What is an AI hallucination?
When a model states something false as if it were fact — a structural feature of next-word prediction, not a bug that’s been fixed.
A hallucination is when an AI system states something false as though it were established fact: a wrong date, a fabricated source, a plausible-sounding number that nobody ever measured. It happens because these systems generate text by predicting what word comes next, not by checking a fact against reality first. Researchers, including the US National Institute of Standards and Technology (NIST), describe this as a structural feature of how the technology works, not a bug that gets patched away with the next release.
Why it happens
A language model produces text by predicting the most statistically likely continuation of what came before it, based on patterns learned from vast amounts of writing. Most of the time that produces accurate answers, because true statements tend to look and sound like the patterns the model has seen. But plausible and true are not the same thing. When a model lacks the actual fact, it does not stop and say so by default; it keeps predicting a fluent-sounding continuation, and a fluent sentence about a made-up court case reads exactly like a fluent sentence about a real one. Stanford HAI’s research on AI reliability has documented this pattern across widely used tools such as ChatGPT and other assistants like Claude, Gemini and Copilot, noting that better training reduces the rate but has not eliminated it in any system studied.
What to do about it
Verify anything that matters before you rely on it, especially names, dates, citations, and numbers, by checking an independent source. Tools that ground their answers in documents or data you supply, rather than relying purely on the model’s trained memory, tend to hallucinate less because there is something concrete to check against. Treat AI output as a confident first draft, not a verified fact, and be most careful exactly when the answer sounds most authoritative.