Documented AI hallucination examples
Cases where an AI system's invented output produced a legal ruling, a sanction or a published correction, each one on the record, none of them anecdotal.
The internet is full of screenshots of chatbots saying silly things. Screenshots are not evidence. What follows are incidents that produced a court decision, a formal sanction or a documented study finding, cases where someone independent checked.
An airline was held liable for its chatbot’s advice
Moffatt v. Air Canada, British Columbia Civil Resolution Tribunal, February 2024.
Jake Moffatt’s grandmother died in November 2022. He asked Air Canada’s website chatbot about bereavement fares and was told he could book at full price and apply for the reduced rate retroactively. He booked, spending roughly CA$1,630. Air Canada’s actual policy, set out on a page the chatbot had itself linked to, does not allow retroactive bereavement claims.
The airline argued the chatbot should be treated as a separate entity responsible for its own statements. The tribunal rejected that outright, holding that while a chatbot has an interactive component, “it is still just a part of Air Canada’s website,” and that it should be obvious the airline is responsible for all the information on it. It found Air Canada “did not take reasonable care to ensure its chatbot was accurate,” and awarded Moffatt CA$812.
This is the closest thing to a landmark ruling on the question of who owns a chatbot’s mistake.
Six court decisions that never existed
Mata v. Avianca, Southern District of New York, 2023.
A brief cited six fabricated opinions produced by ChatGPT, each attributed to a real federal judge. One, presented as an Eleventh Circuit decision, was described by the court as containing legal analysis that was “gibberish.” When challenged, counsel produced copies of the fake opinions rather than verifying them. Judge Castel imposed a $5,000 penalty jointly and severally on the two attorneys and their firm, finding “acts of conscious avoidance and false and misleading statements to the Court.”
Nearly half of AI news answers had a significant problem
European Broadcasting Union / BBC study, published October 2025.
Twenty-two public service media organisations across 18 countries and 14 languages evaluated more than 3,000 responses from ChatGPT, Copilot, Gemini and Perplexity to questions about the news.
The findings:
- 45% of answers had at least one significant issue.
- 20% contained major accuracy problems, including hallucinated details and outdated information.
- 31% had sourcing problems, including attributing claims to news organisations that never made them.
- Counting lower-severity issues, 81% of responses had at least one problem.
Sourcing was the most uneven category. Gemini had a significant sourcing issue in 72% of its responses, against under 25% for the others.
An earlier BBC-only study in February 2025, focused on questions about BBC coverage, found 51% of responses had significant issues, so the picture did not improve markedly across that year.
Fifteen fake citations across five months
Couvrette v. Wisnovsky, District of Oregon, December 2025.
Not a single lapse but a sustained pattern: fifteen AI-generated fake citations and eight fabricated quotations spread across three briefs filed over five months. The court fined lead counsel $15,500 and awarded the defendants their attorney fees.
The through-line
None of these are cases of a model being asked something obscure. They are cases of a model being asked something ordinary (an airline policy, a case citation, what happened in the news) and producing a fluent, well-formatted, entirely wrong answer that nobody could distinguish from a right one by looking at it.
The failures reach the public record only where an independent party had both the means and the motive to check. In courts, that is routine. Almost nowhere else is.