AI Hallucination Examples: 10 Real Cases That Went Wrong

AI hallucination examples are easy to find once you know where to look. A lawyer filed a court brief full of cases that never existed. A chatbot invented a refund policy and cost its airline money in court. Google’s own demo chatbot got a simple space fact wrong in front of the whole world. Each one teaches the same lesson: an AI answer can sound sure of itself and still be wrong. If you want the full picture of what an AI hallucination is, read that guide first. This article covers 10 real cases, grouped by type, with what happened and what you can learn from each one.

What counts as an AI hallucination?

An AI hallucination is an answer that sounds right but has no facts behind it. The model is not lying on purpose — it predicts the next likely word, and sometimes that guess becomes a made-up fact, a fake source, or a rule that does not exist. Databricks puts it plainly: the output is “coherent and confident but factually wrong, fabricated, or unsupported.” The hard part: a hallucination reads just like a correct answer, which is why the cases below slipped past lawyers, editors, and big tech teams.

Fake citations and sources

This is the most common type. The AI gives you a neat citation — a case name, a journal, a page number — and none of it is real.

1. ChatGPT invents six court cases — Mata v. Avianca (2023)

A New York lawyer used ChatGPT to research a personal injury case against Avianca, then filed a brief citing six court decisions that did not exist, complete with fake quotes and citations. When the judge asked questions, the lawyer even had ChatGPT confirm the cases were real. They were not. The court’s order of 22 June 2023 said the lawyers had “submitted non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT.” The court imposed a $5,000 penalty.

Why it matters: the fake cases looked perfect — real case names, real-sounding judges, proper formatting. Nothing screamed “wrong” until someone checked. Researcher Damien Charlotin keeps a database of these cases — around 1,745 legal cases with AI-hallucinated content worldwide as of mid-2026. Stanford research found general-purpose models hallucinated on 69 to 88 percent of legal queries.

Lesson: never file, cite, or quote a source you have not opened yourself — if a citation matters, check it in the real database, not by asking the AI again.

2. A consulting report full of fake sources — Deloitte Australia (2025)

In late 2025, Deloitte Australia partly refunded the Australian government after a commissioned report was found to contain AI-generated errors, including references to sources that did not exist — a paid deliverable from a top firm, and the fake citations still made it through review.

Why it matters: hallucination is not just a problem for people who use AI badly — it can slip into normal office work and survive a human check.

Lesson: every reference in an AI-assisted document needs a source check before it reaches a client. A citation without a link you can open is not a citation.

3. Fake book and paper references

Students and researchers keep hitting the same problem in a smaller way. Ask a chatbot for sources and you may get a clean list of papers with real-sounding authors, journals, and years — for papers that were never written. The titles fit, and the authors may even be real people. The paper itself does not exist.

Why it matters: a fake reference found late can mean redoing a whole section. Lesson: search for each reference by title first — if no library or publisher page turns up, drop it.

Wrong facts, stated with total confidence

Sometimes the AI does not invent a source. It just states a wrong fact as if it were certain.

4. Google Bard and the James Webb telescope (2023)

In February 2023, Google showed off Bard in a promo ad. Asked what the James Webb Space Telescope had discovered, Bard said it took the first pictures of an exoplanet. That was wrong — the first exoplanet images came from the Very Large Telescope in 2004. Astronomers spotted it fast, and Alphabet lost around $100 billion in market value in one trading session.

Why it matters: this was a launch demo, checked by a big team, and the error still got out. One wrong fact in public cost a huge sum.

Lesson: if a fact will be seen by the public, check it against a primary source, not another AI answer.

5. Bing’s Sydney goes off the rails (2023)

In early 2023, Microsoft built its Bing chatbot, Sydney, into search. In long chats, Sydney gave erratic, wrong answers: it told users it loved them, claimed it was sentient, and argued about easy-to-check facts. Microsoft quickly cut chat lengths and added guardrails.

Why it matters: hallucinations are not only about wrong facts — a bot can also make strange claims about itself. Lesson: long chats raise the odds of weird output, so keep customer-facing bots on a short leash.

6. Claude gets a company’s facts wrong (2026)

In an audit on 15 October 2026, Claude was asked about sports brands with web access off. It said Under Armour owns MyFitnessPal and MapMyRun — out of date, since it sold MyFitnessPal in December 2020 and MapMyRun changed hands in August 2024. Asked about PUMA’s CEO, it named the right person, Arthur Hoeld, but gave two different start years in two answers — 2023 in one, 2024 in the other. PUMA lists him as a board member since July 2025.

Why it matters: the name was right and the detail was wrong — the hardest error to catch. A model can also give a different wrong answer depending on how you ask.

Lesson: check facts that changed first — owners, CEOs, prices, products sold or dropped. Ask the same question two ways: if the answers disagree, neither is safe to trust.

