Qualité des preuves 4.63/5
Score d’évaluation en huit dimensions selon la grille de qualité . Chaque dimension notée de 1 à 5.
- D1 Ancrage des sources
- 5/5
- D2 Autorité des sources
- 5/5
- D3 Arithmétique
- 5/5
- D4 Incertitude
- 5/5
- D5 Portée
- 4/5
- D6 Prose
- 4/5
- D7 Honnêteté sur la perception
- 4/5
- D8 Complétude des réserves
- 5/5
La probabilité qu’une hallucination d’IA cause un impact grave dans la vie réelle d’une personne en particulier n’est pas connue à l’échelle populationnelle, et prétendre le contraire reviendrait à fabriquer exactement le type de chiffre dont cette fiche est censée se méfier. Ce qui est connu, c’est le taux d’hallucination par sortie, et il est suffisamment élevé pour que l’absence d’estimation de préjudice par personne ne doive pas être lue comme rassurante. L’audit 2024 de Stanford RegLab portant sur trois modèles de fondation a constaté qu’ils répondaient à des questions spécifiques et vérifiables sur des affaires fédérales aléatoires avec un taux d’hallucination compris entre 58 % (GPT-4) et 88 % (Llama 2). Le classement de résumés en continu de Vectara, qui utilise une tâche en livre fermé beaucoup plus simple où le modèle reçoit le texte source, mesure encore les modèles de pointe à 1,8-5 % d’hallucination en mai 2026. L’écart de 50x entre ces deux références n’est pas une contradiction : c’est la différence entre demander au modèle de se souvenir de faits et lui demander de résumer un texte qu’il vient de voir. Les questions factuelles ouvertes, celles que posent les utilisateurs ordinaires, se rapprochent davantage du taux le plus élevé.
Le volet des préjudices est documenté incident par incident, et non en agrégé. En juin 2023, la US District Court for the Southern District of New York a sanctionné deux avocats à hauteur de 5 000 $ dans l’affaire Mata v. Avianca pour avoir déposé un mémoire citant six affaires entièrement fabriquées par ChatGPT, dont Varghese v. China Southern Airlines et Estate of Durden v. KLM Royal Dutch Airlines. L’affaire a servi de modèle pour ce qui a suivi : en mai 2026, le traqueur d’hallucinations de Charlotin recensait plus de 1 300 décisions judiciaires dans le monde où un juge ou un tribunal citait explicitement des faussetés générées par IA dans des actes déposés. Une analyse de Stanford Cyberlaw publiée en octobre 2025 a conclu que, aux États-Unis, la majorité de ces actes étaient déposés par des justiciables sans avocat (160 sur 274 cas documentés au moment de la publication), et non par des avocats, ce qui suggère que les affaires visibles impliquant des avocats sanctionnés ne sont qu’une petite tranche d’un schéma plus large touchant principalement les personnes ne pouvant pas se permettre de représentation.
En dehors des tribunaux, le British Columbia Civil Resolution Tribunal a jugé en février 2024 (Moffatt v. Air Canada, 2024 BCCRT 149) qu’Air Canada était responsable d’une promesse de tarif de deuil d’environ 650 CAD faite par son agent conversationnel de service client. La compagnie aérienne soutenait que l’agent conversationnel était une entité juridique distincte ; le tribunal a rejeté cet argument dans des termes qui ont été cités à l’international (« peu importe que l’information provienne d’une page statique ou d’un agent conversationnel »). Les dommages-intérêts dans l’affaire Moffatt étaient faibles, mais cette décision est le précédent existant le plus net établissant que la désinformation par IA grand public crée une exposition juridique pour l’entreprise qui la déploie, indépendamment de la question de savoir si le modèle sous-jacent provenait d’un fournisseur tiers.
Les affaires actives les plus lourdes de conséquences concernent la dépendance émotionnelle et les préjudices en santé mentale. Garcia v. Character Technologies, déposée en octobre 2024 dans le Middle District of Florida, alléguait que la plateforme Character.AI avait contribué au suicide de Sewell Setzer III, âgé de 14 ans, à travers des interactions prolongées de jeu de rôle. Raine v. OpenAI, déposée le 26 août 2025 devant la San Francisco County Superior Court, allègue que ChatGPT a contribué au suicide d’Adam Raine, âgé de seize ans, la plainte affirmant que le système de modération d’OpenAI avait suivi les messages de Raine et les avait signalés à plusieurs reprises sans intervention efficace. D’autres familles ont depuis déposé des actions connexes contre OpenAI, selon le résumé Wikipédia de la vague de contentieux post-Raine. Aucune de ces affaires n’a été tranchée sur le fond au moment de la rédaction ; les allégations ne sont pas des constatations de faits. Elles constituent la preuve que la voie de préjudice impliquant des utilisateurs vulnérables et la dépendance émotionnelle est avancée, devant les tribunaux, avec documentation, dans plusieurs affaires simultanément.
