Hallucinated legal citations reach courts often enough that a public database counts the decisions in thousands [1]. As of 9 November 2026, the AI Hallucination Cases database listed 2,286 decisions in which a court or tribunal found or implied that a party relied on hallucinated content, 1,550 of them from the United States [1]. Its author calls the count "necessarily an undercount", because it holds only cases where courts made or implied that finding [3].
The court records behind the count show three ways a citation fails and a range of remedies, from penalties to disqualification and dismissal. The full report reads seventeen decisions from June 2023 to March 2026 in the United States, the United Kingdom, Canada and Australia [4-20], and the rules courts and bars wrote in response.
Hallucinated legal citations by the numbers
A decision can name more than one kind of party, so the database's counts overlap [1].
| Party | Decisions |
|---|---|
| Self-represented litigant | 1,314 |
| Lawyer | 907 |
| Judge | 34 |
| Expert | 18 |
Source: [1], as of 9 November 2026.
By our count of the database's CSV export, it lists 16 decisions from 2023, 61 from 2024, 865 from 2025 and 1,344 from 2026 to 9 November [2]. By the same count, U.S. courts account for between 61% and 72% of each year's decisions [2]. Its author finds cases through referrals, automated monitoring and keyword searches [3], so the newest months may be incomplete. After the United States, the page lists Canada with 245 decisions, Australia with 112, the United Kingdom with 90 and Israel with 58 [1]. The author links the U.S. weight to what he calls "an excellent legal data ecosystem" [3].
Three failures courts describe
The High Court of England and Wales set out what generative tools can produce: they "may cite sources that do not exist. They may purport to quote passages from a genuine source that do not appear in that source" [5]. Courts' findings fall into three failures.
- Fabricated authorities. Cases that do not exist. In Al-Haroun, 18 of 45 citations were to cases that do not exist [5]. In Wadsworth v. Walmart, motions in limine "cited nine cases, but eight did not exist" [11].
- False quotations. Words a real source does not contain. In Noland v. Land of the Free, 21 of 23 case quotations in the opening brief were fabrications [10].
- Misrepresented holdings. Real cases cited for what they do not say. The Fifth Circuit's opinion in Fletcher v. Experian sets out, beside 16 fabricated quotations, five "serious misrepresentations of law or fact", among them a case that "does not mention" the proposition it was cited for [20].
The database tags 1,872 decisions as fabricated, 968 as misrepresented and 606 as false quotes [1]. By our count, 1,037 decisions carry more than one tag [2]. In Couvrette v. Wisnovsky, three briefs cited fifteen nonexistent cases and attributed fabricated quotations to eight real authorities [13]. Invented sources need not be cases. In Kohls v. Ellison, an expert's declaration cited two academic articles that do not exist [18].
What courts ordered
| Date | Decision | Order |
|---|---|---|
| 22 June 2023 | Mata v. Avianca, S.D.N.Y. | $5,000 penalty, letters to the plaintiff and to each judge falsely named as an author [4] |
| 30 January 2024 | Park v. Kim, 2d Cir. | Referral to the court's Grievance Panel [6] |
| 20 February 2024 | Zhang v. Chen, B.C. Supreme Court | Costs of the extra effort borne by counsel personally [8] |
| 24 February 2025 | Wadsworth v. Walmart, D. Wyo. | Pro hac vice admission revoked, a $3,000 penalty and two of $1,000 [11] |
| 6 May 2025 | Lacey v. State Farm, C.D. Cal. | Briefs struck, discovery relief denied, $31,100 to the defense [16] |
| 23 July 2025 | Johnson v. Dunn, N.D. Ala. | Public reprimand, disqualification, referral to the state bar [15] |
| 12 September 2025 | Noland v. Land of the Free, Cal. Ct. App. | $10,000, opinion served on the State Bar [10] |
| 12 December 2025 | Couvrette v. Wisnovsky, D. Or. | Briefs struck, $15,500, claims dismissed with prejudice [13] |
| 18 February 2026 | Fletcher v. Experian, 5th Cir. | $2,500 [20] |
| 23 March 2026 | Couvrette v. Wisnovsky, D. Or. | $94,704.38 in fees and costs [14] |
By our count, 206 U.S. decisions in the database record an amount above its $1 placeholder, with a median of $2,500 and a mean of $7,568 [2]. The U.S. amounts reach $110,204, which combines Couvrette's $15,500 sanction with its fee award, so the field mixes penalties paid to courts with fees paid to the other side [2, 13, 14].
