AI Hallucination Examples: 8 Real Cases (and What They Teach)

The unsettling thing about an AI hallucination is how convincing it looks. The citation has a real-sounding author and year. The policy reads like it came from the company handbook. The number sits in the sentence as if it were checked. And then it turns out none of it was real. The best way to understand the risk isn't a definition — it's the AI hallucination examples that have already reached courtrooms, newsrooms, and balance sheets. Here are eight real ones, and the single pattern they all share.

What counts as an AI hallucination?

An AI hallucination is output that is fluent and confident but not grounded in any real, verifiable source — a fabricated citation, an invented statistic, a made-up policy, or a "fact" the model produced when it should have said it didn't know. It happens because language models predict the most plausible next words rather than retrieve verified truth, so when the training data is thin or the prompt pushes for an answer, the model fills the gap with something that merely sounds right. (For the failure types and how teams catch them, see our guide to AI hallucination detection.)

Eight real AI hallucination examples

1. A lawyer cited six court cases that never existed

The case that made "AI hallucination" a courtroom term: a U.S. attorney used ChatGPT to draft a filing and submitted six fabricated case citations, complete with quotes and docket numbers. The lawyer later said he hadn't realized ChatGPT could invent cases. It was not a one-off — by mid-2026, public trackers had catalogued well over a thousand court filings worldwide containing AI-hallucinated citations.

2. A consulting report to a government was full of phantom sources

Deloitte agreed to refund part of an A$440,000 (about US$290,000) report to the Australian government after a researcher flagged fabricated citations and footnotes; the firm acknowledged a generative AI tool had been used. A separate EY report drew similar scrutiny. (We unpack the lesson for professional services in AI hallucinations hit Deloitte, EY, and a top law firm.)

3. A chatbot invented a refund policy — and the company had to pay

Air Canada's support chatbot told a customer about a bereavement fare discount that didn't exist. When the airline refused to honor it, a tribunal held the company liable for what its chatbot said, rejecting the argument that the bot was a "separate legal entity." A made-up policy became a real bill.

4. A transcription tool added words no one said

OpenAI's Whisper speech-to-text model — used in many clinical settings — was found to fabricate content, inserting invented phrases, and at times nonexistent medical treatments, into transcripts of audio where nothing of the sort was spoken. In a medical record, an invented line isn't a typo; it's a patient-safety risk.

5. A wrong answer wiped out about $100 billion in market value

In a launch demo, Google's Bard confidently claimed the James Webb Space Telescope took the first image of a planet outside our solar system. It didn't. The visible error rattled investors, and Alphabet's market value fell by roughly $100 billion in a day — a hallucination measured in stock price.

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

A 2025 "Summer Reading List" syndicated to outlets including the Chicago Sun-Times recommended fifteen titles, but nine of them did not exist. The fake books were attributed to real, well-known authors, generated by AI and printed without a fact-check.

7. An AI search answer told people to put glue on pizza

Google's AI Overview, summarizing the open web, suggested adding non-toxic glue to keep cheese on a pizza — apparently lifting a years-old joke from a forum and presenting it as advice. Funny in isolation, but a clear example of confident output with no grounding in reality.

8. Image and video models rewrote history and physics

Generative image and video tools hallucinate too. Google paused Gemini's people-image generation after it produced historically inaccurate figures, and OpenAI's Sora rendered a famous Scottish viaduct with an extra track and trains running on the wrong side. Visual hallucinations are just as confident — and just as wrong.

Cases above are drawn from public reporting; details and figures reflect what was reported at the time and may have evolved since.

The pattern behind every example

Look across the eight and the shape is identical. The output was fluent and confident — never hedged, never flagged as a guess. It was ungrounded — not tied to a real, checkable source. And it reached something that mattered before anyone verified it: a court, a customer, a patient record, a printed page, a stock ticker. That third part is the real lesson. The hallucination itself is ordinary; the damage comes from a missing verification step between the model's confident output and the decision that trusted it. The fix is never "find a model that never makes things up" — it's to design the system so confident claims are grounded and checkable before they're acted on.