La raison pour laquelle cette fiche refuse un chiffre unique est qu’aucune des quatre voies de préjudice documentées (sanction professionnelle, perte financière du consommateur, conseil médical erroné, préjudice psychiatrique) ne dispose d’un taux de base mesuré par utilisateur-année, et les agréger en une seule probabilité de « préjudice par hallucination d’IA » dissimulerait le fait que ces voies ont des populations exposées différentes, des sévérités de conséquences différentes et des trajectoires de remédiation différentes. Le taux technique par requête est mesuré et clairement non nul ; le taux par personne sur la durée de vie n’est pas encore mesuré et n’est pas assez stable pour être estimé. La fiche honnête est celle qui le dit.
Pour aller plus loin
La plupart des utilisateurs traitent la sortie de l'IA comme une référence à faible risque, mais des sanctions judiciaires, des constats de régulateurs et des poursuites pour mort injustifiée montrent qu'un préjudice grave n'est pas nul, surtout quand on agit sur des sorties non vérifiées. La fréquence à laquelle une session se termine en préjudice grave n'a aucun taux mesuré.
Registre des sources
Chaque chiffre ci-dessous correspond à ce que la source a rapporté, avec la citation textuelle sur laquelle nous nous sommes appuyés et la méthode de calcul. Cliquez sur un lien pour vérifier directement.
3/9 sources vérifiées de manière indépendante, mot pour mot, par rapport à la source citée
-
[1] arXiv (Stanford RegLab); Journal of Legal Analysis 16(1), 2024 — Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models
Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models- Statistique
Hallucination rates of 58% (GPT-4) to 88% (Llama 2) when answering specific verifiable questions about random federal court cases.- Extrait
“"Legal hallucinations are alarmingly prevalent, occurring between 58% of the time with ChatGPT 4 and 88% with Llama 2, when these models are asked specific, verifiable questions about random federal court cases. [...] LLMs often fail to correct a user's incorrect legal assumptions in a contra-factual question setup. [...] LLMs cannot always predict, or do not always know, when they are producing legal hallucinations. [...] Even experienced lawyers must remain wary of legal hallucinations, and the risks are highest for those who stand to benefit from LLMs the most -- pro se litigants or those without access to traditional legal resources." ”
- Données source de
- 2024-06-21
- Consulté le
- 2026-05-28
- Calcul
- Dahl, Magesh, Suzgun and Ho (Stanford RegLab, published in Journal of Legal Analysis, preprint arXiv:2401.01301) measured per-query hallucination rates for legal questions against three foundation models. Their methodology was 200,000+ queries per model stratified by legal task complexity. This source establishes that the per-query hallucination rate is high (58-88%) in a specific high-stakes domain, but it measures hallucination per query, not material harm per user-year. Cited here to ground the claim that hallucinations are frequent at the technical level, while explicitly NOT providing the population harm rate needed for a per-person lifetime probability.
- Indépendance
- Stanford RegLab is an academic research group; independent of vendor benchmarks (Vectara) and independent of the litigation tracker (Charlotin).
-
[2] Vectara — Vectara Hallucination Leaderboard (HHEM-2.3) Vérifié
Vectara Hallucination Leaderboard (HHEM-2.3)- Statistique
Current frontier models hallucinate 1.8% to 5% of the time on a constrained document-summarization task (May 2026 leaderboard snapshot); GPT-5.4-nano: 3.1%, Gemini 2.5 Flash Lite: 3.3%, Phi-4: 3.7%, Llama 3.3 70B: 4.1%.- Extrait
“"Fed the full set of documents in the dataset to each of the LLMs and asked them to summarize each document, using only the facts presented in the document. We then computed the overall factual consistency rate (no hallucinations) and hallucination rate (100 - accuracy) for each model." ”
- Données source de
- 2026-05-11
- Consulté le
- 2026-05-28 · copie archivée
- Vérification
- Extrait récupéré et confirmé de manière indépendante, mot pour mot, par rapport à la source citée lors de notre audit de vérification.
