Real AI hallucination examples
Short answer
The clearest AI hallucination examples come from court filings. In Mata v. Avianca (June 22, 2023), lawyers cited six court decisions that ChatGPT had invented and were fined $5,000. Similar fake citations have since reached law firms, a government report and search results, and courts have addressed AI-hallucinated content in 2,241 decisions.
The cases at a glance
| Case | What the AI got wrong | Consequence | Date | Source |
|---|---|---|---|---|
| Mata v. Avianca (S.D.N.Y.) | Six non-existent court decisions generated by ChatGPT | $5,000 sanction | Jun 22, 2023 | Opinion and Order on Sanctions |
| Wadsworth v. Walmart (D. Wyo.) | 8 fake cases in motions, drafted with Morgan & Morgan's in-house AI | $3,000 fine and pro hac vice revoked for the drafter, $1,000 each for two others | Feb 24, 2025 | LawNext |
| Johnson v. Dunn (N.D. Ala.) | Fabricated citations from ChatGPT in a large-firm filing | 3 Butler Snow attorneys disqualified and referred to the bar | Jul 23, 2025 | EDRM |
| Deloitte report for the Australian government | Fabricated references and a made-up court quote | Partial refund of an AU$440,000 contract | Oct 6, 2025 | The Register |
| Meta AI on breaking news | Said in some answers that the July 2024 Trump assassination attempt did not happen | Meta called it hallucination and said it was fixing it | Jul 30, 2024 | Meta |
 Google AI Overviews | Viral answers recommending glue on pizza and eating rocks | Google made more than a dozen technical fixes | May 30, 2024 | |
| Grok 3 in AI search | 154 of 200 answers linked to error pages | Broken or invented URLs, even beside correct answers | Mar 6, 2025 | Tow Center |
Hallucinations in court
Mata v. Avianca: the case that started the count
In a suit against the airline Avianca in a New York federal court, the plaintiff's lawyers filed a brief that cited six court decisions. None existed. ChatGPT had generated them, and they read like real opinions. When the court questioned them, the lawyers kept standing by the fake opinions instead of withdrawing them. On June 22, 2023, Judge P. Kevin Castel imposed a $5,000 sanction (Mata v. Avianca, Inc., Opinion and Order on Sanctions).
Lesson: a citation that looks complete is not evidence that it exists. Search the case name in a court database before you rely on it.
Wadsworth v. Walmart: an in-house tool, the same failure
The risk is not limited to public chatbots. In a case against Walmart and Jetson Electric Bikes in federal court in Wyoming, attorneys at Morgan & Morgan filed motions in limine citing 8 cases that did not exist. The drafts came from MX2.law, the firm's in-house AI. On February 24, 2025, Judge Kelly Rankin fined the drafting lawyer $3,000 and revoked that lawyer's pro hac vice admission, and fined two other attorneys $1,000 each, as reported by LawNext.
Lesson: a private or firm-built AI tool hallucinates too. Every lawyer who signs a filing owns its citations.
Johnson v. Dunn: when an AI policy is not enough
Large firms with written AI policies have been caught as well. In Johnson v. Dunn, Judge Anna Manasco of the Northern District of Alabama found fabricated citations generated with ChatGPT in a filing by Butler Snow lawyers. On July 23, 2025, she disqualified three attorneys from the case and referred them to the bar, according to a write-up by a retired judge on EDRM.
Lesson: a policy only works if someone checks each citation against the source before filing.
The running count
These are not isolated cases. Damien Charlotin's AI Hallucination Cases database tracks legal decisions that address AI-hallucinated content. It listed 2,241 such decisions when checked on October 29, 2026. The database counts decisions, not every fake citation filed, so the true number of fabricated references is higher.
Hallucinations in reports and newsrooms
Deloitte's report for the Australian government
Deloitte Australia produced a report for the Department of Employment and Workplace Relations that contained fabricated references and a made-up quote from a court judgment. University of Sydney academic Christopher Rudge spotted the errors. The revised report disclosed that it had used Azure OpenAI's GPT-4o, and Deloitte agreed to a partial refund of the AU$440,000 contract (The Register, October 6, 2025).
Lesson: fake references survive review when no one opens them. One reader who checks the footnotes can catch what a whole team missed.
