AI hallucination examples: 65 documented cases with real consequences

Key Findings

• In Mata v. Avianca, a federal judge in Manhattan fined two lawyers and their firm US$5,000 on June 22, 2023 for filing judicial opinions with fake quotes and citations created by ChatGPT, the first widely reported AI-citation sanction. (Mata v. Avianca, Opinion and Order on Sanctions (S.D.N.Y.))
• As of February 18, 2026, the U.S. Court of Appeals for the Fifth Circuit noted that researcher Damien Charlotin had identified 239 cases of AI hallucination by lawyers in U.S. courts, excluding self-represented litigants and judges. (Fletcher v. Experian, No. 25-20086 (5th Cir. 2026))
• A federal magistrate judge in Oregon ordered two lawyers to pay a combined US$110,000 in fines and fees over filings full of fake cases in December 2025, the largest AI-hallucination penalty in that court, according to the ABA Journal. (ABA Journal)
• British Columbia's Civil Resolution Tribunal held Air Canada liable for its chatbot's invented bereavement-refund policy on February 14, 2024 and ordered it to pay the customer CA$812.02. (Moffatt v. Air Canada, 2024 BCCRT 149)
• Alphabet lost about US$100 billion in market value, a 7.7% share drop, after a February 2023 Bard promotion wrongly said the James Webb Space Telescope took the first image of a planet outside the solar system. (Global News)
• Deloitte agreed in October 2025 to refund the final installment of a A$440,000 Australian government contract after its report was found to cite nonexistent sources and a fabricated court quote; the revised report disclosed use of GPT-4o. (The Register)
• On Vectara's hallucination leaderboard updated October 22, 2026, summarization hallucination rates across 111 models ranged from 1.8% to 24.2%, as scored by Vectara's HHEM-2.3 evaluator. (Vectara Hallucination Leaderboard)
• OpenAI's April 2025 system card reported that o3 hallucinated on 33% of PersonQA questions and o4-mini on 48%, compared with 16% for the older o1 model (vendor-reported). (OpenAI o3 and o4-mini System Card)
• OpenAI's GPT-5 system card (August 2025) reported a SimpleQA hallucination rate of 0.40 for gpt-5-thinking without web access, versus 0.46 for o3 and 0.52 for GPT-4o (vendor-reported). (GPT-5 System Card)
• GPTZero reported in January 2026 that it found more than 100 hallucinated citations in papers accepted to the NeurIPS 2025 AI conference, after checking 4,841 accepted papers. (GPTZero)

Figures

Documented AI hallucination cases with real consequences

What is an AI hallucination?

An AI hallucination is an output from a generative AI system that is fluent and confident but false or unsupported: a court case that does not exist, a quote no one said, a policy a company never had, or a statistic with no source. The term covers both invented facts (fabrication) and real facts misattributed or distorted, such as a genuine study cited for a finding it does not contain.

Hallucinations differ from ordinary software bugs because the output looks correct. In most cases in this dataset the false content was caught by someone outside the system: opposing counsel who could not find a cited case, a reader who knew a book did not exist, or a neurologist who knew the "basilar ganglia" is not part of the brain.

Why AI models hallucinate

Large language models generate text by predicting likely next words from patterns in training data. They do not look facts up unless connected to a search or retrieval system, so a plausible-sounding citation can be produced as easily as a real one.

In a September 2025 paper, Why Language Models Hallucinate, OpenAI researchers argued that hallucinations arise as ordinary statistical errors during pretraining and persist because most evaluations grade models like students on a multiple-choice exam, rewarding a guess over an admission of uncertainty. Their proposed fix is to change scoring so that confident errors are penalized more than abstentions.

The cases here show three practical triggers: questions about specific people or legal authorities where training data is thin (defamation and fake-citation cases), answers stitched from low-quality sources (Google's glue-on-pizza answer drew on a joke forum post), and summarization that compresses or merges facts (Apple's news-alert summaries).