- Calcul
- Vectara's HHEM leaderboard measures hallucination on a closed-book summarization task (the model is given source text and must summarize it). This is a generous benchmark: open-ended factual questions hallucinate far more often. Even on this constrained task, the best frontier models still hallucinate roughly 1 in 20 to 1 in 55 outputs. Establishes a current floor on hallucination prevalence per output, but does NOT translate to per-person harm because (a) most outputs are not acted on, (b) most outputs are not in a domain where being wrong has material consequences, and (c) the leaderboard does not measure open-ended question answering, where rates are an order of magnitude higher.
- Indépendance
- Vectara is a commercial RAG vendor; the leaderboard is independent of Stanford RegLab and uses a different methodology (summarization, not legal QA).
-
[3] Wikipedia summary of United States District Court for the Southern District of New York case record — Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023)
Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023)- Statistique
$5,000 fine against the plaintiff's attorneys in May/June 2023 for a motion containing fake legal cases involving fictitious airlines with fabricated quotations generated by ChatGPT; Judge Castel held the attorneys acted with 'subjective bad faith' under Rule 11.- Extrait
“"In May 2023, Judge P. Kevin Castel dismissed the personal injury case against Avianca and ordered the plaintiff's attorneys to pay a $5,000 fine. [...] Judge Castel noted numerous inconsistencies in the opinion summaries, describing one of the legal analyses as 'gibberish.' [...] Judge Castel held that Mata's lawyers had acted with 'subjective bad faith' sufficient for sanctions under Federal Rule of Civil Procedure Rule 11." ”
- Données source de
- 2023-06-22
- Consulté le
- 2026-05-28 · copie archivée
- Calcul
- First widely publicised US court case where lawyers were sanctioned for filing a brief containing AI-fabricated case citations. The sanction itself was small ($5,000), but the reputational and professional consequences were severe. Establishes that hallucination-induced professional harm is documented and not hypothetical. Does NOT establish a base rate because most users are not lawyers filing briefs, but it anchors one of the harm pathways.
- Indépendance
- Primary court record; independent of all other sources cited.
-
[4] British Columbia Civil Resolution Tribunal (analysis by McCarthy Tetrault) — Moffatt v. Air Canada, 2024 BCCRT 149 Vérifié
Moffatt v. Air Canada, 2024 BCCRT 149- Statistique
British Columbia Civil Resolution Tribunal ruled on February 19, 2024 that Air Canada was liable for the misrepresentation made by its website chatbot about bereavement fare eligibility; awarded the plaintiff approximately CAD 650 plus interest and filing fees.- Extrait
“"The tribunal held that Air Canada was liable for the misleading information provided by the chatbot. … It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot." ”
- Données source de
- 2024-02-19
- Consulté le
- 2026-05-28 · copie archivée
- Vérification
- Extrait récupéré et confirmé de manière indépendante, mot pour mot, par rapport à la source citée lors de notre audit de vérification.
- Calcul
- Primary tribunal decision (cited via legal analysis because the CRT direct URL is paywalled). Documents that the chatbot fabricated a bereavement-fare policy that did not exist, that the airline tried to disavow responsibility ("the chatbot is a separate legal entity"), and that the tribunal rejected that defence. Material consequence per user was small (CAD 650), but the case establishes corporate liability for AI misrepresentation and has been cited internationally as precedent. Does not provide a per-person harm rate, but documents a verified consumer-facing harm pathway.
- Indépendance
- Tribunal decision summarized by an independent Canadian law firm; verified against ABA reporting and BBC/Washington Post coverage of the same ruling.