Altered quotes in news answers
Newsrooms have measured the problem directly. In the BBC's first study, 19% of AI answers citing BBC content introduced factual errors, and 13% of quotes attributed to BBC articles were altered or missing from the original (BBC, February 11, 2025). In the follow-up EBU/BBC study, Gemini had significant problems in 20% of its 290 responses that contained direct quotes, against an average of 12% across the four assistants (October 21, 2025).
Lesson: never paste an AI quote into your work. Find the words in the original first.
Meta AI and a breaking news event
After the July 2024 assassination attempt on Donald Trump, Meta AI told some users the event had not happened. Meta acknowledged the answers, called them hallucinations, described them as an industry-wide issue and said it was fixing them (Meta, July 30, 2024).
Lesson: models know least about the newest events. For breaking news, go to a news source, not a chatbot. More on Meta's assistant is on is Meta AI accurate?
Repeating propaganda as fact
NewsGuard found Microsoft Copilot repeated false news claims 36.67% of the time in its August 2025 audit, and also cited content from the Pravda network via VK (NewsGuard, September 4, 2025). When an assistant searches the web, it can pick up a false narrative and pass it on with confidence.
Lesson: check who the source is, not just whether there is one.
Hallucinations in search results
Google AI Overviews: glue on pizza
In May 2024, screenshots spread of Google's AI Overviews suggesting glue to keep cheese on pizza and recommending eating rocks. Google said AI Overviews performed on par with featured snippets in its own tests (vendor claim) and made more than a dozen technical fixes (Google, May 30, 2024).
The issue did not end there. In the New York Times and Oumi analysis reported by Futurism (April 8, 2026), AI Overviews answered 91% of SimpleQA questions correctly, but 56% of answers were not supported by the sources they cited. Google said the study has serious holes. See are Google AI Overviews accurate?
Lesson: a citation under a summary does not mean the summary came from it. Click through.
Grok and invented links
In the Tow Center test of AI search tools, Grok 3 answered 94% of source queries wrong, and 154 of its 200 answers led to error pages, from fabricated or broken URLs (March 6, 2025). Even its correct answers often linked to made-up addresses. More on is Grok accurate?
Lesson: a link that fails to load is a red flag for the whole answer.
What the examples have in common
Every case above shares three features:
- The output looked right. Fake cases had names and docket numbers; fake links had real-looking URLs.
- A quick check would have caught it. One search for the case, the quote, the link or the event shows the answer is wrong.
- The cost landed on the person who used the answer. Courts sanctioned lawyers, not chatbots, and a consultancy refunded fees, not a model maker.
The pattern matches the research. Language models produce likely-sounding text, and they are rewarded for answering rather than admitting doubt, as why does AI hallucinate? explains. Fake references are the most common failure in professional settings, covered in detail on does ChatGPT make up sources?. The fix in every case is the same short routine on how to fact-check an AI answer.
No AI answer is safe to file, publish or cite until someone has checked it against the original source.
Questions
What is the most famous AI hallucination example?
Mata v. Avianca is the best-known case. Lawyers in a New York federal court filed a brief citing six decisions that ChatGPT had invented, then stood by them when questioned. On June 22, 2023, the judge imposed a $5,000 sanction (Opinion and Order). It is still the case most lawyers cite when warning about AI.
How many court cases involve AI hallucinations?
Damien Charlotin's AI Hallucination Cases database listed 2,241 legal decisions addressing AI-hallucinated content when checked on October 29, 2026. It counts decisions rather than every fake citation, and the number changes daily, so it is a floor, not a full tally of fabricated references in court filings.
Have lawyers been punished for AI hallucinations?
Yes. Lawyers in Mata v. Avianca were fined $5,000 in June 2023. In Wadsworth v. Walmart, the drafting lawyer was fined $3,000 and lost pro hac vice admission in February 2025. In Johnson v. Dunn, three Butler Snow attorneys were disqualified and referred to the bar in July 2025.
Do AI search engines hallucinate too?
Yes. Across 8 AI search tools, more than 60% of answers naming a news article's source were wrong in the Tow Center test (March 2025). Google AI Overviews gave answers not supported by their cited sources 56% of the time in the New York Times and Oumi analysis reported in April 2026.
What can I learn from these examples?
Open every source before you rely on an answer. Each case above involved output that looked complete, and each would have been caught by one search for the case, quote or link. Treat AI as a first draft, verify at the origin, and use the five steps on how to fact-check an AI answer.
Sources
- Independent study
- Independent study
- Vendor claim
- News audit
- Vendor claim
- News audit
- Independent study
