---
title: "Hallucinations de l'IA en justice : le tracker des sanctions, 2 202 affaires"
url: https://www.haqq.ai/fr/blog/ai-legal-hallucination-audit
locale: fr
type: blog
description: "2 202 décisions de justice impliquent des citations inventées par l'IA, selon la base Charlotin (oct. 2026). Sanctions par pays, courbe de croissance, prévention."
category: "ai-legal-tech"
keywords: "AI hallucination cases, how many AI hallucination cases, lawyers sanctioned for AI, AI fake case citations, AI sanctions tracker, Charlotin AI Hallucination Cases database, Mata v Avianca, Couvrette v Wisnovsky, Withers v City of Aberdeen, Ayinde v Haringey, UKUT 81 AI supervision, Tan Hai Peng SGHC 49, Argentina IA citas inexistentes, AI court sanctions 2026"
published: 2026-05-22
updated: 2026-10-27
english_version: https://www.haqq.ai/blog/ai-legal-hallucination-audit
source: src/data/blogPosts.ts
generator: scripts/generate-page-markdown.ts
---

# Hallucinations de l'IA en justice : le tracker des sanctions, 2 202 affaires

2 202 décisions de justice impliquent des citations inventées par l'IA, selon la base Charlotin (oct. 2026). Sanctions par pays, courbe de croissance, prévention.

> **Note:** In short: Damien Charlotin's AI Hallucination Cases database lists 2,202 court and tribunal decisions involving AI-fabricated citations or content, as of its 19 October 2026 update. That is up from about 200 in mid-2025. US courts account for 1,501. The largest single-matter penalty we can source is about $109,700. We re-verify this page monthly.

> **Note:** **Tracker: 2,202 cases.** As of 19 October 2026 (database "last updated" date, read by HAQQ on 20 October 2026). Source: [Damien Charlotin, AI Hallucination Cases database](https://www.damiencharlotin.com/hallucinations/). Previous check: 2,046 on 21 September 2026. Change: +156 cases in 28 days, about 5.6 per day.

## Key facts

- 2,202 decisions worldwide involving hallucinated AI material (Charlotin database, updated 19 October 2026).
- By country: United States 1,501; Canada 234; Australia 112; United Kingdom 76; Brazil 41; India 16; Italy 15; France 14; Argentina 10; Singapore 8 (Charlotin database jurisdiction filter, 19 October 2026).
- By filer: pro se litigants 1,267; lawyers 872; judges 32; experts 17 (Charlotin database party filter, 19 October 2026).
- Growth: about 200 cases in mid-2025, 719 in January 2026 (Scientific American, 2026), 1,227 in early April 2026 (PlatinumIDS, April 2026), 1,458 on 22 May, 1,598 on 9 June, 2,046 on 21 September and 2,202 on 19 October 2026 (HAQQ checks of the live database).
- Largest single-matter penalty we can source: about $109,700 in sanctions and fees in Couvrette v. Wisnovsky, D. Oregon (ABA Journal, 2026).
- US courts imposed over $145,000 in AI-filing penalties in Q1 2026 (ComplexDiscovery, April 2026).
- Withers v. City of Aberdeen, N.D. Mississippi, 8 June 2026: fake citations on both sides, trial canceled, two lead attorneys suspended from the district for two years (Bloomberg Law, June 2026).
- Paid legal research tools hallucinated on 17% (Lexis+ AI) to 33% or more (Westlaw AI-Assisted Research) of benchmark queries (Stanford RegLab and HAI, Journal of Empirical Legal Studies, 2025).

## How many AI hallucination cases are there?

2,202, as of the database's 19 October 2026 update. The database is maintained by Damien Charlotin, a research fellow at HEC Paris. Its inclusion bar is strict: a decision is listed only where a court or tribunal found, or clearly implied, that a party relied on hallucinated material. Accusations alone do not count.

We check the number ourselves each month rather than quoting secondary coverage. On 22 May 2026 the count was 1,458. On 9 June it was 1,598, a pace of just under eight new cases a day. Between 9 June and 21 September the pace fell to about 4.3 a day (2,046). Between 21 September and 19 October it picked up again to about 5.6 a day (2,202).

