Introduction
Walk into any district court registry today and you will find case bundles that look identical to the ones filed five years ago, except for one thing: a growing number of them cite judgments that were never delivered. Not misquoted judgments, not misapplied ones, but judgments that simply do not exist. This is the strange, very present cost of artificial intelligence entering Indian legal research faster than the profession has built the habits to check it. The Supreme Court, several High Courts, and even a tax tribunal have all encountered the same phenomenon in the space of about two years: AI tools inventing case law with the same confident formatting as a genuine citation. This article looks specifically at that problem , AI hallucination in judicial and litigation research , as the sharpest illustration of why India’s embrace of legal-research AI cannot proceed without a firm, enforceable structure of human verification.
Building the Machinery: How Indian Courts Are Using AI
India’s judiciary has adopted AI through the e-Courts Mission Mode Project Phase III (2023), which integrates artificial intelligence, machine learning, optical character recognition, and natural language processing into court administration.[1] Key initiatives include SUPACE, an AI-assisted legal research tool for judges,[2] SUVAS, which has translated more than 36,000 Supreme Court judgments into Indian languages,[3] and newer applications such as LegRAA and TERES for research assistance and real-time transcription.[4]
These initiatives respond to India’s mounting judicial backlog, which exceeded 5.4 crore pending cases nationwide, including more than 87,000 before the Supreme Court as of July 2025.[5] AI-assisted tools reduce the time spent on repetitive administrative tasks while preserving judicial decision-making through a human-in-the-loop approach, reflected in the Supreme Court’s adoption of real-time transcription for Constitution Bench hearings.[6]
When the Machine Invents: The Hallucination Problem
The same generative AI that efficiently summarizes judgments can also fabricate convincing but non-existent legal authorities, a phenomenon known as AI hallucination, where language models generate plausible responses without verifying factual accuracy.[7]
Indian courts have repeatedly encountered fabricated AI-generated citations. In December 2024, the Bengaluru Income Tax Appellate Tribunal recalled a ₹669 crore tax order after discovering it relied on fictitious Supreme Court and Madras High Court decisions.[8] Similar issues arose before the Bombay High Court in tax proceedings,[9] an eviction dispute involving fabricated citations and apparent AI-generated drafting,[10] a Delhi High Court petition falsely attributing paragraphs to Raj Narain v. Indira Nehru Gandhi,[11] and even a Bengaluru trial court, where the Karnataka High Court ordered an inquiry after a judge relied upon two non-existent Supreme Court decisions.[12]
The problem first gained international attention in Mata v. Avianca, Inc. (2023), where a U.S. federal court sanctioned lawyers for filing AI-generated fictitious precedents.[13] Indian courts have since referred to the decision as an early warning of the risks posed by unverified AI-generated legal research.
Indian courts initially approached generative AI cautiously. In Jaswinder Singh v. State of Punjab (2023), the Punjab and Haryana High Court used ChatGPT only as a supplementary research tool while making it clear that the judgment rested solely on judicial reasoning.[14] Likewise, in Md. Zakir Hussain v. State of Manipur, the Manipur High Court noted the government’s reliance on ChatGPT but again questioned the reliability of AI-generated legal research.[15]Seen together, these early, relatively cautious encounters with generative AI look almost quaint next to the scale of the fabrication problem that followed within three years , a trajectory that should temper any assumption that the technology’s reliability will simply improve on its own before the profession has built the verification habits to catch it.
Pooja Ramesh Singh and the Supreme Court’s Reckoning
The Supreme Court confronted the issue directly in Pooja Ramesh Singh, arising from insolvency proceedings in which tribunal decisions relied upon AI-generated, non-existent judicial precedents.[16] On 2 July 2026, the Court set aside both tribunal orders, holding that a decision founded upon “non-existent, fake and hallucinated material” generated through AI is “no decision in the eyes of the law,” irrespective of whether the fabricated authority was central or incidental to the outcome.[17]
The Court drew a comparison to the Bhopal gas tragedy to describe how quickly an unregulated technological hazard can spread through a system built on trust, while making clear that it was not rejecting AI’s use in adjudication altogether. Its own words captured the balance it was trying to strike: AI could be adopted “in aid of adjudication,” but only with “total and absolute control over adjudication, with a human in the loop, at every stage.”[18]
The Response So Far: Policies Without Teeth
Institutional responses have followed, though unevenly. The Supreme Court’s Centre for Research and Planning released a White Paper on Artificial Intelligence and the Judiciary in November 2025, requiring that any AI-generated summary, precedent list, or legal ratio be independently verified by a judge or court staff before being relied upon.[19] The Court has also circulated draft AI rules, in June 2026, that would bar AI from deciding cases outright and require lawyers to formally disclose when AI tools have been used in preparing a filing.[20]
At the High Court level, Kerala moved first in July 2025 with a policy barring judicial officers from using AI tools to arrive at findings or draft orders, and specifically warning against feeding case data into public cloud-based tools such as ChatGPT; Gujarat followed with a comparable policy in April 2026, and the Punjab and Haryana High Court circulated similar internal guidance around the same period.[21]