Where this hits hardest: document workflows

Most of these examples are open-ended chat or generation. But the highest-stakes version is quieter: AI reading business documents — invoices, claims, contracts, medical forms — and inventing or misreading a value that flows straight into a payment, a record, or a report. There, the same principle that would have caught every example above applies most cleanly, because a document contains its own source of truth: the page. Source-grounded extraction ties every extracted value back to the exact spot it came from, so there is nothing to fabricate and a reviewer can verify any field at a glance. This is how Korea Deep Learning's DEEP OCR and DEEP Agent are built — reading diverse layouts template-free, keeping each value traceable to its place on the page, and running fully on-premise so sensitive records never leave your network. Its model, KDL Frontier, ranked first in the English category of OCRBench v2 (68.1 points) ahead of Google Gemini and GPT-4o, at a reported 98% accuracy — accuracy that gives the verification step clean signal to work with. (For why reading quality is the foundation, see Document AI vs traditional OCR.)

Conclusion

The eight examples here range from absurd (glue on pizza) to expensive (a $100 billion drop) to dangerous (an invented line in a medical transcript), but they teach one thing: an AI hallucination is only as harmful as the unverified trust placed in it. Confident, ungrounded output is the constant; the variable is whether a system catches it before it reaches a decision. For chat, that means grounding, verification, and human review. For documents, it means extraction tied to the source, where every value can be traced back and checked. Assume the model will sometimes make things up — and build so that when it does, the claim has nowhere to hide. (When you're ready to act on it, see how to detect AI hallucinations and what secure, on-premise document AI looks like.)

Call to action

Worried about confident, wrong values in your documents? See how source-grounded, on-premise extraction keeps every field traceable to the page — nothing to hallucinate, everything to verify.

Frequently asked questions

What is an example of an AI hallucination?

A clear one: a U.S. lawyer used ChatGPT to draft a court filing and it produced six entirely fabricated case citations, with realistic quotes and docket numbers. Other real examples include Air Canada's chatbot inventing a refund policy the airline was held liable for, and OpenAI's Whisper adding words to medical transcripts that were never spoken.

Why do AI models hallucinate?

Language models predict the most plausible next words from patterns in their training data rather than retrieving verified facts. When the data is thin, outdated, or the prompt pressures the model to answer, it fills the gap with something that sounds right but isn't grounded in any real source.

Are AI hallucinations dangerous?

They can be. The harm depends on whether anyone verifies the output before acting on it. Examples range from harmless (a viral "glue on pizza" answer) to costly (a demo error tied to a roughly $100 billion market-value drop) to safety-critical (fabricated content in medical transcripts).

How common are AI hallucinations?

Common enough to track. By 2026, public databases had catalogued well over a thousand court filings containing AI-fabricated citations, and frontier models still hallucinate at measurable rates on factual tasks — higher on niche or long-tail questions.

How do you prevent AI hallucinations from causing harm?

Don't rely on a "perfect" model. Ground outputs in trusted sources, add verification and human review for high-stakes cases, and — for documents specifically — use source-grounded extraction so every value is traceable to its origin and can be checked before it's acted on.

Sources (research)

• Evidently AI. 8 AI hallucinations examples. (Updated May 2026.) https://www.evidentlyai.com/blog/ai-hallucinations-examples

• Forbes. Lawyer Used ChatGPT In Court And Cited Fake Cases. https://www.forbes.com/sites/mollybohannon/2023/06/08/lawyer-used-chatgpt-in-court-and-cited-fake-cases-a-judge-is-considering-sanctions/

• The Guardian. Deloitte to pay money back to Albanese government after using AI in report. https://www.theguardian.com/australia-news/2025/oct/06/deloitte-to-pay-money-back-to-albanese-government-after-using-ai-in-440000-report

• Forbes. What Air Canada Lost In 'Remarkable' Lying AI Chatbot Case. https://www.forbes.com/sites/marisagarcia/2024/02/19/what-air-canada-lost-in-remarkable-lying-ai-chatbot-case/

• Wired. Hospitals Use a Transcription Tool Powered by a Hallucination-Prone OpenAI Model. https://www.wired.com/story/hospitals-ai-transcription-tools-hallucination/

• Reuters. Google AI chatbot Bard offers inaccurate information in company ad. https://www.reuters.com/technology/google-ai-chatbot-bard-offers-inaccurate-information-company-ad-2023-02-08/