Two caveats on reading the number. First, it counts only what a judge caught and wrote up; a fake citation nobody checks never reaches the database, so 2,202 is a floor. Second, most entries are not lawyers. Self-represented litigants account for 1,267 of the decisions and lawyers for 872. The lawyer cases are the ones that produce sanctions, suspensions and regulator referrals, and they are the ones this page tracks in detail.

## How fast is the count growing?

**Documented AI hallucination cases over time**

Cumulative decisions in the AI Hallucination Cases database (Charlotin). Sources: Scientific American, PlatinumIDS, HAQQ checks of the live database.

- Mid-2025: ~200
- January 2026: 719
- Early April 2026: 1,227
- 22 May 2026 (our audit): 1,458
- 9 June 2026: 1,598
- 21 September 2026: 2,046
- 19 October 2026 (latest): 2,202

_May, June, September and October figures are HAQQ's own checks against the live database. The implied rate over the latest window is about 5.6 new cases per day._

Scientific American (2026) put the database at about 200 cases a year earlier and 719 in January 2026. PlatinumIDS counted 1,227 in April 2026, 811 of them in US courts. Our four checks since May extend the same line. The daily rate moves around from month to month, but it has not gone below four.

## Which lawyers have been sanctioned for AI fake citations?

The table below lists the decisions we treat as landmarks, verified against the judgment or primary reporting. Read top to bottom, it shows the penalty arc: a fine in 2023, fee awards and revoked admissions in 2025, then suspensions, a canceled trial and personal costs against supervising lawyers by 2026.

| Case | Court | Year | Penalty |
| --- | --- | --- | --- |
| Mata v. Avianca | S.D.N.Y. (US) | 2023 | $5,000 fine |
| Wadsworth v. Walmart | D. Wyoming (US) | 2025 | $3,000 fine for drafter, $1,000 each for 2 signers, pro hac vice revoked |
| R (Ayinde) v Haringey [2025] EWHC 1383 (Admin) | Divisional Court (England and Wales) | 2025 | Contempt threshold met; referral to the Bar Standards Board |
| Coomer v. Lindell | D. Colorado (US) | 2025 | $3,000 each for 2 lawyers |
| Acevedo c/ Caceres Mareco | Cámara de Apelación Civil y Comercial de Morón, Sala I (Argentina) | 2025 | Appeal declared deserted; warning; bar referral |
| M.J.L. c/ Peugeot Citroen Argentina | Cámara de Apelaciones de General Roca (Argentina) | 2025 | Warning to both sides; bar referral |
| UK v SSHD [2026] UKUT 81 (IAC) | Upper Tribunal, Immigration and Asylum Chamber (UK) | 2025 (reported 2026) | One solicitor referred to the SRA; one had self-reported |
| Couvrette v. Wisnovsky | D. Oregon (US) | 2025 | About $109,700 in sanctions and fees |
| Tan Hai Peng Micheal v Tan Cheong Joo [2026] SGHC 49 | High Court (Singapore) | 2026 | S$5,000 personal costs each against counsel and the assisting solicitor |
| Whiting v. City of Athens | 6th Circuit (US) | 2026 | $15,000 each for 2 lawyers, plus fees and double costs |
| Withers v. City of Aberdeen | N.D. Mississippi (US) | 2026 | Trial canceled; 2-year district suspensions; fines of $1,000 to $3,500 for every lawyer of record |

## How do courts treat AI hallucinations by jurisdiction?

Courts in the United States, England and Wales, Singapore and Argentina reached the same rule without coordinating: the lawyer who files a document must check every authority in it, and that duty cannot be handed to a tool, a junior or a vendor. Where they differ is in who pays and how much.

