, Writ Petition (Bom. H.C., order dated Oct. 27, 2025) (Colabawalla & Jamsandekar, JJ.); Moneylife, AI ‘Hallucination’ in Tax Assessment: Bombay HC Sets Aside Rs27.91-Crore Order Passed on Non-existent Case Laws (Oct. 27, 2025), https://moneylife.in/article/ai-hallucination-in-tax-assessment-bombay-hc-sets-aside-rs2791crore-order-passed-on-nonexistent-case-laws/78673.html.
[20] Greenopolis Welfare Ass’n v. Narender Singh (Del. H.C., petition withdrawn Sept. 2025); Raj Narain v. Indira Nehru Gandhi, AIR 1975 SC 865.
[21] Tulip Kanth, Zero-Tolerance for Using AI-Generated Precedents Without Verification: Supreme Court Sets Aside NCLT, NCLAT Orders Citing Fake Judgments, Verdictum (July 2, 2026), https://www.verdictum.in/supreme-court/pooja-ramesh-singh-v-jammu-and-kashmir-bank-ltd-2026-insc-668-ai-generated-precedents-1616998.
[22] LiveLaw, Supreme Court Sets Aside NCLT Judgment for Using AI Hallucinated Citations, Asks BCI to Examine Issue (July 2026), https://www.livelaw.in/supreme-court/supreme-court-sets-aside-nclt-judgment-for-using-ai-hallucinated-citations-asks-bci-to-examine-issue-539616.
[23] The Wire, Fake AI Cases ‘Entered’ an NCLT Insolvency Order. The Supreme Court Quashed It. But Who Is Accountable? (July 2026), https://m.thewire.in/article/law/fake-ai-cases-entered-an-nclt-insolvency-order-the-supreme-court-quashed-it-but-who-is-accountable.
[24] Ibid at 22.
[25] Ko v. Li, 2025 ONSC 2766 (Can.); Ontario, Rules of Civil Procedure, R.R.O. 1990, Reg. 194, r. 4.06.1(2.1), as amended by O. Reg. 384/24.
[26] Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), 2023 WL 4114965 (S.D.N.Y. June 22, 2023).
[27] Ko v. Li, 2025 ONSC 2766 (Can.); Ontario, Rules of Civil Procedure, R.R.O. 1990, Reg. 194, r. 4.06.1(2.1), as amended by O. Reg. 384/24.
[28] Ibid.
[29] Ibid at 23.
[30] Supreme Court of India, Draft Regulations for Use of Artificial Intelligence in Courts, 2026 (June 3, 2026), https://cdnbbsr.s3waas.gov.in/s3ec0490f1f4972d133619a60c30f3559e/uploads/2026/06/2026060342.pdf.
[31] Lawbeat, Supreme Court Releases Draft AI Rules for Courts; Lawyers Must Disclose Use of AI in Pleadings (June 4, 2026), https://lawbeat.in/top-stories/supreme-court-releases-draft-ai-rules-for-courts-lawyers-must-disclose-use-of-ai-in-pleadings-1598628.
[32] Kashmir Observer, SC Warns Against AI-Generated Fake Legal Precedents (July 2, 2026), https://kashmirobserver.net/2026/07/02/sc-warns-against-ai-generated-fake-legal-precedents/.
[33] KMG Wires Pvt. Ltd. v. National Faceless Assessment Centre, Writ Petition (Bom. H.C., order dated Oct. 27, 2025) (Colabawalla & Jamsandekar, JJ.); Moneylife, AI ‘Hallucination’ in Tax Assessment: Bombay HC Sets Aside Rs27.91-Crore Order Passed on Non-existent.
[34] Matthew Dahl et al., Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models, arXiv:2401.01301 (2024).
[35] Adam Tauman Kalai & Ofir Nachum, Why Language Models Hallucinate (OpenAI, 2025), https://cdn.openai.com/pdf/d04913be-3f6f-4d2b-b283-ff432ef4aaa5/why-language-models-hallucinate.pdf.
[36] The Wire, Fake AI Cases ‘Entered’ an NCLT Insolvency Order. The Supreme Court Quashed It. But Who Is Accountable? (July 2026), https://m.thewire.in/article/law/fake-ai-cases-entered-an-nclt-insolvency-order-the-supreme-court-quashed-it-but-who-is-accountable.
[37] Ko v. Li, 2025 ONSC 2766 (Can.); Ontario, Rules of Civil Procedure, R.R.O. 1990, Reg. 194, r. 4.06.1(2.1), as amended by O. Reg. 384/24.
[38] Supreme Court of the Republic of Singapore, Registrar’s Circular No. 1 of 2024, Guide on the Use of Generative Artificial Intelligence Tools by Court Users (Sept. 23, 2024).
[39] Supreme Court of India, Draft Regulations for Use of Artificial Intelligence in Courts, 2026 (June 3, 2026), https://cdnbbsr.s3waas.gov.in/s3ec0490f1f4972d133619a60c30f3559e/uploads/2026/06/2026060342.pdf.
[40] Ko v. Li, 2025 ONSC 2766 (Can.); Ontario, Rules of Civil Procedure, R.R.O. 1990, Reg. 194, r. 4.06.1(2.1), as amended by O. Reg. 384/24.
[41] Lawbeat, Supreme Court Releases Draft AI Rules for Courts; Lawyers Must Disclose Use of AI in Pleadings (June 4, 2026), https://lawbeat.in/top-stories/supreme-court-releases-draft-ai-rules-for-courts-lawyers-must-disclose-use-of-ai-in-pleadings-1598628.
[42] R (Ayinde) v. London Borough of Haringey; Al-Haroun v. Qatar National Bank QPSC [2025] EWHC 1383 (Admin); R (Ayinde) v. London Borough of Haringey [2025] EWHC 1040 (Admin).
[43] Supreme Court of India, Draft Regulations for Use of Artificial Intelligence in Courts, 2026 (June 3, 2026), https://cdnbbsr.s3waas.gov.in/s3ec0490f1f4972d133619a60c30f3559e/uploads/2026/06/2026060342.pdf.