AI hallucination examples: 4 documented cases and how to catch them

The AI hallucination examples worth a business owner's time are the ones with a paper trail: a court order, a city's own announcements, a corrected government report, a newspaper's account of what went wrong.

When we checked on November 3, 2026, a public database kept by researcher Damien Charlotin listed 2,262 court and tribunal decisions that dealt with AI-invented material, 1,533 of them in the US. In 897 of those decisions, the person who relied on the invented material was a lawyer.

What is an AI hallucination?

An AI hallucination is content an AI system states with confidence that is false or made up. NIST, the federal standards agency, calls it confabulation in its Generative AI Profile and defines it as "the production of confidently stated but erroneous or false content."

NIST also explains why it happens. AI models "generate outputs that approximate the statistical distribution of their training data," predicting the next word in a sentence. A prediction can be right or wrong, and it reads the same either way. NIST adds that the output can include "confabulated logic or citations that purport to justify or explain the system's answer." That is why an invented source can look more convincing than no source at all.

The costly hallucinations sound right. The fake court decisions below came with citations and quotes, and the fake books came with real authors.

Four AI hallucination examples with real consequences

The four cases below ended in a court penalty, a shutdown, a corrected report and a public apology.

Lawyers filed six court decisions that didn't exist

In March 2023, lawyers suing the airline Avianca in federal court in Manhattan filed papers that cited court decisions an AI chat app had made up. The decisions came with case names, citations and quotes. The judge found that the lawyers "continued to stand by the fake opinions after judicial orders called their existence into question."

On June 22, 2023, Judge P. Kevin Castel imposed a $5,000 penalty on the two lawyers and their firm. He also ordered them to write to their client and to every judge falsely named as the author of the six fake opinions. His order makes a point any business using AI can apply: "Technological advances are commonplace and there is nothing inherently improper about using a reliable artificial intelligence tool for assistance. But existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings."

Looking up the cases in a court database before filing would have caught all six.

A city's AI assistant gave business owners the wrong rules

In October 2023, New York City opened a business website with a pilot AI assistant for owners, the mayor's office announced. It was built to answer questions about "compliance with codes and regulations," and it drew on more than 2,000 of the city's business web pages and articles.

In a report published in March 2024, The Markup said that in its tests the assistant answered that employers could take a cut of their workers' tips and that no city rule required businesses to accept cash. New York's Labor Law bars employers from taking any part of workers' gratuities, and a 2020 city law makes it unlawful for food stores and retail businesses to refuse cash, with a few exceptions. The city kept the assistant running and soon added disclaimers telling users to "not use its responses as legal or professional advice," according to The Markup.

In January 2026, the new mayor called it "functionally unusable," and the city took it down the next month. Drawing on 2,000 official pages didn't stop it from getting the rules wrong. Check any answer about a law or regulation against the rule itself.

A consulting firm's report to a government department had faulty citations

Australia's Department of Employment and Workplace Relations hired Deloitte for an independent review of its Targeted Compliance Framework. A note on the department's site says the report "was updated on 26 September 2025 to address a small number of corrections to references and footnotes," replacing the version dated July 4, 2025.

Deloitte's own account is in a letter to the Department of Finance, released under freedom of information. The firm's team used AI tools to "complete and format citations" in drafts of the report, the letter says, and the AI output was reviewed before it went in. That review "corrected some errors" but missed others.

After a newspaper questioned some of the sources, an investigation "confirmed that the use of the generative AI tools had resulted in inaccurate outputs," the firm wrote. A further review "identified further inaccurate outputs," the letter adds. Reading a report through tells you whether its citations look right. Only opening the sources tells you whether they are.

A newspaper printed a reading list of books that don't exist

On May 18, 2025, the Chicago Sun-Times ran a 64-page special section licensed from King Features. Its summer reading list "recommended 15 book titles with reviews, though 10 of the titles were fake while the authors were real," the paper's own review found. In its first statement, the paper said a freelancer working with King Features had used AI to help build the section. It also said the section "was inserted into our paper without review from our editorial team," and that readers weren't told it came from a third party.

The same review found "multiple errors" in all 10 stories in the section that named sources. They included experts whose named employers had no record of them, and a quote from a well-known chef who, according to her assistant, never gave it. The paper removed the section from its e-paper and didn't charge subscribers for that edition. Under its new policy, a new standards team with newsroom editors reviews licensed content.

Work you buy from a freelancer or an agency carries your name once you publish it, so it needs the same checks as work you wrote.

What kinds of things does AI make up?

Between them, the four cases show four kinds of invented detail. A study of AI transcription adds a fifth.