### United States: from fines to careers

US courts produce most of the database (1,501 of 2,202 decisions) and most of the money. Mata v. Avianca (S.D.N.Y., June 2023) set the reference point: six invented cases from ChatGPT, a $5,000 fine. In Wadsworth v. Walmart (D. Wyoming, February 2025), eight of nine citations came from Morgan & Morgan's in-house platform, MX2.law, and did not exist; the drafting lawyer lost his pro hac vice admission (LawSites, 2025). Branding a tool as legal AI did not move the verification duty off the lawyer.

In Coomer v. Lindell (D. Colorado, July 2025), a brief with nearly 30 defective citations cost each of two lawyers $3,000 (NPR, 2025). In Couvrette v. Wisnovsky (D. Oregon), 15 fake citations and eight fabricated quotations across three summary judgment briefs produced about $109,700 in combined sanctions and fees, which the magistrate judge called "a notorious outlier in both degree and volume" (ABA Journal, 2026).

Appellate courts then raised the ceiling. In Whiting v. City of Athens, the Sixth Circuit found more than 24 fake citations, ordered two lawyers to pay $15,000 each in punitive sanctions under Federal Rule of Appellate Procedure 38, plus the other side's fees and double costs (Sixth Circuit Appellate Blog, 2026). On 8 June 2026, in Withers v. City of Aberdeen, Judge Sharion Aycock canceled a trial after both sides filed nonexistent cases, suspended the two lead out-of-state lawyers from the Northern District of Mississippi for two years, revoked pro hac vice admissions and notified state bars (Bloomberg Law, 2026).

### England and Wales: the authoritative-source rule

In R (Ayinde) v London Borough of Haringey, heard with Al-Haroun v Qatar National Bank, the Divisional Court (the President of the King's Bench Division and Johnson J) handed down [2025] EWHC 1383 (Admin) on 6 June 2025. Five fake cases in a homelessness judicial review brought a junior barrister before the court. The court found the threshold for contempt was met, declined to start proceedings that time, and referred her to the Bar Standards Board (Burges Salmon, 2025).

> "Freely available generative artificial intelligence tools... are not capable of conducting reliable legal research." (R (Ayinde) v Haringey [2025] EWHC 1383 (Admin))

The judgment also listed what counts as an authoritative source: the government's legislation database, the National Archives database of judgments, the official Law Reports and the databases of reputable legal publishers. That list is a practical specification for any research tool a barrister relies on.

The Upper Tribunal (Immigration and Asylum Chamber) went further on supervision in UK v Secretary of State for the Home Department [2026] UKUT 81 (IAC), promulgated 17 November 2025. Under its Hamid jurisdiction it held that a supervising lawyer remains responsible for the accuracy of delegated work and must make sure fee-earners understand the risks of non-specialist AI. It referred one solicitor to the Solicitors Regulation Authority; the other had already self-reported. It also warned that uploading confidential documents to an open-source tool such as ChatGPT can breach confidentiality and waive privilege, and should be referred to the Information Commissioner's Office.

### Singapore: personal costs for both lawyers

In Tan Hai Peng Micheal v Tan Cheong Joo [2026] SGHC 49 (judgment dated 6 March 2026), S Mohan J dealt with two fictitious cases in the defendants' closing submissions. Counsel Goh Peck San filed them; solicitor Amarjit Singh Sidhu, brought in to help with research, had supplied them. The court held that counsel's duty to verify was non-delegable, ordered S$5,000 in personal costs against each of them on an indemnity basis, refused to deal with the issue privately, and barred both from charging their clients for the time spent on it.

### Argentina: warnings and bar referrals at appellate level

Argentina has 10 decisions in the database. The appellate rulings we could verify end in warnings and referrals to the local bar; none imposed a fine. In M.J.L. c/ Peugeot Citroen Argentina, the General Roca civil appeals chamber found nonexistent citations in both sides' briefs, imposed no sanction, issued a formal warning and referred the matter to the Colegio de Abogados de General Roca. In Acevedo c/ Caceres Mareco, Sala I of the Morón civil and commercial appeals chamber declared the appeal deserted because its grounds rested on citations nobody could locate, and also referred the matter to the bar (Marval O'Farrell Mairal, October 2025).

