What is an AI hallucination?

AI models generate text by predicting what is likely to come next, not by checking facts against a trusted record. The result is often useful, but sometimes the model produces something that sounds right and is simply wrong: a court case that does not exist, a product feature that was never built, a quote nobody said, or a web link that leads nowhere. This is called a hallucination.

For example, someone asks an assistant for sources to support a market report, and it returns a neat list of articles with titles, authors, and dates. Some of them turn out not to exist. Nothing in the answer signals which ones are real, because the model writes invented and genuine details in the same confident tone.

Hallucinations are more likely when a question asks for specific details the model was not given, such as exact figures, citations, recent events, or niche facts, and when a prompt pushes for an answer rather than allowing the model to say it does not know. Supplying the source material, asking the tool to quote from it, and connecting it to search or company documents all reduce the risk.

Hallucinations can be reduced but not ruled out, so checking remains essential. Figures, names, dates, legal and medical statements, and anything that will be published or sent to a customer should be verified against the original source. Tools that show where each claim came from make this checking faster, but the links themselves still need opening.

An example

An assistant drafts a policy summary that cites a regulation section that does not exist, which a reviewer catches by checking the original document.