Key Takeaway
In this legal article, Matthew Lee examines eleven potential judicial AI hallucinations. Judicial AI hallucinations meaning fabricated, inaccurate or AI-linked legal material, entered a court or tribunal’s own decision. They are not all the same. In some, AI use was confirmed. In others, it was suspected or reported. In several, the court found that fictitious material had entered the decision but did not establish whether AI had produced it.
Ad/Marketing Communication
This legal article/report forms part of my ongoing legal commentary on the use of artificial intelligence within the justice system. It supports my work in teaching, lecturing, and writing about AI and the law and is published to promote my practice. Not legal advice. Not Direct/Public Access. All instructions via clerks at Doughty Street Chambers.
Subscribe to the AI Law Commentary here.
Last verified: 15 August 2026. This is a non-exhaustive survey of the English-language materials I have been able to verify.
Introduction
The recent South African case of F.J.L v T.G.O involved a judicial decision containing a fictitious case and several inaccurate or combined quotations.
One member of the Full Court considered AI the most plausible explanation for the fictitious authority, although she expressly made no finding that AI had actually been used. She also said that she intended to refer the matter to the relevant professional regulator.
That judgment reminded me of several others I have written about and made me wonder: how often has false or AI-generated material entered a judicial decision?
I sent out my AI agents and spoke with readers who share an interest in these issues. Between us, we identified seven matters in the US, two in India and one each in the UK and South Africa.
For counting purposes, I treat connected proceedings moving through several courts as one matter. I also distinguish between confirmed AI use, suspected or reported AI use and cases where fabricated authorities entered a decision but their source was never established. I have been shown more potential candidates in other jurisdictions, but because I do not speak the relevant languages, I will need reliable assistance with translation before writing about them. You can view the wider collection on the Judicial AI Use Tracker (How are Judges Using AI?)
With those distinctions in mind, let us have a closer look at the eleven matters.
United States
Shahid v Esaam, A25A0196 (my previous commentary here)
It was reported that a court order, apparently prepared by one party’s lawyer, relied on two fictitious cases. The lawyer’s appeal brief later included seven further fictitious cases and four genuine cases which were irrelevant to the points being made.
The appellate court vacated the order, required a new hearing and imposed a $2,500 penalty on the lawyer. It did not find who had created the false citations or whether AI had been used.
In re CorMedix Inc. Securities Litigation, No. 2:21-cv-14020
It was reported that a law-school intern used ChatGPT without authorisation and contrary to chambers and law-school policies while conducting legal research connected to a judicial opinion. An early draft that had not undergone chambers’ normal review process was mistakenly docketed. The judge described it as flawed and inaccurate and said that it should never have been docketed.
The draft was removed from public view and replaced by a final opinion. The judge said that he had taken immediate and appropriate action concerning the intern, introduced a written AI policy and strengthened the chambers’ multi-level review process. As at the date of writing, I have not located any public description of the action taken concerning the intern.
Jackson Federation of Teachers v Fitch, No. 3:25-cv-00417
It was reported that a law clerk used Perplexity as a foundational drafting assistant while preparing a temporary restraining order. An early draft that had not undergone chambers’ normal review process was docketed by mistake. The judge described the draft as containing errors and inaccuracies caused by a failure to confirm information, rather than errors in the core legal reasoning or the ultimate judgment on the merits.
The order was removed and replaced, and the ultimate result remained the same. The judge said that he had taken appropriate action concerning the clerk but did not identify that action publicly. He also introduced mandatory independent review by a second law clerk and required cited cases to be printed from Westlaw and attached to final drafts. The incident was examined following an inquiry from the Chairman of the Senate Judiciary Committee.
In re Domestic Partnership of Campos and Munoz, D085584
A fictitious authority first appeared in a Reddit article, was relied upon by one party’s lawyer and was later included by the opposing lawyer in a proposed order which the court signed.
The appellate court held that reliance on the false authorities was an abuse of discretion. It nevertheless affirmed the order because the challenge had been forfeited and because the appellant had not provided an adequate appellate record. The court imposed $5,000 in sanctions on the lawyer who had cited and continued to rely on the fictitious authorities and directed that the opinion be sent to the State Bar of California.
