What is an AI hallucination?

Output that is fluent, confident and wrong — and the design choices that reduce it.

What is an AI hallucination?

A hallucination is model output that is stated confidently but is factually wrong or entirely invented — a citation that does not exist, a policy that was never written, a figure with no source. It happens because language models generate plausible text rather than retrieving verified facts.

Why it happens

The model has no mechanism for distinguishing 'I know this' from 'this is what such an answer usually looks like'. Both produce equally fluent text, and nothing in the output signals which one occurred.

This is why hallucination is not a bug to be patched but a property to be designed around.

How to design against it

Ground answers in retrieved documents and require a citation for each claim.
Give the model an explicit refusal path — 'not covered in the material I have'.
Build an evaluation set of real questions and measure before launch.
Keep a human review gate on anything consequential or customer-facing.
Log everything, so a wrong answer can be found and traced rather than discovered by a customer.

Can it be eliminated?

Not entirely with current technology. It can be reduced substantially, and — more importantly — made visible. A system that refuses when unsure is far more useful than one that is confidently wrong occasionally and gives no warning.