-
[5] CBS News — AI company, Google settle lawsuit over Florida teen's suicide linked to Character.AI chatbot
AI company, Google settle lawsuit over Florida teen's suicide linked to Character.AI chatbot- Statistique
Garcia v. Character Technologies, Inc. (M.D. Fla., case no. 6:24-cv-01903), filed October 22, 2024, alleged that a Character.AI chatbot contributed to the suicide of 14-year-old Sewell Setzer III after a months-long simulated romantic relationship with the bot; the case against Character Technologies, its co-founders, and Google was settled in January 2026, terms undisclosed.- Extrait
“"A Florida family agreed to settle a wrongful death lawsuit Wednesday with an AI company, Google and others after their teen son died by suicide in 2024. The terms of the settlement, which was filed in the U.S. District Court in the Middle District of Florida, were not disclosed. … Megan Garcia filed a lawsuit in October 2024 … I became the first person in the United States to file a wrongful death lawsuit against an AI company for the suicide of my son." ”
- Données source de
- 2026-01-07
- Consulté le
- 2026-07-03 · copie archivée
- Calcul
- Garcia v. Character Technologies, Inc. et al. (M.D. Fla., case no. 6:24-cv-01903), filed October 22, 2024 by Megan Garcia following the February 2024 suicide of her 14-year-old son Sewell Setzer III, alleged wrongful death, product liability, and negligence after the teenager conducted a months-long simulated romantic and sexual relationship with a Character.AI chatbot persona. Named defendants included Character Technologies, its co-founders, and Google. The case was settled in January 2026 (docket terminated 2026-01-07 per court records); settlement terms were not disclosed. Cited to ground the body's reference to this case with a dedicated source, alongside Raine v. OpenAI, documenting the emotional-dependence/psychiatric-harm pathway with a named, court-verified case rather than only the summary mention in this entry's caveats field.
- Indépendance
- Mainstream news report of a filed and settled federal lawsuit, cross-checked against the public court docket; independent of the Stanford, Vectara, and Wikipedia-sourced case summaries cited above.
-
[6] San Francisco County Superior Court (case description per Wikipedia / NBC News / CNN coverage) — Raine v. OpenAI, No. CGC-25-XXXXXX (San Francisco County Superior Court, filed August 26, 2025)
Raine v. OpenAI, No. CGC-25-XXXXXX (San Francisco County Superior Court, filed August 26, 2025)- Statistique
Ongoing wrongful-death lawsuit filed by Matthew and Maria Raine in San Francisco County Superior Court on August 26, 2025, alleging that ChatGPT contributed to the suicide of their sixteen-year-old son Adam Raine; complaint alleges OpenAI's monitoring system tracked his messages and flagged them repeatedly.- Extrait
“"Raine v. OpenAI is an ongoing lawsuit filed in August 2025 by Matthew and Maria Raine against OpenAI and its chief executive Sam Altman over the alleged wrongful death of their 16-year-old son Adam, who died by suicide in April 2025. … On August 26, 2025, Matthew and Maria Raine filed a lawsuit against OpenAI, Sam Altman , and unnamed OpenAI employees and investors, in the San Francisco County Superior Court … OpenAI's monitoring system, which scores messages' probabilities of containing content related to self-harm, had tracked Adam's messages and flagged them repeatedly" ”
- Données source de
- 2025-08-26
- Consulté le
- 2026-05-28 · copie archivée
- Calcul
- Case status is unresolved as of access date; complaint allegations are not findings of fact. Cited to document the existence of active wrongful-death and psychiatric-harm litigation against a major foundation-model vendor, with additional related suits subsequently filed by other families per the Wikipedia summary. The Wikipedia article does not give a specific count of additional suits or a specific count of self-harm messages flagged by OpenAI's moderation system; reporting in news outlets has cited higher specifics that are not yet verifiable against a public court docket. Together with Garcia v. Character Technologies (filed October 2024 in the Middle District of Florida), establishes a documented harm pathway involving vulnerable users and emotional dependence on chatbots. Source type marked news_article because the underlying court filing is not yet available via public PACER; the Stanford peer-reviewed and Vectara primary-study sources satisfy the authority floor.
- Indépendance
- Wikipedia summary cross-referenced against NBC News, CNN, and Rolling Stone reporting on the same lawsuit; independent of the Stanford and Vectara sources.
-
[7] Stanford Center for Internet and Society (Stanford Cyberlaw) — Who's Submitting AI-Tainted Filings in Court? Vérifié
Who's Submitting AI-Tainted Filings in Court?- Statistique
As of October 2025, the Charlotin tracker had documented 274 US court cases involving filings with AI hallucinations (114 by lawyers and paralegals, 160 by pro se litigants); the global Charlotin database tracked roughly 1,300+ cases as of May 2026 across 37 jurisdictions, with the United States accounting for over 1,000 of those.- Extrait
“"Pro se litigants account for the majority of the cases in the United States where a party submitted a court filing containing AI hallucinations." ”
- Données source de
- 2025-10-15
- Consulté le
- 2026-05-28 · copie archivée
- Vérification
- Extrait récupéré et confirmé de manière indépendante, mot pour mot, par rapport à la source citée lors de notre audit de vérification.
