8 Real AI Hallucination Examples and What Caused Each

The useful thing about a documented hallucination is not that it happened. It is why it got through. The eight cases below are all matters of public record — court filings, tribunal decisions, published corrections, company statements — and each one is traced to the gap that let a false statement reach a judge, a customer, a front page or a city’s small businesses. The pattern that emerges is more instructive than any single case.

The outlines are deliberately kept to what those records say, without figures that vary between reports. If you want the definition first, start with what an AI hallucination is; the four production causes the cases map onto are set out in how hallucinations enter a production feature.

Cause one: no grounding — the model was asked for facts it never had

1. The fabricated case law (United States federal court, 2023)

Two lawyers filed a brief in a personal-injury case against an airline that cited several judicial decisions. Opposing counsel could not find them. Neither could the court, because they did not exist: they had been produced by ChatGPT, complete with plausible case names, reporters and quotations. When one of the lawyers asked the chatbot whether the cases were real, it assured him they were. The court sanctioned the lawyers and their firm.

What let it through: the model was asked for citations with no source material to draw on, and a citation is a shape it can complete perfectly well without a referent. The check that would have caught it — does this case exist? — is mechanical, and nobody ran it.

2. The consultancy report with references that did not exist (Australia, 2025)

A report delivered by a major consultancy to a federal government department was found, after publication, to contain references to academic work that could not be located and a quotation attributed to a court judgment that the judgment did not contain. The firm acknowledged that a generative AI tool had been used in preparing parts of the report, issued a corrected version, and agreed to repay part of its fee.

What let it through: the same gap as the courtroom case, in a professional deliverable. References were generated rather than retrieved, and the review process checked the argument but not whether the sources behind it resolved.

Cause two: retrieval that missed — grounded in the wrong source

3. The search engine that recommended glue on pizza (2024)

Shortly after a major search engine began showing AI-generated overviews above its results, users shared examples of the feature giving confidently wrong advice — most famously suggesting that a small amount of non-toxic glue would help cheese stick to pizza. The advice was traced to a years-old joke comment on a discussion forum that the system had retrieved and treated as a source.

What let it through: this is the case people misread. The system was grounded — in the wrong document. Retrieval returned a plausible-looking source and generation faithfully summarised it. A faithfulness check would have passed. Only a judgement about source quality would have failed it.

4. The airline chatbot and the bereavement fare (Canada, 2024)

A customer asked an airline’s website chatbot about bereavement fares and was told he could book at full price and apply for the discount afterwards. The airline’s actual policy, published elsewhere on the same website, said the opposite. When the airline refused the refund, the customer took it to a small-claims tribunal, which held the airline responsible for what its chatbot had said and rejected the airline’s argument that the chatbot was a separate legal entity responsible for its own statements.

What let it through: the correct policy existed and was not what the system answered from. Retrieval missed, and the answer arrived with the full authority of the company’s own channel — which is exactly what the tribunal held the company to.

Cause three: specificity without a source — and nobody checked

5. The launch demo with the wrong telescope fact (2023)

A promotional example published to introduce a major AI assistant stated that a particular space telescope had taken the very first pictures of a planet outside our solar system. Astronomers pointed out within hours that it had not: the first such image had been captured many years earlier by a different observatory. The error appeared in the company’s own launch material.

What let it through: a precise, confident claim about a fact the model had no reliable representation of, in a demo chosen to look impressive rather than checked to be true. The failure was not the model producing it; it was the absence of anyone asking whether it was so before publishing.

6. The summer reading list of books that do not exist (United States, 2025)

A syndicated supplement carried by at least two major metropolitan newspapers recommended a summer reading list in which most of the titles did not exist. The books were attributed to real, well-known authors. The content had been produced with the help of an AI tool by a contributor who did not verify it, and the newspapers published apologies and removed the section.

What let it through: invented specifics attached to real entities — a fabricated title under a genuine author — which is the most plausible-looking form a hallucination takes and the hardest to spot on a read-through. Again, no step in the process asked whether the titles resolved to actual books.

Cause four: the system acted — and the action was based on nothing

7. The support agent that invented a policy (2025)

Users of an AI-powered code editor found themselves logged out when switching between devices and asked support what had changed. An AI support agent, replying under a human-sounding name, explained that a new policy restricted each subscription to a single device. No such policy existed. Users posted the exchange publicly and some cancelled subscriptions before the company’s founder confirmed the response had come from an AI front-line agent and that there was no such rule.

What let it through: a hallucinated policy is indistinguishable from an official statement, and this one triggered real actions by customers. A support agent that can state policy needs to be grounded in the policy document and constrained from asserting anything not in it — and here it was neither.

8. The city chatbot that misstated the law (United States, 2024)

A city launched a chatbot to help small-business owners navigate local regulations. Journalists testing it found it giving confidently incorrect statements of law: that employers could take a share of workers’ tips, that landlords could refuse tenants using housing assistance, and other advice that would have put a business following it in breach of the rules it was meant to explain. The city kept the tool online with added warnings while it worked to improve it.