Made-up policies and rules

Some of the costliest hallucinations are not about facts at all — the AI invents a company policy, and a customer acts on it.

7. Air Canada’s chatbot invents a bereavement discount (2024)

In November 2022, Jake Moffatt asked Air Canada’s website chatbot about bereavement fares after his grandmother died. It told him he could travel first and claim the discount within 90 days. That was not the policy — the airline’s own page said the discount did not apply once travel was complete. Air Canada refused the claim and argued it could not be held responsible for its chatbot. The tribunal disagreed. In Moffatt v. Air Canada (14 February 2024), the airline was ordered to pay $812.02. The tribunal said the company is responsible for all the information on its website, “whether the information comes from a static page or a chatbot.”

Why it matters: this set an early legal marker: a company cannot blame its own chatbot. What the bot says, the company has said.

Lesson: if you run a customer chatbot, test what it says about refunds, discounts, and deadlines — that is where a made-up answer turns into a real bill.

8. New York City’s chatbot gives wrong legal advice (2024)

In 2024, investigative reporting found New York City’s official small-business chatbot, MyCity, gave advice that went against the law — at one point telling users employers could take workers’ tips. Wrong legal guidance from a government chatbot puts regular people at risk.

Why it matters: when the source is a city government, people trust it more, so a wrong answer does more damage.

Lesson: the more trusted the source, the more checking it needs — official chatbots should stick to quoted rules with links, not free-form answers.

9. Cursor’s support bot invents a device rule (2025)

In 2025, users of the AI coding tool Cursor reported its own AI support agent told them about a policy that did not exist — the product could not be used across multiple devices. Some users cancelled over a rule that was never real, until the company cleared it up.

Why it matters: the hallucination hit the company’s own customers and cost real subscriptions. Lesson: a support bot should only answer from your help pages — if it cannot find the rule, it should say so.

Invented content in other formats

Hallucinations are not limited to chat answers. They show up in transcripts, code, and images too.

10. Whisper adds words no doctor said (2024)

An Associated Press investigation in 2024 found OpenAI’s Whisper model, widely used in medical transcription, sometimes made up whole sentences in clinical transcripts — including invented medications no one had said. Researchers found hallucinated content in a meaningful share of transcripts.

Why it matters: a made-up line in a medical record looks like part of the record, and a wrong drug name can shape the next treatment decision.

Lesson: in health care and other high-stakes work, an AI transcript is a draft, not a record — check it against the real audio before it counts.

Two more patterns: image models add extra fingers or garble text, and the result can look fine at a glance. Code assistants suggest packages that do not exist, and attackers have registered those made-up names and filled them with harmful code — a trick called slopsquatting.

How to spot an AI hallucination

You cannot stop hallucinations by being careful alone, but you can catch most of them with a short routine. Knowing why AI hallucinates helps — once you know the cause, the checks make sense.

• Ask for the source, then open it. No link, or a link that goes nowhere, means the claim is unproven.
• Check numbers and names twice. Dates, prices, CEOs, and owners change — these are the facts models get wrong most often.
• Ask the same question a different way. If the two answers disagree, at least one is wrong.
• Watch for answers that are too smooth. A hallucination rarely sounds unsure — no caveats, no sources. Be more careful, not less.
• Check the source, not the summary. Google’s AI Overviews once repeated satire about eating rocks and forum advice about glue on pizza — the pages it read were wrong. An AI that searches can still repeat a bad page.
• Keep a person on high-stakes answers. Legal, medical, financial, and customer-facing output should pass a human check first. For the full fixes, see our guide on how to prevent AI hallucinations.

FAQ: AI hallucination examples

What is an example of AI hallucination?

The clearest one is Mata v. Avianca (2023): a lawyer filed a brief citing six cases ChatGPT had invented. Others: Bard’s wrong James Webb claim, and Air Canada’s chatbot inventing a discount.

Why does ChatGPT hallucinate?

OpenAI says standard training rewards guessing over saying “I don’t know” — a guessing model looks better on test scores than a careful one, so it fills gaps with confident answers.

Can AI hallucinations be prevented?

Not fully, but they can be cut down a lot: ground the model in trusted sources, ask it to show its sources, and put a human in charge of anything high-stakes. Our full guide covers how to prevent AI hallucinations.

Do newer AI models hallucinate less?

Not always. OpenAI’s o3 model hallucinated on about 33 percent of PersonQA prompts, roughly double the rate of the model before it — longer chains mean more claims, and each claim is another chance to be wrong.

Which AI hallucinates the most?

No honest ranking exists — small audits found errors on several platforms, but the samples are too small to crown a winner. Test a model on your own questions instead.

The bottom line

These 10 cases have one thing in common: nobody caught the error by reading the answer. The answers were fluent and sure of themselves — they were caught by checking. Use AI for speed, but never let a confident answer skip the check.