Patterns across the cases

Counts below are computed from the 65 rows in this dataset. They describe what has been documented and reported, not how often AI systems hallucinate overall.

• Fake legal citations are the largest category. 32 cases involve court filings or rulings; 9 of them are from 2026, and 3 involve orders drafted in judges' own chambers.
• ChatGPT is the most frequently named system, appearing in 16 rows. In many rows the specific tool was never disclosed.
• Consequences escalated over time. Early legal sanctions were a few thousand dollars (US$5,000 in Mata v. Avianca, 2023); by 2025 and 2026 courts imposed US$30,000 to US$110,000 penalties, disqualified lawyers and dismissed appeals.
• Organizations are held to what their AI says. Air Canada's argument that its chatbot was responsible for its own words was rejected, Deloitte refunded fees, and Cursor refunded a user over a policy its support bot invented.
• Most errors were found by outsiders, such as opposing counsel, journalists, academics or affected people, rather than by the deploying organization's own checks.

Hallucination rates by model

Published hallucination rates measure different things and cannot be compared across benchmarks. Figures below are reported exactly as published.

Vectara Hallucination Leaderboard (HHEM-2.3), read on the leaderboard page last updated October 22, 2026. Vectara asks each model to summarize a set of documents and uses its own evaluation model to score how often the summary introduces content not supported by the source. Selected rows from 111 models listed (source):

OpenAI system cards (vendor-reported). SimpleQA is a set of short fact-seeking questions; PersonQA asks questions about public facts on people. OpenAI reports a hallucination rate and an accuracy for each. These are OpenAI's own measurements, published in its o3 and o4-mini system card, GPT-4.5 system card and GPT-5 system card.

OpenAI's o3 card noted that o3 makes more claims overall, producing more accurate and more inaccurate claims, and that more research is needed on why. The GPT-5 card also reported that gpt-5-thinking had a hallucination rate 65% smaller than o3, measured as the share of factual claims with errors on prompts representative of ChatGPT production traffic.

Methodology

Inclusion: a case is listed only when an AI system produced false or fabricated information that led to a documented consequence (a sanction, refund, retraction, correction, apology, removal, disqualification, complaint or lawsuit), and when it is confirmed by a primary record (court order, tribunal decision, company statement, research report) or by at least one reputable news outlet reporting specifics. Every row was checked against the source it cites in October 2026.

Dates are the date of the ruling, disclosure or first public report, at the precision the source allows (day, month or year). For judges' orders, the date is the order or the disclosure as reported.

Legal cases are identified by case name and court rather than by the names of individual lawyers. Defamation-type cases describe what the AI output said, not any allegation about the person, and state the outcome as recorded (for example, the Georgia court ruled for OpenAI in the Walters case, which is not listed because no consequence beyond the lawsuit followed).

The AI system column uses the name given by the court or reporting. Where a court found AI use likely but no tool was named, or where officials did not confirm AI use (as in the MAHA report), the column says so.

Domain is one of legal, customer service, search, news/publishing, government, health, academic or other. Counts by domain are computed from the rows in this table and reflect what has been publicly documented, which is skewed toward court filings because judges publish written orders.

Benchmark figures are copied as published. The Vectara figures were read from its public leaderboard repository, last updated October 22, 2026. OpenAI figures come from its own system cards and are labeled vendor-reported. No rates were recalculated or averaged.

Exclusions: incidents where a chatbot was manipulated into an offer or insult (prompt injection or jailbreaks) rather than producing false information, and widely shared AI errors without a documented consequence, were left out. Some cases were left out because no reputable source with specifics could be fetched, including reported school-closure report errors in Tromsø, Norway.

Cite and Reuse This Dataset

The data is free to reuse under CC BY 4.0: copy it, chart it or build on it, with a link back to this page. Download the CSV or JSON.