The gap in all of this is structural rather than accidental. Every one of these instruments binds judges and court staff. None of them binds the advocate drafting the petition in the first place, and the Bar Council of India’s own conduct rules , Part VI, Chapter II, which sets out an advocate’s duty to the court, including the mandate under Rule 3 that no advocate shall influence a court’s decision through illegal or improper means , were framed long before generative AI existed and do not speak to citation fabrication in specific terms.[22]
India also has no dedicated AI statute; the field is governed instead by a patchwork of general instruments, including the Digital Personal Data Protection Act, 2023, which touches only the data-handling side of AI tools and says nothing about the reliability of their outputs.[23] District courts, where the overwhelming majority of India’s litigation is actually decided, remain almost entirely outside the reach of these High Court and Supreme Court instruments, even though nothing suggests the hallucination problem is confined to the appellate tiers.[24]
The Missing Piece: A Framework for Verification
What is conspicuously absent is a mandatory, system-wide checkpoint rather than a scattering of advisories. One proposal gaining attention among commentators is a Citation Integrity Module built into the e-filing process itself , the same portal that has already processed over 92 lakh e-filed cases as of September 2025 , which would cross-reference every cited judgment against verified legal databases at the moment of submission, checking not only that the case exists but that the paragraph or ratio being quoted actually appears in it.[25]
A filing that fails this check would not be rejected outright but would carry a discrepancy flag visible to the judge, preserving access to courts while making fabrication far harder to smuggle through unnoticed. Such a mechanism would also close the more dangerous variant of this problem: a citation that is real but attributes to it a paragraph or holding that was never written, an error nearly impossible to catch without reading the entire original judgment.[26]
Conclusion
None of this argues for turning away from AI in Indian legal research. SUVAS has made tens of thousands of judgments legible to citizens who do not read English; SUPACE has begun to relieve judges of hours spent manually sorting precedent; and the ambition behind Phase III of the e-Courts Project answers a caseload crisis that has no other realistic solution at this scale. But the last two years of hallucinated citations, recalled orders, and disciplined litigants show what happens when the profession treats an AI-generated citation with the same trust as one drawn from a reporter checked by a human hand. The Supreme Court’s own language in the Pooja Ramesh Singh matter , zero tolerance, human in the loop, at every stage , is the right instinct. Turning that instinct into an enforceable, uniformly applied verification requirement, reaching down to the district courts where most of India litigates, is the work that remains.
THIS ATICLE IS WRITTEN BY JANICE VINOLIA FROM SAVEETHA SCHOOL OF LAW, SIMATS, CHENNAI
REFERENCE :
[1] IndiaAI, ‘From Backlogs to Breakthroughs: The Integration of AI in India’s Judiciary’; Dr. Syama Prasad Mookerjee Research Foundation, ‘Artificial Intelligence in the Indian Judiciary: SUPACE, SUVAS, and the Limits of Assistive Automation’ (2026).
[2] Oxford Institute of Technology and Justice, ‘India: Increasing Use of AI Across the Justice System’; SPMRF (n 1).
[3] ICTworks, ‘How 5 Judicial AI Solutions Are Proving Skeptics Wrong in India’ (2026); PIB, ‘From Digitisation to Intelligence: How AI is Enhancing Access to Justice in India’ (2026).
[4] LiveLaw, ‘Regulating Artificial Intelligence in Indian Judiciary: From Institutional Experimentation to a National Framework’ (2026); Bharatlaw.ai, ‘AI Legal Research India: What Lawyers Need to Know About SUPACE, LegRAA, and the Rise of Judicial AI Tools’ (2026).
[5] SPMRF (n 1).
[6] IndiaAI (n 1).
[7] Arxiv, ‘Ethical Challenges of Using Artificial Intelligence in Judiciary’ (2025); LiveLaw, ‘Phantom Precedents: The Rise of AI-Generated Case Law in Indian Courts’ (2026).
[8] iPleaders, ‘AI-Hallucinated Case Law: How Fake Citations Are Getting Lawyers Sanctioned in India (and How to Avoid It)’ (2026); Buckeye Trust v PCIT, ITAT Bengaluru, order dated 30 December 2024, recalled under s 254(2) of the Income-tax Act 1961.
[9] MediaNama, ‘10 Cases That Show Indian Courts Have an AI Hallucination Problem’ (2026); KMG Wires Pvt Ltd v National Faceless Assessment Centre, Bombay High Court, October 2025.
[10] Deepak v Heart & Soul Entertainment Ltd, Bombay High Court, January 2026; MediaNama (n 9).
[11] Greenopolis Welfare Association matter, Delhi High Court, September 2025, as reported in MediaNama (n 9) and LiveLaw, ‘Phantom Precedents’ (n 7).
[12] MediaNama (n 9).
[13] Mata v Avianca, Inc, US District Court, Southern District of New York, sanctions order, June 2023, as discussed in iPleaders (n 8).
[14] Jaswinder Singh v State of Punjab (2023), Punjab and Haryana High Court, as discussed in iPleaders (n 8).
[15] Md. Zakir Hussain v State of Manipur, order dated 23 May 2024, Manipur High Court, as discussed in iPleaders (n 8).
[16] LiveLawBiz, ‘Supreme Court Likens AI Hallucinations to Gas Behind Bhopal Gas Tragedy, Sets Aside CIRP Orders Over Fake Case Law’ (2026).
[17] Pooja Ramesh Singh v Jammu and Kashmir Bank Ltd, Civil Appeal No 11950 of 2025, decided 2 July 2026 (SCC Online Blog report).
[18] ibid; LiveLawBiz (n 16).
[19] Law and Other Things, ‘Phantom Precedents: AI Hallucination, the Verification Vacuum, and the Case of Mandatory Pre-Filing Reform in Indian Courts’ (2026).
[20] MediaNama (n 9).
[21] iPleaders (n 8).
[22] iPleaders (n 8); LiveLaw, ‘Phantom Precedents’ (n 7), discussing Bar Council of India Rules, Part VI, Chapter II.
[23] Digital Personal Data Protection Act 2023 (India); Oxford Institute of Technology and Justice (n 2), noting the absence of dedicated AI legislation in India.
[24] Law and Other Things (n 19).
[25] Law and Other Things (n 19), citing e-Courts e-filing portal data as of September 2025.
[26] Law and Other Things (n 19).