A Rosario appeals chamber (Sala II, August 2025) took the same preventive line, calling the lawyer's conduct good faith but describing unverified delegation of case-law research as reckless while the tools still hallucinate (Diario Judicial, 2025). The procedural consequence is the one Argentine lawyers feel: an appeal built on fake authority fails the concrete-critique requirement and is lost.

### United Arab Emirates: the Gulf's first AI costs order

In *Arabyads Holding Ltd v Gulrez Alam* [2025] ADGMCFI 0032, decided in December 2025, the ADGM Court of First Instance ordered the defendant's former lawyers to pay AED 282,508 in wasted costs, on an indemnity basis. Their defence relied on non-existent or misapplied authorities, and the court called the failure to verify AI-generated research reckless and a breach of the ADGM Rules of Conduct (Matouk Bassiouny & Hennawy; Morgan Lewis). How the three UAE court systems treat AI is in [Onshore, DIFC, ADGM](https://www.haqq.ai/blog/uae-ai-rules-for-lawyers-onshore-difc-adgm).

## How have AI sanctions escalated?

**Largest sourced penalty in a single matter, by milestone**

US dollar amounts as reported. Sources: Mata v. Avianca order; Whiting (Sixth Circuit Appellate Blog); Couvrette (ABA Journal).

- Mata v. Avianca (2023): $5,000
- Whiting v. City of Athens (2026), punitive sanctions: $30,000
- Couvrette v. Wisnovsky (2025-26), combined: ~$109,700

_Withers v. City of Aberdeen (June 2026) carried smaller fines but added two-year suspensions and a canceled trial, which no dollar bar captures._

The money matters less than the other consequences. The 2025 and 2026 decisions add revoked admissions, regulator referrals, a contempt finding, a canceled trial, district suspensions and personal costs against the lawyer who signed without checking. A fine ends a bad week. A two-year suspension from a federal district ends a practice there.

## Why do lawyers keep filing fake AI citations?

Base rates explain most of it. Stanford RegLab's study found that purpose-built paid research tools hallucinated on 17% (Lexis+ AI) to 33% or more (Westlaw AI-Assisted Research) of its benchmark queries, and an earlier Stanford study found general chatbots wrong on 69% to 88% of legal queries. Our own 300-task benchmark of frontier models found 24% of graded answers cited or applied law that did not support the point (HAQQ ran that benchmark and sells a legal AI product, so weigh our numbers accordingly). A tool that is wrong one time in four, used daily by busy lawyers, fills this database on schedule.

The rest is delegation. Wadsworth shows a firm platform inventing eight of nine cites. UKUT 81 and SGHC 49 show the person who signed relying on someone else to have checked. Every jurisdiction above landed on the same answer: the signature carries the duty.

## How can lawyers avoid AI citation sanctions?

- Open every cited authority in a primary source (official reports, court databases, legislation portals) before filing. A citation that only exists in the AI output is a fake until proven otherwise.
- Check that the passage says what the brief claims. Several sanctioned briefs cited real cases for propositions they never stated.
- Put the check at signature. Supervisors in the UK and Singapore were held responsible for work they delegated and did not check.
- Keep confidential documents out of consumer chatbots; the Upper Tribunal treats that as a confidentiality and privilege problem.
- Prefer tools that retrieve and cite sources you can click over tools that answer from memory, and keep a record of what was verified.

This is how we built HAQQ's [legal research](https://www.haqq.ai/product/legal-research): Justinian retrieves the source, cites it, and flags when it cannot find support, and material work routes to a lawyer before it leaves the firm. That narrows the gap and leaves the duty to check with the lawyer.

For a step-by-step version you can adopt as firm policy, see [The Maker Cannot Be the Checker: A Verification Protocol for Legal AI](https://www.haqq.ai/blog/legal-ai-verification-protocol).

## How is this tracker kept current?

Once a month we re-read the live database and primary reporting and check four things: the total count, the largest sourced penalty, new landmark decisions (appellate, cross-border or a new kind of sanction) and the implied daily rate. A number we cannot re-verify comes off the page.