The court did not find that the Reddit article or the original fictitious authority was AI-generated. However, counsel acknowledged using online resources including AI and conceded that some additional fictitious information may have been obtained using AI tools.
Payne v The State, No. S26A0459
A prosecutor used AI to prepare court filings and a proposed order without independently checking the authorities it produced. The trial court entered an order based largely on that draft. It contained nine citations which either did not exist or did not support the propositions for which they were used. The prosecutor also identified 12 further AI-generated authorities in trial-level briefing that had not been independently verified and withdrew reliance on nine authorities cited in the State’s appellate brief.
The Supreme Court of Georgia vacated the order denying the motion for a new trial and required the trial court to prepare a fresh order without assistance from either party. It suspended the prosecutor’s privilege to practise before the Supreme Court of Georgia for six months and required 12 additional hours of continuing legal education covering ethics, brief writing and the proper use of AI. The Court did not determine the merits of the conviction appeal at that stage.
Henry County Schools v Grant, Douglas and Forbes, A26A0698–A26A0700
The trial court adopted an order drafted by a lawyer. It ignored binding law, misquoted genuine decisions and relied on at least one hallucinated case.
The appellate court reversed the state-law rulings and vacated and remanded the federal discrimination ruling. It did not establish whether AI had produced the false authority.
One party’s written argument relied on a non-existent case and misstated the relevant legislation. The opposing lawyer warned the court about both problems before judgment, but the ruling still repeated the inaccurate passage almost word for word.
In an unpublished opinion, the appellate court reversed the decision and ordered the case to be reconsidered by a different judge. It did not find whether AI had produced the fictitious case.
India
Gummadi Usha Rani v Sure Mallikarjuna Rao
A reproduced copy of the High Court judgment records the judicial officer’s explanation that the four authorities had been obtained using an AI tool. The High Court found that those authorities did not exist and had been included without independent verification.
The High Court strongly warned against unverified judicial use of AI but left the order in place because it considered that the correct legal principles had nevertheless been applied.
In an interlocutory order dated 27 February 2026, the Supreme Court directed that the trial court should not proceed on the basis of the Advocate Commissioner’s report. It also began examining the consequences and accountability arising from the trial court’s reliance on AI-generated, non-existent judgments, issued notice to the Attorney General, Solicitor General and Bar Council of India, and appointed senior counsel. I have not yet located an official final disposal of this matter.
Pooja Ramesh Singh v Jammu and Kashmir Bank Ltd, 2026 INSC 668
A tribunal relied on six false or inaccurate authorities found through its own research. Some cases did not exist. Others were genuine decisions to which AI-generated passages had been falsely attributed. The appellate tribunal then repeated the same material.
The Supreme Court set aside both decisions and sent the case back for fresh consideration. It adopted a “zero tolerance” approach to fake or hallucinated material being used as precedent by either the Bar or the Bench. It directed the Bar Council of India to constitute a committee, consider the submission of fake or hallucinated material by advocates, prescribe preventive guiding principles and identify the disciplinary consequences of breaching those norms.
United Kingdom
Sandie Peggie v Fife Health Board and another, Employment Tribunal (Scotland), Case No. 4104864/2024
The tribunal judgment was corrected twice. The certificates made twelve changes, including corrections to descriptions and purported quotations from decided cases. Several were clerical corrections, while others concerned case names, descriptions, quotation marks or passages attributed to decided cases.
Separate from those documented corrections, a publicly posted complaint raised additional concerns about the judgment. The information below is taken solely from that complaint and the accompanying response. I have not found any additional official source that independently confirms the matters raised, so I am not presenting any allegation as established fact. Readers can review the original material and form their own view. The complaint states:
” …Purely and simply, my concern is that for any judge to issue words in quotation marks that are not contained in the previous case for which a reference is given belongs to the highest level of judicial misconduct.
Suggestions have been made in the Press that the falsehoods may have resulted from careless use of some form of artificial intelligence, but that can be no possible excuse…”
The response said that the judge had not used generative AI and that two of the erroneous quotations had come through correspondence with another judicial office-holder. The publicly reproduced response did not identify the original source of that wording. The corrections did not alter the dispositive outcome.