Fines and deterrence in the courts' own words
Two courts said in 2025 that earlier sanctions had failed to deter [15, 17]. Johnson v. Dunn held that "Fabricating legal authority is serious misconduct that demands a serious sanction", and added: "If fines and public embarrassment were effective deterrents, there would not be so many cases to cite" [15]. In HoosierVac, the magistrate judge recommended $15,000, finding that penalties on other attorneys "have evidently failed to act as a deterrent", and the district judge imposed $6,000 [17]. Couvrette adopted a valuation of "$500 per non-existent case and $1,000 per fabricated quotation" and warned that "With a known price tag, an attorney could decide" the risk was worth taking [13]. No source reviewed measures whether any of these sanctions deterred.
Self-represented litigants, experts and judges
Rule 11 binds "an attorney or unrepresented party" [22]. In Harber, the UK First-tier Tribunal accepted that a self-represented appellant had not known her nine cited decisions were not genuine, and still called the practice "a serious and important issue" [7]. The Court of Appeals of Georgia vacated a trial court order that "relied upon non-existent case law" [12]. A Senate Judiciary Committee release of 23 October 2025 says two federal district judges admitted that "their staff used generative AI to draft error-ridden orders" [27].
Courts outside the United States
The High Court's Divisional Court ruled on 6 June 2025 that lawyers who use these tools for research "have a professional duty therefore to check the accuracy of such research by reference to authoritative sources" [5]. It found the threshold for contempt met in Ayinde, decided against contempt proceedings and directed that its judgment go to the Bar Council, the Law Society and the Council of the Inns of Court [5]. In Canada, Zhang v. Chen called citing fake cases "an abuse of process" [8], and Ko v. Li ended a contempt proceeding on conditions that included professional development courses and no bill to the client for the work [19]. In Australia, Dayal referred a solicitor to the Victorian legal regulator, a decision "not intended to be punitive" [9].
Rules courts and bars wrote
The Second Circuit held that Rule 11 requires attorneys to "read, and thereby confirm the existence and validity of, the legal authorities on which they rely", and that no special rule was needed [6]. The Fifth Circuit proposed a certification rule, received thirteen comments and declined it [20, 21]. Other courts chose disclosure. In the Northern District of Texas, "A brief prepared using generative artificial intelligence must disclose this fact on the first page" [23], and the Federal Court of Canada requires a declaration under a notice of 7 May 2024 [24]. A New South Wales practice note, in force from 3 February 2025, requires authors who used generative tools to verify that every citation exists, is accurate and is relevant, and says the verification "must not be solely carried out by using a Gen AI tool or program" [25]. The State Bar of California's 2026 guidance says the "duty of candor to the tribunal cannot be delegated to AI" [26].
Which check catches which failure
A citation lookup answers whether an authority exists. Free Law Project's lookup service, offered "as a guardrail to help prevent hallucinated citations", reports whether a citation was found, and it does not look up statutes or law journals [31]. A citation that resolves still has to be read. The California Court of Appeal published Noland "as a warning" that no paper filed in any court should contain citations "that the attorney responsible for submitting the pleading has not personally read and verified" [10]. Courts named reading the source as the check for quoted words and holdings [6, 10, 11].
Automated checkers cover part of this. In a 2026 benchmark of 1,300 brief excerpts with injected errors, the headline agentic setup reached 84.4% recall and a 55.0% F1 score, and "all models struggle with subtle error categories" [30]. Studies of the tools themselves found hallucination rates of 58% to 88% for the general-purpose models tested on questions about federal cases [28], and 17% to 33% for the commercial legal research tools tested [29]. Our review of LLM contract review benchmarks reads accuracy scores for a related legal task.
Our methods note on checking LLM citations against their sources sets out the check in five steps, from existence to faithfulness. Rillor's document and regulation intelligence work turns rules, contracts and filings into records a system can check, with each field linked to the passage it came from and a person reviewing every record. Our worked example from rule text to a checkable model shows the same pattern on SEC Rule 201: an LLM proposes fields tied to exact phrases, and a person reviews them.
Read the full report
The full report adds the database's counts by year and legal field, each order's terms, the rules courts and bars wrote, the detection results and the limits of the records.