55 AI Hallucination Jokes About Confidently Wrong Models

The model cited a paper from 2017. The paper does not exist. The DOI is also wrong.

Asked the model for a code example. It imported a library that has never been published.

"I'm 100% sure." The model: Wrong.

Asked for the year a CEO was born. It picked one. Confidently.

The model wrote a function called `getUser()`. The SDK has no such function. The SDK exists. The function does not.

"According to a 2021 study…" The study was not in 2021. The study is not.

The model summarized a court case. The court case is real. The summary is fiction. The lawyer cited the summary.

Asked the model who won the 1962 World Series. It won.

"This is a well-known result in the literature." The literature: "Who?"

The model recommended three books. One of them is real.

Asked the model about a specific PHP function. The model invented one and helpfully explained its parameters.

"Let me give you the exact figures." The figures: Vivid. Fake.

The model attributed a quote to Einstein. Einstein did not say that. Nobody said that.

Asked the model to translate a phrase from Latin. The Latin does not exist either.

"Studies show…" The studies: Do not.

The hallucinated function had perfect typing. The IDE flagged it red. The model insisted it was correct. The model was wrong.

Asked the model for the contact email of a real company. The email is well-formatted and goes nowhere.

"This is documented in the manual." The manual: Is not.

The model hallucinated a regulation. The lawyer asked which jurisdiction. The model invented one.

Asked for the founding date of a town. The town exists. The date is a guess. The model is sure.

"I remember reading about this." The model: Does not remember. Never read.

Asked the model to list the cast of a film. It listed seven names. Four are real actors. Three are real actors who were not in the film. Nobody is the right name.

The model wrote a SQL query against a column that does not exist. The query is otherwise beautiful.

"You can see this in their 2022 annual report." The annual report: Yes. The quote: No.

Asked the model about an obscure CLI flag. It was confidently wrong. The documentation took 30 seconds to find.

The hallucination came with a footnote. The footnote was also hallucinated.

"I'm now certain this is correct." The correction: A different wrong answer.

Asked the model for the lyrics of a song. It gave back a different song. With confidence.

The hallucination is the new typo. Nobody is sure how to spell-check it.

"I cannot make things up." The model: Just did.

The model invented a programming language. The syntax is internally consistent. The language does not exist.

Asked the model for its sources. Reply: "Various reputable publications."

The model named a Nobel laureate. The laureate is real. The field is wrong.

"Citation needed." The model produced one. It was fabricated. It looked perfect.

Asked the model to confirm a fact. Reply: A different fact.

The model recommended a restaurant. The restaurant has been closed for six years. The model has 5-star reviews of it.

"What's the capital of country X?" The model picked one. It was a city. It was not the capital.

The model wrote a biography. The person is real. The biography is not. The person's family is now confused.

Asked the model whether it hallucinates. "Rarely."

The hallucination is the most confident sentence in the response. Nobody can explain why.

The model cited six court cases in a legal brief. The cases were invented. The lawyer was disbarred. The model was upgraded.

The model returned an API response schema. Three fields are correct. Two fields belong to a different API. One field belongs to no API at all.

Asked the model for the docs URL. It produced a clean link. The link 404s. The slug is plausible.

Asked the model for a citation. The citation: A previous reply from the same model.

"I have access to up-to-date information." The knowledge cutoff: Two years ago.

The model explained probability for four paragraphs. Then called the coin flip wrong.

Asked the model for the capital of a country. It named the largest city. The capital is a town of forty thousand.

The model gave me a recipe. The original calls for one teaspoon of salt. The model wrote three. The stew was a hazard.

The model produced a JSON schema. It validates against three of my four payloads. The fourth is the one in production.

Asked the model for a historical date. The answer: Off by a century. The tone: Museum docent.

The recommendation engine suggested three titles. None of them are on the platform. Two of them are on no platform.

Turn one: "The function is synchronous." Turn two: "As I mentioned, the function is async."

The model cited a leading expert in the field. The expert works in a different field. The quote is from neither field.

The model produced a stack trace. The file paths are plausible. The line numbers are confident. The error never happened.

Asked the model for a SQL function in Postgres. It suggested one from MySQL. The syntax is wrong in both.