An AI hallucination is an answer that sounds confident and convincing but is false or made up: a fact, a quote, a law, a link or a book that does not exist. It happens because a language model builds the most likely text without checking if it is true. It is not a rare glitch: it is part of how it works.
A question from the app. Pick an answer.
If it predicts what is likely, why does it make things up?
A false answer does not imply an intent to deceive. It can generate convincing text without support.
A likely sentence can be false. Predicting tokens does not guarantee fact checking.
A hallucination does not mean the system is broken: generating text does not guarantee accuracy.
You ask for a reading list for a school paper and get five books with author, year and publisher. Two do not exist. Or you ask for the source of a fact and get a link that looks real, but the page is not there. The tone does not change: what is made up is written with the same confidence as what is true.
The name is a metaphor: the AI does not see things or get confused like a person. It fills in something that fits the shape of the answer, even when it does not fit reality.
A language model writes by picking one token after another based on what usually comes next. If the fact you want was not in what it learned, a most likely next piece still exists, and it writes it. A citation with author and year has a very recognizable shape, so the model can build a perfect one that is empty inside.
Some situations make it worse: you ask about something newer than what the model learned, the detail is very specific (an article number, an exact date) or the instructions push it to always answer, like an assistant told "never say you do not know".
Many models saw far more text about global topics than about your town. Ask about a musician well known in your region, the bus routes in your neighborhood or the books of an author from your country and the chance of getting something made up goes up: there are few examples, and the model fills in with what seems typical.
Give it the source in the prompt: paste the text or document and ask it to answer only from that (automated, that is RAG). Give it permission to say "I do not know". Ask it to quote the exact sentence each fact comes from, so you can look it up. And open the links and check numbers, names and quotes before you use them. None of this guarantees it, but it lowers the risk a lot.
"I ask if it is sure and it says yes." That proves nothing: it can confirm what it made up just as calmly. "If it searches the web, it stops making things up." It helps, but it can misread a page or mix up two. "It only happens with hard questions." It also happens with simple questions about facts it does not know.