- 20 October 2026: count updated from 2,046 to 2,202 (database update of 19 October), +156 in 28 days, about 5.6 a day. Added country and filer breakdowns. Folded in our earlier jurisdiction report: UKUT 81, verified SGHC 49 orders (S$5,000 each), and the Argentine appellate decisions (warnings and bar referrals). Corrected the Wadsworth fine breakdown.
- 21 September 2026: count updated from 1,598 to 2,046, +448 in 104 days, about 4.3 a day.
- 11 June 2026: rebuilt as a tracker. Count updated from 1,458 to 1,598. Added Couvrette, Whiting, Withers, Wadsworth, Coomer, Ayinde and Tan Hai Peng.
- 22 May 2026: original audit published at 1,458 cases.

General legal information, not legal advice; check the rules of the court and regulator you appear before.

## Sources and further reading

- [Damien Charlotin: AI Hallucination Cases database](https://www.damiencharlotin.com/hallucinations/)
- [R (Ayinde) v Haringey and Al-Haroun v QNB [2025] EWHC 1383 (Admin), Judiciary of England and Wales](https://www.judiciary.uk/judgments/ayinde-v-london-borough-of-haringey-and-al-haroun-v-qatar-national-bank/)
- [UK v SSHD [2026] UKUT 81 (IAC), National Archives](https://caselaw.nationalarchives.gov.uk/ukut/iac/2026/81)
- [Tan Hai Peng Micheal v Tan Cheong Joo [2026] SGHC 49, eLitigation](https://www.elitigation.sg/gd/s/2026_SGHC_49)
- [Marval O'Farrell Mairal: Argentine courts warn over fake AI citations](https://www.marval.com/publicacion/tribunales-advierten-por-citas-falsas-creadas-con-ia-17356)
- [ABA Journal: Oregon federal judge hands down $110K penalty for AI errors](https://www.abajournal.com/news/article/oregon-federal-judge-hands-down-110000-penalty-for-ai-errors)
- [Bloomberg Law: lawyers on both sides in Mississippi case punished for AI errors](https://news.bloomberglaw.com/litigation/lawyers-on-both-sides-in-mississippi-case-punished-for-ai-errors)
- [Sixth Circuit Appellate Blog: sanctions for fake citations in Whiting v. City of Athens](https://www.sixthcircuitappellateblog.com/recent-cases/sixth-circuit-sanctions-attorneys-for-fake-citations-what-does-this-mean-for-use-of-ai/)
- [LawSites: Morgan & Morgan attorneys sanctioned in Wadsworth v. Walmart](https://www.lawnext.com/2025/02/federal-judge-sanctions-morgan-morgan-attorneys-for-ai-generated-fake-cases-in-court-filing.html)
- [NPR: AI hallucination in Mike Lindell case](https://www.npr.org/2025/07/10/nx-s1-5463512/ai-courts-lawyers-mypillow-fines)
- [Burges Salmon: professional conduct and AI in Ayinde v Haringey](https://www.burges-salmon.com/articles/102krxw/professional-conduct-and-ai-ayinde-v-haringey/)
- [Scientific American: why lawyers keep citing fake cases invented by AI](https://www.scientificamerican.com/article/why-lawyers-keep-citing-fake-cases-invented-by-ai/)
- [ComplexDiscovery: the AI sanction wave, $145K in Q1 penalties](https://complexdiscovery.com/the-ai-sanction-wave-145k-in-q1-penalties-signals-courts-have-lost-patience-with-genai-filing-failures/)
- [PlatinumIDS: inside the AI hallucination crisis hitting courts](https://blog.platinumids.com/blog/ai-hallucination-crisis-courts-2026)
- [Stanford HAI / RegLab: legal models hallucinate 1 out of 6 queries or more](https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries)
- [Our fake-citation search experiment](https://www.haqq.ai/blog/legal-ai-hallucination-fake-citation)
- [Can lawyers use AI? Bar rules country by country](https://www.haqq.ai/blog/can-lawyers-use-ai)
- [Our 300-task frontier model benchmark (the 24% citation gap)](https://www.haqq.ai/blog/best-ai-for-legal-work-benchmark)
- [AI and client confidentiality for lawyers](https://www.haqq.ai/blog/ai-client-confidentiality-lawyers)