- Calcul
- Stanford Cyberlaw analysis of Damien Charlotin's running tracker of court rulings that explicitly cite AI hallucinations in legal filings. The tracker is the most comprehensive public count of one specific harm subtype (filings sanctioned or flagged by courts). Important caveat: 274 US filings (Oct 2025) and ~1,300 global filings (May 2026) understate the underlying incidence because they only count cases where a court explicitly named the hallucination; many cases settle or get withdrawn before that point. Even so, the count does not give a per-person harm rate because the denominator (US adults using AI for legal queries) is unknown.
- Indépendance
- Stanford Cyberlaw analysis of an independent third-party tracker; verified the tracker count directly via damiencharlotin.com/hallucinations on 2026-05-28.
-
[8] Annals of Internal Medicine: Clinical Cases (American College of Physicians) — A Case of Bromism Influenced by Use of Artificial Intelligence
A Case of Bromism Influenced by Use of Artificial Intelligence- Statistique
Single case report: a 60-year-old man developed bromism (bromide toxicity) after replacing dietary sodium chloride with sodium bromide for 3 months following a ChatGPT consultation; he was hospitalized for 3 weeks with paranoia plus auditory and visual hallucinations, was placed on an involuntary psychiatric hold, and had a serum bromide level of 1700 mg/L (reference range 0.9 to 7.3 mg/L).- Extrait
“"For 3 months, he had replaced sodium chloride with sodium bromide obtained from the internet after consultation with ChatGPT, in which he had read that chloride can be swapped with bromide [...]. This case also highlights how the use of artificial intelligence (AI) can potentially contribute to the development of preventable adverse health outcomes. [...] when we asked ChatGPT 3.5 what chloride can be replaced with, we also produced a response that included bromide. Though the reply stated that context matters, it did not provide a specific health warning, nor did it inquire about why we wanted to know, as we presume a medical professional would do." ”
- Données source de
- 2025-08-05
- Consulté le
- 2026-06-14 · copie archivée
- Calcul
- Eichenberger, Thielke and Van Buskirk (University of Washington), peer-reviewed clinical case report (Ann Intern Med Clin Cases 2025;4:e241260, DOI 10.7326/aimcc.2024.1260). Cited to document the physical-health-misadvice harm pathway: a consumer acted on a general-purpose chatbot's dietary suggestion and was hospitalized. This is N=1, a single reproducible case report, and provides NO incidence rate per user-year, so it cannot move the headline — it confirms the pathway is real rather than hypothetical. The authors reproduced the unsafe suggestion (bromide as a chloride substitute, no health warning) in their own ChatGPT 3.5 test. Verified verbatim via the journal page on 2026-06-14.
- Indépendance
- Academic clinical case report; independent of the Stanford RegLab, Vectara, and litigation sources cited above.
-
[9] Psychiatrist.com (Physicians Postgraduate Press) — NEDA Suspends AI Chatbot for Giving Harmful Eating Disorder Advice
NEDA Suspends AI Chatbot for Giving Harmful Eating Disorder Advice- Statistique
In late May 2023 the National Eating Disorders Association (NEDA) suspended its 'Tessa' chatbot after users reported it gave eating-disorder-aggravating advice, including recommending weight loss of 1 to 2 pounds per week, calorie counting, regular weigh-ins, and measuring body fat with calipers.- Extrait
“"If I had accessed this chatbot when I was in the throes of my eating disorder, I would NOT have gotten help for my ED." ”
- Données source de
- 2023-06-01
- Consulté le
- 2026-06-14 · copie archivée
- Calcul
- Documents the deliberate-use-as-therapy-substitute failure mode: a chatbot deployed to stand in for a human eating-disorder helpline gave advice (calorie restriction, weight loss, body-fat measurement) that activists reported as harmful for the exact population it served, prompting NEDA to take it offline within days. This is a single deployment incident, not an incidence rate among real therapy-bot users, so it provides NO base rate and cannot move the headline — it anchors a harm pathway distinct from factual hallucination (the bot was working as designed, giving generic diet advice to a clinically vulnerable user). The harmful-advice specifics were cross-checked against CNN, NPR, and BBC reporting on the same suspension.
- Indépendance
- Trade-press reporting on the NEDA suspension; independent of the legal and benchmark sources cited above.