In one sentence

A hallucination is when an AI produces information that sounds confident and plausible but is not true.

The simple model

A model generating text is producing a likely-sounding answer, not looking up a fact. Most of the time the likely-sounding answer is also correct. Sometimes it is not, and the model has no built-in sense that it has crossed that line. The result is a fluent, confident statement that is simply wrong: an invented policy detail, a made-up reference, a plausible but incorrect number.

The danger is precisely that it does not look like an error. It looks like every other answer.

Why you care

Hallucination is the central risk of generation and of any AI answering questions. A person reading a confident answer tends to trust it. If the workflow acts on it, or a customer receives it, the invented information becomes a real problem.

This is why fluency must never be mistaken for accuracy, and why answers that matter need grounding and checking.

How to reduce it

You cannot fully eliminate hallucination, but you can make it much less likely and much easier to catch:

  • Ground answers in real sources. Have the AI answer from retrieved, current material rather than from memory. See retrieve then answer.
  • Require citations. An answer that cites its source can be checked in seconds. See cite the source.
  • Instruct it to say "I do not know" when the sources do not contain the answer, rather than inventing one.
  • Keep a person in the loop for consequential answers.
  • Test for it: include cases in your test set where the honest answer is "not found", and check the AI does not fabricate.

A small example

An assistant confidently told a customer about a "30-day price-match guarantee" that did not exist. It had invented a plausible policy. After grounding it in the actual policy documents, requiring citations, and instructing it to say when it could not find something, the invented answers stopped, and any that slipped through were easy to spot because they lacked a real source.

What this does not solve

Grounding and citation greatly reduce hallucination; they do not guarantee truth. A model can still misread a real source, and a wrong source produces a wrong answer even when cited. Treat every consequential answer as checkable, and check it.