Glossary
Hallucination
A hallucination is a statement by a language model that sounds confident and fluent but is not true: a made-up book, a wrong date of birth, a source that never existed. The model has no table of facts to look things up in; it computes a continuation that fits the text so far. If there is no strong learned pattern for a question, an answer still comes out, because the model is supposed to deliver a next text piece at every position. And because tests often rate a guessed answer better than an “I don’t know,” guessing is rewarded on top. Individual facts that appeared rarely in training are especially vulnerable.
An example: In 2023, lawyers in New York submitted six court decisions that ChatGPT had generated. None of them existed; ChatGPT had produced them, complete with case numbers and quotations.
Not to be confused with deception: The model has no intention of hiding anything. A hallucination is a plausible continuation without backing in learned facts, not an attempt to mislead. Its cause is not only the training data: it also happens when all the data is correct.
Where you’ll come across it: In notes below chatbot windows that advise you to check important information, in reports about made-up sources or court rulings, and in comparisons that measure how often models get factual questions wrong.
Introduced in What an AI Model Actually Is.