AI Hallucination Examples: 4 Real Cases and How to Spot Them

An AI hallucination is an answer that sounds sure of itself and is false or has nothing behind it: a case that was never decided, a quote nobody said, a date that is off by two years. The model isn't lying. It is predicting plausible words, and a made-up detail reads exactly like a real one. Below are four documented cases, what they share, and a short check you can run on any AI draft before it goes out.

Four real examples of AI hallucinations

1. The legal brief with invented cases (2023). Two New York lawyers filed a brief in Mata v. Avianca that cited court opinions ChatGPT had made up. On June 22, 2023, Judge P. Kevin Castel fined the lawyers and their firm $5,000 and found bad faith, citing conscious avoidance of signs the cases were fake and misleading statements to the court. The judge noted there is nothing inherently improper about using a reliable AI tool for help. The failure was not checking what it produced.

2. The chatbot that invented a refund rule (2024). An Air Canada chatbot told a customer he could claim a bereavement fare after his trip. The airline's own web page said refunds weren't available for travel that had already happened. In February 2024 a British Columbia tribunal held Air Canada responsible for what its chatbot said and rejected the idea that the bot was a separate entity. A hallucination in a customer-facing tool becomes your statement.

3. The consulting report with fake sources (2025). Deloitte Australia produced a review for the Department of Employment and Workplace Relations that contained fabricated citations and a made-up court quote. An academic noticed that some of the cited works didn't exist. Deloitte revised the report, disclosed that it had used a generative AI model, and agreed to repay the final installment of its A$440,000 contract. Even professional teams with review processes missed it.

4. The product demo with a wrong "first" (2023). In Google's launch promo for its Bard chatbot, the bot said the James Webb Space Telescope took the very first picture of a planet outside our solar system. Astronomers pointed out that the first exoplanet image came from a different telescope in 2004, and Alphabet's share price dropped that day. The sentence was smooth, specific and easy to believe, which is what makes this kind of error hard to see.

What these cases have in common

Every one of them hid inside normal-looking text. Four patterns show up again and again:

• Specifics with a name on them. Case names, report titles, authors and quotes. A citation looks right whether or not it exists.
• Superlatives and firsts. "The first," "the only," "the largest." They sound authoritative and are easy to get wrong.
• Rules and policies. Refund terms, legal limits, deadlines. The model fills a gap with what such a rule would plausibly say.
• Nobody opened the source. In each case the claim would have failed a two-minute lookup.

A five-minute check for your own draft

You don't need a special tool to do this. It works on any AI-assisted draft, from an email to a report.

1. Highlight every checkable claim. Names, numbers, dates, quotes, "first/only/most" statements, rules, prices and anything with a source attached.
2. Mark the ones that would hurt if wrong. A client figure or a legal reference matters far more than a general observation.
3. Open the source for each. Search the exact title, case name or quote. If the model gave a link, open it and read the line it supports.
4. Treat "can't find it" as a finding. If you can't find a source that says the same thing, cut the claim, reword it as unconfirmed, or replace it with one you can cite.
5. Re-read the fixed text. Corrections often leave awkward seams, and a draft that is accurate should also sound like you.

For a longer version with a prompt that asks a model to list its own doubts, see our 6-step checklist for fact-checking ChatGPT. And if your concern is how a draft reads rather than whether it is true, the 10-point checklist for spotting AI habits in your writing covers that side.

Doing it in Unflag

When there are a lot of claims, Fact Check does the first pass for you. It is part of Unflag Plus, so free accounts can't use it.

1. Open Fact Check and sign in. On the web it needs at least 50 words, and Plus accepts up to 10,000 characters per input.
2. Paste the AI draft and run the check.
3. Unflag extracts the claims in your text and searches the web for each one. You get an overall confidence score, a summary, an analysis for each claim and the sources.
4. Each claim gets a verdict: True, False, Partially True, Misleading or Unverifiable.
5. Open the sources behind anything important. "Unverifiable" means exactly that, not "false," and a verdict is a lead to confirm, not a replacement for reading the source.
6. Correct the draft, then use Rewrite or Humanize AI to make it read naturally. Humanize AI tries to keep figures such as prices and percentages, but check your numbers again after any rewrite.

As of October 2026, Unflag Plus is $8.99 a month or $45.99 a year on the web. The apps are billed weekly or yearly through the stores.

Do newer AI models still hallucinate?

Yes, though how often depends on the model, the topic and whether it can search. Models that can browse or read sources you give them make fewer unsupported claims than a model answering from memory. None of the vendors claim the problem is gone, so the habit of checking names, numbers and quotes still applies. Treat any rate you see quoted as the vendor's or researcher's own figure for one test, not a promise about your draft.

Where hallucinations hide in everyday writing

• Work emails and reports: statistics, competitor facts and "industry standard" claims.
• Cover letters and bios: dates, job titles and awards the model filled in.
• Blog posts and marketing copy: "studies show" lines and invented expert quotes.
• Summaries: a point added that the source never made.
• Medical, legal and money topics: use AI to prepare questions, and a qualified professional or official source for the answer.

What this check can't do

No tool, including Unflag, can promise every error is caught. A web-based fact checker is limited by what it can find, and a claim can be both unverifiable and true. If your course, employer or publisher has an AI policy, follow it, including any rule about saying when AI helped.

How we put this together

We read the published reporting on each case (Bloomberg Law, Daily Hive, TechRadar and Space.com, as of October 2026), compared the steps against Unflag's own Fact Check flow in the product and its code, and checked Unflag's pricing as of October 2026. Reports differ on some details, such as the exact amounts in the Air Canada and Deloitte cases, so we only quote figures that sources agree on. The checklist was written for this post and no tool output is attached to it.

The short version: a hallucination reads like any other sentence, so check the names, numbers, quotes and "firsts" yourself, and use Fact Check when there are too many to chase by hand.

Frequently asked questions

What is an AI hallucination?

An AI hallucination is output that sounds confident and is false or unsupported: an invented citation, a wrong date, a quote nobody said, or a rule that doesn't exist. The model predicts plausible words, so a made-up detail can read exactly like a true one.

What is an example of an AI hallucination?

In 2023, two New York lawyers filed a brief citing court cases that ChatGPT had invented, and a federal judge fined them $5,000. In 2025, a consulting firm refunded part of a fee after a government report contained fabricated citations and a made-up court quote.

Why does AI make things up?

A language model writes the most likely next words, not a verified fact. When it has no solid answer it still produces something fluent. It is most likely to slip on citations, exact numbers, quotes, recent events and niche topics where it has little data.

How can I tell if ChatGPT is hallucinating?

You can't tell from tone, because wrong answers sound as sure as right ones. Look for specifics you can check: names, numbers, dates, quotes and sources. Search each one outside the chat. If you can't find a source that says the same thing, treat the claim as unconfirmed.

Can an AI hallucination be fixed?

Not fully. Giving a model real sources and letting it search reduces mistakes, but doesn't remove them. The reliable fix is a human step: find the claims that matter, check them against primary sources, then correct or cut what you can't confirm.

Does Unflag catch AI hallucinations?

Unflag's Fact Check, part of Unflag Plus, pulls out the claims in a text, searches the web for each one and returns a verdict, a confidence score and sources. It can miss things or be wrong, so open the sources for any claim that matters before you rely on it.