## Questions fréquentes

### Combien y a-t-il d'affaires d'hallucinations d'IA ?

2 202 décisions de cours et tribunaux au 19 octobre 2026, selon la base AI Hallucination Cases de Damien Charlotin. Elle ne recense que les décisions dans lesquelles un tribunal a constaté ou clairement laissé entendre le recours à des éléments hallucinés, de sorte que le nombre réel est plus élevé. Les tribunaux américains en représentent 1 501.

### Combien d'affaires d'hallucinations d'IA impliquent des avocats plutôt que des parties se représentant elles-mêmes ?

872 des 2 202 décisions impliquent des avocats et 1 267 des parties agissant sans avocat (pro se), selon le filtre par partie de la base au 19 octobre 2026. Un plus petit nombre implique des juges (32), des experts (17) et des procureurs (5).

### Quelle est la plus lourde sanction pour de fausses citations générées par l'IA ?

La plus lourde pénalité dans une seule affaire que nous puissions sourcer est d'environ 109 700 $ de sanctions et de frais cumulés dans Couvrette v. Wisnovsky (D. Oregon), où l'avocat avait déposé 15 fausses citations et huit citations textuelles fabriquées dans trois mémoires (ABA Journal, 2026).

### Un avocat peut-il être suspendu pour avoir cité de fausses affaires générées par l'IA ?

Oui. Dans Withers v. City of Aberdeen (N.D. Mississippi, juin 2026), le juge a suspendu deux avocats principaux du droit d'exercer dans le district pendant deux ans et a annulé le procès. Les tribunaux révoquent aussi des admissions pro hac vice et signalent des avocats aux régulateurs, comme dans Ayinde v Haringey (Bar Standards Board) et UK v SSHD [2026] UKUT 81 (Solicitors Regulation Authority).

### Un avocat superviseur est-il responsable des hallucinations d'IA d'un collaborateur junior ?

Les tribunaux d'Angleterre et de Singapour répondent oui. L'Upper Tribunal, dans [2026] UKUT 81 (IAC), a jugé que les superviseurs restent responsables de l'exactitude du travail délégué. Dans Tan Hai Peng v Tan Cheong Joo [2026] SGHC 49, la High Court de Singapour a mis à la charge personnelle, à hauteur de 5 000 S$ de dépens, à la fois l'avocat plaidant qui avait déposé des affaires fictives et le solicitor qui les lui avait fournies.

### Comment les avocats peuvent-ils éviter les sanctions lorsqu'ils utilisent l'IA ?

Ouvrez chaque source citée dans une source primaire avant le dépôt, vérifiez que le passage dit ce qu'affirme le mémoire, faites ce contrôle au moment de la signature plutôt que de vous fier à la personne qui a rédigé, gardez les documents confidentiels hors des chatbots grand public, et privilégiez les outils qui citent des sources consultables plutôt que ceux qui répondent de mémoire.

### Quelles affaires judiciaires ont donné lieu à des sanctions pour des hallucinations d'IA dans des actes de procédure ?

L'exemple le plus clair dans le Golfe est Arabyads Holding Ltd v Gulrez Alam [2025] ADGMCFI 0032. En décembre 2025, l'ADGM Court of First Instance a condamné les anciens représentants juridiques du défendeur à payer 282 508 AED au titre des frais inutilement exposés (wasted costs), sur une base indemnitaire, pour avoir déposé des recherches assistées par l'IA non vérifiées. Comme dans d'autres affaires de sanctions liées à l'IA, le problème était l'absence de vérification, et non l'usage de l'IA.

---

Canonical page: https://www.haqq.ai/fr/blog/ai-legal-hallucination-audit
Whole-site context for agents: https://www.haqq.ai/llms.txt · https://www.haqq.ai/agents.md

Generated at build time from the site's own content data. Do not edit by hand.
