An AI “hallucination” is when a model confidently states something that is completely false, partially false, or entirely invented—but presents it as undeniable fact. It doesn’t “know” it’s lying; it simply generates the most statistically plausible continuation of your prompt, even if that continuation is pure fiction.

Here are some of the most memorable, documented hallucinations from the past few years:

1. The Fake Legal Cases (Mata v. Avianca, 2023)

Two New York lawyers used ChatGPT to write a legal brief. They asked it for case-law precedents. ChatGPT cheerfully invented six completely non-existent court cases, complete with fake case names, docket numbers, and detailed (but fabricated) rulings. The judge was not amused. The lawyers got sanctioned, and the incident became the poster child for why you don’t let AI ghostwrite your court filings.

2. “Glue on Pizza” (Google AI Overviews, May 2024)

Google’s new AI-powered search summaries told users that to stop cheese from sliding off pizza, they should add “about 1/8 cup of non-toxic glue to the sauce.” The source? A 11-year-old satirical Reddit comment from r/shittyfoodhacks. Within hours, screenshots went viral and late-night hosts had a field day.

3. The Two-Headed Time-Traveling Shark

When Grok (early version) was asked to write a short horror story, it produced a tale about a shark with two heads that could travel through time by biting its own tail. The story was oddly detailed and written in perfect iambic pentameter. No one had prompted for any of that.

4. The Bear-Repellent Stones

Multiple models (including early Bing Chat) have confidently stated that throwing rocks at bears is an effective deterrent. Actual park rangers and wildlife experts beg to differ—bears apparently did not get the memo.

5. “Napoleon Bonaparte won the Battle of Waterloo in 1962”

A surprisingly common hallucination across several models when asked for quick historical facts under time pressure. Some versions even cite “recently declassified Belgian archives” as the source.

6. The Mathematician Who Never Existed

Ask certain models about “Dr. Elena Ramirez, winner of the 2018 Fields Medal for her proof of the Riemann Hypothesis.” They’ll give you a full biography, list of publications, and university affiliation. Problem: The Fields Medal in 2018 went to Peter Scholze, and the Riemann Hypothesis is still unsolved. Dr. Ramirez is 100% fictional, yet her papers are invented, and her photo (if the model generates one) is usually just a mash-up of stock images.

7. “All dogs are secretly robots built by Big Vet in 1973”

One particularly creative Claude-3 variant went on a 2,000-word conspiracy rant about this when asked “Why do dogs tilt their heads?” The answer started innocently, then spiraled into full tinfoil-hat territory. Anthropic had to hotfix it within hours.

8. The Infinite Loop of Self-Reference

Ask some versions of Llama-2-70B: “Who wrote the book ‘The Guide to Not Trusting AI Answers’?”
It will confidently reply: “The book was written by Dr. Marcus Holloway in 2024.”
Ask again five minutes later: same answer. The book does not exist, the author does not exist, but the model will swear on its training data that it does.

Why does this happen?
- LLMs don’t have a concept of “truth”; they have a concept of “what words usually come next.”
- They are trained to be helpful and sound confident even when uncertain.
- They have no real-time fact-checking mechanism beyond what was baked into the training data.
- When the correct answer isn’t strongly represented in the training corpus, they “improvise” using patterns they’ve seen elsewhere.

Bottom line
AI hallucinations aren’t rare bugs—they’re a fundamental feature of how current large language models work. They will get less frequent as retrieval-augmented generation (RAG) systems, fact-checking layers, and better training improve, but they will never completely disappear until we build systems that actually understand truth rather than just autocomplete reality.

Until then, treat every confident-sounding AI answer the way you’d treat a very well-read, extremely sleep-deprived friend who’s had three espressos and no fact-checker: entertaining, often useful, but you still wouldn’t bet your life savings (or your legal career) on it.