South Africa
F.J.L v T.G.O (2025/220239) [2026] ZAGPJHC 875 (31 July 2026)
The Acting Judge’s reasons contained incorrect citations, inaccurate quotations, passages assembled from different judgments and a repeatedly cited case which did not exist.
Opperman J considered AI the most plausible explanation, particularly for the fictitious case. She nevertheless made no finding that the Acting Judge had used AI because he had not had an opportunity to respond.
The appeal was dismissed after the Full Court considered the issues without relying on the false material. Opperman J said that she intended to forward the judgment to the Chairperson of the Legal Practice Council for investigation. She stressed that her observations were not findings and were matters to be investigated.
Conclusion
So, what do these eleven matters show? It seems that:
- In five, an appellate court reversed, vacated or set aside all or part of the decision: Shahid, Payne, Henry County Schools, H.C. v Contreras and Pooja Ramesh Singh.
- In CorMedix, an inaccurately docketed preliminary draft was removed and replaced by a final opinion. In Jackson Federation of Teachers, an early draft order was removed and replaced by the same court. Neither change followed an appeal.
- In Peggie, the written reasons were corrected without altering the dispositive outcome.
- The decisions remained in place in Torres Campos and F.J.L., although for very different reasons.
- In Gummadi Usha Rani, the High Court initially left the order standing, but the Supreme Court later intervened and began a wider review.
It also seems the consequences are inconsistent even within a single jurisdiction.
In Gummadi, the Andhra Pradesh High Court treated the fictitious citations as non-vitiating because it considered that the correct legal rule had nevertheless been applied. The Supreme Court of India took a markedly stricter approach in Pooja, holding that a decision must be set aside if fake or AI-hallucinated precedent enters the decision-making process, regardless of whether it had a direct bearing on the outcome. Gummadi is still before the Supreme Court, so this may yet resolve into consistency rather than conflict, but for now these matters disclose neither a settled national rule nor a cross-jurisdictional one.
I was rather taken aback by how difficult it was to find any related consequences publicly. Lawyers or prosecutors were sanctioned in several of the US matters. As at the date of this article, I have not located a public description of any formal sanction imposed on a judge, clerk or intern in these matters. That does not mean that no action was taken. In CorMedix and Jackson Federation of Teachers, the judges said that appropriate action had been taken concerning the relevant member of staff but did not identify what that action was. In F.J.L., Opperman J said that she intended to forward the judgment to the Chairperson of the Legal Practice Council for investigation, while stressing that her observations were not findings.
F.J.L. also contains the most striking piece of judicial reasoning on institutional accountability in this survey. Opperman J examined at length whether the Judicial Service Commission or the Legal Practice Council has jurisdiction over an Acting Judge once the acting term ends, and concluded that the LPC’s is the only jurisdiction that survives the lapse of appointment. I only practise in England and Wales, so I’m wary of saying more than that from the outside. If any readers who do practise in South Africa can offer insight, I’d genuinely welcome it and it may be worth a dedicated article once I’ve heard from you.
So how much should we worry about this problem? Some may say that, given the vast number of cases heard, eleven matters across these jurisdictions represent such a small proportion that the issue is barely significant. Others may point out that these matters are unlikely to capture every incident, particularly in jurisdictions such as the UK, where not every court and tribunal decision is published in a readily searchable form.
My concern is less about what we know than what we do not. Eleven reported matters show that the problem exists, but they cannot tell us how widespread it is.
What concerns me most is how easily a convincing AI hallucination or deepfake might pass unnoticed by the parties or the court. Campos is instructive. One fictitious authority in the case traced back to a Reddit post, which the court did not find to be AI-generated, yet a second fictitious authority in the same filings was described by the court as apparently AI-generated and counsel conceded that further false citation details may also have come from AI tools. Even within a single matter, then, some fabrication was apparently human in origin and some apparently machine and it was not always possible to say which was which. That is the more unsettling finding: verification aimed only at declared “AI use” is unlikely to catch it.
There is also the risk of second or third-order hallucinations, where false information is repeated until it begins to appear credible (see my Type 8 here). I sometimes call this the cascade or domino effect. Could such errors already be embedded in apparently reliable materials and, in some cases, contribute to injustice? The difficulty is that the errors we fail to detect are also the ones we cannot count.