The last row comes from researchers who ran an AI speech-to-text model on thousands of recorded speech clips, averaging about 10 seconds, in April and May 2023. In their study, 1.4% of the clips produced text nobody had said. They also found that "38% of hallucinations include explicit harms," such as perpetuating violence, making up inaccurate associations or implying false authority.

How often does AI hallucinate?

No single rate applies, because it depends on the task and the tool. Law is one field where researchers have measured it, because every answer can be checked against real cases.

Stanford researchers tested AI legal research tools from LexisNexis and Thomson Reuters, which look up real case law before answering. The tools hallucinated "between 17% and 33% of the time," counting an answer as hallucinated if it was "either incorrect or misgrounded." A misgrounded answer cites a source that doesn't say what the answer claims, or doesn't apply. The tools did better than a general-purpose AI chat model. Some of their makers had advertised "hallucination-free" citations, and the researchers wrote that "the providers' claims are overstated" in their paper.

NIST's view is that "confabulations are a natural result of the way generative models are designed." Better tools lower the rate, but none of the tools in the Stanford study came close to zero.

How to catch AI hallucinations before they go out

The checks below come from the failures above. Give each check to a named person, and run it on anything that goes to a customer, a client, a court or the public.

1. Open every source the draft cites. Find the case in a court database, the book on the publisher's site and the study by its title. If you can't find it in a few minutes, cut it. NIST's suggested action is to "review and verify sources and citations in GAI system outputs."
2. Match every quote to its original. A quote from a real person needs a published source or the person's confirmation. AI writes what the person would plausibly say, so an invented quote reads like a real one.
3. Trace every number to its document. Check prices, dates and percentages against the file they came from, and recalculate totals yourself.
4. Check answers about laws and rules against the rule itself. That means the statute, the regulator's page or your own written policy, not the AI's summary of it.
5. Give customer-facing AI your current documents, and let it say it doesn't know. Load the price list, policies and terms it should answer from. When the answer isn't there, it should pass the question to a person instead of filling the gap. For review replies, our Google reviews guide has reply rules to start from.
6. Test it with questions your documents don't answer. Before switching it on, ask about a refund you don't offer or an area you don't serve. The right answer to those is a hand-off to a person.
7. Treat AI transcripts and meeting notes as drafts. Before they go to a client, check names, amounts, deadlines and promises against the recording or your own notes.
8. Log every error you catch. Note the task, what was invented and which check found it. After a month, the log shows which tasks need tighter instructions and which need a person reading every line.

If your team doesn't have written rules yet, item 17 in our AI readiness assessment covers who reviews AI drafts and what they compare them with. For the rest of a one-page policy, see what goes in an AI policy.

Optimitor's AI content work links every fact and number in a draft to its source, and nothing goes on your website until a person on your team approves it.

Frequently asked questions

What is an example of an AI hallucination?

A common example is a citation to a source that doesn't exist, such as the six court decisions New York lawyers filed in Mata v. Avianca in 2023. Other examples are harder to spot: a policy your business doesn't have, a quote its supposed speaker never gave, or a sentence in meeting notes that nobody spoke. They sound plausible, which is how they get through.

How do you check if AI is hallucinating?

Check each claim against an original source the AI didn't write. Open every citation, match quotes to where they were published and trace numbers to the document they came from. Asking the same AI whether it's right is no check. In the Avianca case, the court found the AI told the lawyer that the cases it supplied were real and could be found in legal research databases, according to the sanctions order.

Will AI hallucinations go away?

Not entirely. NIST describes confabulation as a natural result of how generative models are designed. Models do improve: when the transcription researchers re-ran their clips in December 2023, after an update, only 12 of the 187 clips that had produced invented text still did. They wrote that the results suggest the model is improving but "still regularly and reproducibly hallucinates," so keep a person checking anything that leaves the business.

Is a business responsible for what its AI tells customers?

A British Columbia tribunal ruled on a related case in 2024, Moffatt v. Air Canada. The airline's chat assistant told a customer they could claim a bereavement fare after traveling, which the airline's policy didn't allow. The tribunal held Air Canada responsible for "all the information on its website" and ordered it to pay C$812.02. The decision doesn't say whether the assistant used AI, and it doesn't decide US law, so ask a lawyer about your own situation.

Should a small business stop using AI because of hallucinations?

No. The judge in Mata v. Avianca wrote that there is "nothing inherently improper about using a reliable artificial intelligence tool for assistance." In the cases above, the invented detail got past whoever checked the work, or nobody checked it at all. Use AI for first drafts, and make a named person responsible for checking what goes out.