Behavior & Risk
Hallucination
When an AI model states something false or fabricated with the same confidence as a correct answer.
Language models generate text by predicting what's statistically plausible, not by checking a fact database — so a fluent, confident-sounding sentence and a fabricated one can look identical from the outside. Hallucinations range from invented citations to wrong dates to entirely made-up events or people.
Techniques like RAG and tool use — letting a model call a search engine or calculator instead of guessing — reduce hallucination rates, but no released model has eliminated it, which is why fact-checking AI output against primary sources still matters.