Short answer
An AI hallucination is an answer from an AI system that sounds confident and is false, or is not supported by anything the system was given. It happens because language models produce plausible text, not looked-up facts, and many are trained and scored to prefer a confident guess over saying they do not know.
A plain example
Ask a general chatbot what your company's refund window is, and it may give a clear number even though it has never seen your policy. The answer is fluent, specific and invented. That is a hallucination, and a person who trusts it can act on it.
How it differs from an old answer
An old answer was true once and nobody updated it. A hallucination was never true. The first needs an owner and a review date. The second needs a bot that answers only from approved sources and says so when it has none.
What reduces it
Limit the bot to approved answers, show the source and date, give it a fixed 'I don't know' line, route unknown questions to a person, and test with questions that have no answer. These reduce hallucinations; they do not remove them.
Related
Want early access and new checklists by email?
Leave your email and we write to you when your spot is ready. Or sign in now and the first 100 founders start right away.
Common questions
No. The system has no intent. It produces plausible text, and sometimes that text is false.
Not completely. Vendors that publish guidance say the techniques reduce them but do not eliminate them, so critical information should still be checked by a person.