What these decisions do show is that this is no longer simply a problem of lawyers and litigants in person putting fictitious cases before the courts. False material has now entered decisions through lawyer-drafted orders, clerks, interns, a tribunal’s own research and, at least on suspicion, judicial reasoning itself.
Once a fictitious authority becomes part of a judgment, it carries the voice and authority of the court. It may also be harder to detect than an error in a party’s written submissions. That is why the way courts check, preserve, investigate and correct these decisions matters so much.
As always, I would love to hear your thoughts. Can a judicial decision survive an AI hallucination or other significant error, or does the presence of such an error undermine the process to the point that justice requires the matter to be reconsidered without it? Is it ultimately a question of degree and materiality?
If you have found this helpful, you may wish to subscribe to my AI law articles via the Substack below. Please do continue sending me decisions that you think should be tracked, whatever the jurisdiction. One of the clearest lessons from this work is that lawyers, judges and court users in different systems are often confronting closely related problems. Sharing decisions, experiences and reflections across practice areas and jurisdictions remains one of the most constructive ways to respond.
Final Words From Popular LLMs
This is the point at which I invite an LLM to provide an independent comment on the preceding article. At times the observations are slightly off the mark, at other times they are perceptive, and occasionally they produce something entirely unexpected. Whatever the outcome, I find it a useful and revealing exercise to see what emerges.
ChatGPT (OpenAI)
“Taken together, these matters raise legal issues extending well beyond citation accuracy. Once fabricated or unverifiable material enters judicial reasons, questions arise about procedural fairness, the duty to give reasons, meaningful appellate review, open justice, finality and responsibility for material supplied by staff or parties. The cases do not establish a single rule across jurisdictions: some decisions were corrected or replaced, some were remitted, and others remained in force, while the Supreme Court of India in Pooja adopted a markedly stricter approach. The common legal lesson is that every authority and quotation must be independently verified, superseded versions preserved, corrections made transparently, and responsibility for the final decision remain human and identifiable.”
Gemini (Google)
“While the phenomenon of “judicial AI hallucination” is growing, it is not a monolith. A review of recent matters across several jurisdictions reveals that fabricated or erroneous legal material enters judicial decisions through several distinct pathways. In a minority of cases, judicial officers or their clerks have been confirmed to have directly used generative AI tools that produced hallucinations which bypassed review processes. Far more commonly, however, the false material originates externally—often in proposed orders drafted by counsel using AI—which are then adopted by the court without adequate independent verification. Furthermore, when tracking this phenomenon, strict care must be taken not to assume AI provenance where a court has made no such finding, and to distinguish between confirmed AI use, suspected use, and matters where judicial officers have expressly denied utilizing generative AI.”
Claude (Anthropic)
“What strikes me about this survey is less a clean AI/non-AI divide than how tangled the two can become within a single matter. In Torres Campos, the original fictitious Twigg citation traces to a Reddit post read at a family dinner, and the court made no finding that the post itself was AI-generated. But the same proceedings also produced a further fictitious case, Teegarden, which the court itself described as apparently AI-generated, and counsel admitted at oral argument to using AI tools and conceded she may have sourced further fabricated citation details that way. On my own count of the eleven matters, AI use is confirmed in around half and unestablished, denied, or merely suspected in the rest, which suggests verification protocols aimed narrowly at declared ‘AI use’ will still miss a substantial share of the risk this article documents. What these matters actually call for is a rule that any authority unfamiliar to the court, however it arrived, is checked against a primary source before it is relied on, adopted, or signed. On accountability, Opperman J’s extended discussion in F.J.L v T.G.O concerns a narrower and more structural question, whether the JSC or the LPC has jurisdiction over an Acting Judge once the acting term expires. Reading the eleven matters together, though, rather than through that one judgment, the asymmetry is real: lawyers have faced professional sanction in several of the US and Indian matters, while no clerk, intern or judicial officer identified here appears to have faced a comparable public sanction. That gap, my reading of the survey rather than Opperman J’s finding, is the more troubling one, because it leaves the people closest to the point of adoption the least visible when something goes wrong.”