Citation

agitracker.io (2026). AI hallucination examples: 65 documented cases with real consequences. Updated October 29, 2026. https://agitracker.io/data/ai-hallucination-examples/

HTML link

<a href="https://agitracker.io/data/ai-hallucination-examples/">AI hallucination examples: 65 documented cases with real consequences</a> (agitracker.io)

Frequently Asked Questions

What are some real examples of AI hallucinations?

Well-documented examples include lawyers in Mata v. Avianca filing ChatGPT-invented cases (US$5,000 sanction, 2023), Air Canada's chatbot inventing a refund policy (CA$812.02 ordered, 2024), Google's AI Overviews suggesting glue on pizza (2024), the Chicago Sun-Times printing a reading list of nonexistent books (2025), and Deloitte refunding part of an Australian government contract over a report with fabricated citations (2025).

What is the most famous AI hallucination case?

Mata v. Avianca is the most cited. In 2023, lawyers in a personal-injury suit against the airline filed a brief citing court decisions that ChatGPT had invented, then stood by them after the court questioned their existence. Judge P. Kevin Castel fined the lawyers and their firm US$5,000 on June 22, 2023, and courts still cite the case when sanctioning similar conduct.

How often do AI models hallucinate?

It depends on the model and the task. On Vectara's summarization leaderboard (October 22, 2026), rates ranged from 1.8% to 24.2% across 111 models. On harder fact-recall tests the rates are much higher: OpenAI reported its o3 model hallucinated on 33% of PersonQA questions and its o4-mini on 48% (April 2025, vendor-reported). Benchmarks measure different things, so figures are not directly comparable.

Is there a database of AI incidents?

Yes. The AI Incident Database, run by the Responsible AI Collaborative, catalogs AI harms of all kinds, and lawyer-researcher Damien Charlotin maintains a database of court decisions involving AI hallucinations, which had 239 U.S. lawyer cases as of February 2026 according to a Fifth Circuit order. This page is narrower: 65 hallucination cases with verified consequences, each linked to a court record or reputable report.

Can a company be held liable for what its chatbot says?

In at least one ruling, yes. In Moffatt v. Air Canada (February 2024), British Columbia's Civil Resolution Tribunal rejected Air Canada's argument that its chatbot was a separate entity responsible for its own actions, found the airline liable for negligent misrepresentation, and ordered it to pay CA$812.02. Defamation suits over chatbot output have had mixed results, and a Georgia court ruled for OpenAI in Walters v. OpenAI in May 2025.

What happens to lawyers who cite fake AI cases?

Penalties range from warnings to large fines and disqualification. Courts have imposed fines from under US$1,000 to a combined US$110,000, ordered fee payments to opponents, removed lawyers from cases, referred them to bar regulators, dismissed appeals as frivolous and, in Colorado, suspended a lawyer. Courts increasingly say verifying every citation is a non-delegable duty.

Why does ChatGPT make up fake citations?

Language models generate text that fits learned patterns, so a citation with a plausible case name, court and year can be produced without any real source behind it. OpenAI researchers argued in September 2025 that training and evaluation reward guessing over saying "I don't know." Connecting a model to search or a verified legal database reduces, but does not remove, the risk.

Research Using This Dataset

• AI Hallucination Cases Reach Higher Courts as Penalties Grow: Analysis of 65 documented AI hallucination cases: sanctions moved to appellate and supreme courts, penalties grew, and courts stopped naming the AI tool.

Related Datasets

• Claude system prompt history and every other officially published AI system prompt: An index of officially published system prompts from Anthropic, xAI, OpenAI, Google, Meta, Mistral, Microsoft and Alibaba, with dates, lengths and notable rules.
• AI Model Release Timeline: Every Major LLM From GPT-1 to GPT-6 (2018-2026): Release dates, open-weight status, context windows, input types and launch API prices for 159 notable LLMs from 2018 to October 2026, each sourced to the lab.

Version History

• October 29, 2026: First published with 65 records, each linked to its source.

Spotted an error or a missing record? Every row links to its source; corrections are made and dated on this page. See all datasets.