AI Hallucination Examples: Real Cases That Made the News

An AI hallucination is output that is fluent, confident, and false — an invented court case, a citation to a paper that doesn't exist, a company policy the chatbot made up on the spot. These aren't hypotheticals: every example below is documented, several cost real money or careers, and each one was catchable with a five-minute check.

1. The lawyer who cited six fake cases (Mata v. Avianca, 2023)

The canonical example. A New York attorney used ChatGPT for legal research and filed a brief citing six precedents — all invented by the model, complete with realistic citations and quotes. The court sanctioned the lawyers, and "did you check whether the case exists?" became a standing question in courtrooms. Dozens of similar sanctions have followed since; fabricated authority remains the single most common professional AI failure.

2. Fake citations in a U.S. government health report (2025)

The "MAHA" federal health report was found to contain references to studies that don't exist — nonexistent papers attributed to real researchers, who publicly disavowed them. Detection researchers (GPTZero among others) flagged the pattern as classic model-fabricated citation structure. A national-level document, undone by unchecked references.

3. Deloitte's partial refund on a consulting report (2025)

Deloitte Australia agreed to partially refund a government contract after a ~A$440k report was found to contain AI-hallucinated citations and a fabricated quote attributed to a court judgment. The firm re-issued the report with corrections — and the story became the reference case for "AI slop" in professional services.

4. Air Canada's chatbot invents a refund policy (2024)

An airline support chatbot told a customer he could apply for a bereavement fare refund retroactively — a policy that didn't exist. A tribunal ordered the airline to honor the invented policy, rejecting the argument that the chatbot was "a separate legal entity." Hallucinations aren't just embarrassing; they can be binding.

5. Bard's telescope error in its own launch demo (2023)

Google's first Bard demo confidently claimed the James Webb Space Telescope took the first picture of an exoplanet (it didn't — that was 2004, by a ground telescope). Alphabet's market value dropped roughly $100 billion the next day. The lesson aged well: fluency is not accuracy, even on stage.

6. Invented sources in student bibliographies (ongoing)

The version teachers see weekly: essays citing plausible-looking papers — real journal, real-sounding authors, valid-looking DOI — that resolve to nothing. Models generate citation-shaped text; they do not look sources up. This is now one of the fastest ways students get referred for academic-integrity review, independent of any AI detector. (Writing one? Our annotated bibliography guide covers doing it right.)

7. Confident statistics with no origin

Ask a model for supporting data and it will supply precise numbers — "a 2023 Stanford study found 47%…" — that trace back to nothing. Fabricated statistics are harder to spot than fabricated citations because they often go uncited entirely. Rule of thumb: any number an AI gives you is a claim to verify, not a fact to repeat.

How to catch hallucinations before they cost you

• Verify every citation resolves. Paste references into our free fake-citation checker — it confirms DOIs resolve and sources exist. This single step would have prevented cases 1, 2, 3, and 6.
• Trace statistics to a primary source. If you can't find the study in 5 minutes of searching, treat the number as fiction.
• Check quotes verbatim. Models paraphrase and attribute; real quotes match the source word-for-word.
• Treat confident specificity as a red flag, not reassurance. Hallucinations are precise. Vague answers are often safer ones.
• Know what you're reading. If you're reviewing submitted work, the AI detector shows which passages read as generated — fabrications cluster in generated text.

Frequently asked questions

What is an AI hallucination?

Output that is grammatically fluent and factually false — invented sources, fake events, made-up policies — produced because language models predict plausible text rather than retrieve verified facts.

How common are AI hallucinations?

Rates vary by model and task, but citation-style tasks are the worst case: studies have repeatedly found large fractions of model-generated references are wrong or nonexistent. Newer models hallucinate less, not never — verification is still on you.

Is there a tool to check for AI hallucinations?

For the most damaging category — fabricated references — yes: our citation checker verifies sources are real, free. For claims and statistics, the tool is a search engine and the primary source.