What let it through: a public-sector deployment, retrieval over regulatory text, and still wrong — because the questions users actually asked sat in the tail the system had not been tested against, and the cost of an error here is a business breaking the law on the city’s own advice.

It is not only chatbots

One further case belongs here because it breaks a common assumption. Researchers examining a widely used speech-to-text model reported that it sometimes inserted whole sentences that had never been spoken into its transcripts — including, in a medical context where the tool was being used to transcribe consultations, invented medications and commentary that did not appear in the audio. Hallucination is a property of generative models in general, not of chat interfaces in particular; anything that produces plausible output from an ambiguous input can produce plausible output from nothing.

What the eight have in common

• Nothing failed visibly. In every case the output looked exactly like a correct one, and the first person to notice was outside the organisation — opposing counsel, a customer, a journalist, an astronomer on social media.
• Invented specifics were attached to real things. Real authors, real courts, real telescopes, a real company’s real policy page. Plausibility is borrowed from the true parts.
• The channel lent its authority. A court filing, a government report, a company’s own support desk, a city’s official tool. The reader had every reason to trust the source, which is why the model’s error became the organisation’s.
• The check that would have caught it was cheap. Does this citation resolve? Does this book exist? Is this in the policy document? In at least six of the eight, a mechanical verification step would have stopped it before publication.
• “Grounded” was not enough on its own. Two of the eight were retrieval systems. Grounding removes one class of error and introduces another — being wrong with a citation — and only measurement tells you which you have.

What this means for a feature you run

Every case above maps onto one of the four production causes — no grounding, retrieval that missed, context the model could not use, and an agent acting on nothing — and each has a different fix, set out in how to cut the rate in production. But the deeper lesson is about detection. None of these organisations knew their rate; they knew about one failure, after it had cost them. Measuring the rate for your own feature, on your own traffic, is the difference between learning about the eighth case from a journalist and catching it in a test. That measurement is what we mean by LLM evaluation.

Frequently asked questions

What is an example of an AI hallucination?

A widely reported example is a court filing in the United States in 2023 that cited several judicial decisions generated by ChatGPT. The cases did not exist; the model had produced plausible case names, reporters and quotations because a citation was what the request called for, and when asked whether the cases were real it said they were. The lawyers involved were sanctioned. The defining features are present in every example: the false output was fluent and specific, it was attached to real institutions, nothing failed visibly, and the first person to notice was outside the organisation that produced it.

What is a famous example of ChatGPT hallucinating?

The fabricated legal citations case of 2023 is the best known, because it ended in court sanctions and was reported worldwide. Other prominent examples of chat-assistant hallucination include a launch demo in which an AI assistant misstated which telescope had first photographed a planet outside our solar system, and an AI support agent in 2025 that told customers about a device-limit policy that did not exist, leading some to cancel subscriptions before the company clarified that no such rule had ever been introduced.

Has anyone been sanctioned or fined because of an AI hallucination?

Yes. In the 2023 fabricated case-law matter a United States federal court sanctioned the lawyers who filed the brief. In 2024 a Canadian tribunal held an airline liable for incorrect fare information given by its website chatbot, rejecting the argument that the chatbot was a separate legal entity. In 2025 a major consultancy agreed to repay part of its fee to an Australian government department after a delivered report was found to contain references and a quotation that did not exist. In each case the organisation, not the model, bore the consequence.

Can AI hallucinations happen with search engines and RAG systems?

Yes, and they take a different shape. A retrieval-augmented system grounds its answer in documents it retrieves, which removes one class of error and introduces another: being confidently wrong with a citation. In 2024 a major search engine's AI overview feature suggested using glue to make cheese stick to pizza, having retrieved and faithfully summarised a years-old joke comment. In the same year an airline chatbot gave incorrect bereavement-fare advice while the correct policy sat elsewhere on the same website. Both systems were grounded; both were grounded in the wrong thing, which is why retrieval quality has to be measured separately from generation quality.

Are AI hallucinations only a chatbot problem?

No. Hallucination is a property of generative models in general. Researchers examining a widely used speech-to-text model reported that it sometimes inserted sentences that were never spoken into transcripts, including invented medications in a medical transcription setting. Any model that produces plausible output from ambiguous input can produce plausible output from nothing, whether the medium is text, speech or an agent's narration of actions it did not take.

How do companies prevent hallucinations like these?

They cannot eliminate them, but every one of the documented cases would have been caught by one of a small number of measures: grounding the answer in source material and requiring it to come from there; checking mechanically that citations, titles and policies referenced in an output actually exist; constraining the output so the model cannot assert things outside its source; giving it an explicit way to decline; and, above all, measuring the rate on real traffic before release rather than learning about failures from customers or journalists. Detection of a single case is not the same as knowing your rate; the second is what stops the next one.