What Is an AI Hallucination? (Why Models Make Things Up)

Updated July 23, 2026 · Data as of 2026-07-09

A hallucination is output that is fluent, confident, and wrong: an invented citation, a function that does not exist, a plausible statistic with no source. The model is not lying, it is producing likely text, and likely is not the same as true.

The cause sits in the mechanism. A language model predicts what should come next given everything so far. When the material it needs is missing, unclear, or absent from training, the most probable continuation is still a well-formed sentence, so it writes one. Nothing in the process distinguishes between recalling a fact and constructing something shaped like one, which is why the tone of a fabricated answer is identical to the tone of a correct one.

There are reliable danger zones. Specific citations, page numbers, case law, and academic references are hallucinated often, because their form is highly predictable and their content is not. So are library functions and API parameters that sound exactly like what should exist. Recent events past the training cutoff are another, as are questions built on a false premise, where the model tends to accept your framing and elaborate on it rather than push back.

The mitigations that work are structural rather than verbal. Grounding the answer in supplied documents and demanding citations lets you check the source rather than trust the sentence. Asking a model to run code, or to consult a search tool, replaces recall with verification. And a plain instruction that saying you do not know is an acceptable answer measurably reduces confident invention. What does not work is asking the model whether it is sure: the follow-up is generated by the same process as the original.

Why it matters when you choose a tool

The right question is not which tool hallucinates least, since all of them do it, but which tool makes errors cheap to catch. Prefer products that cite sources you can open, that search rather than recall for anything current, and that execute code rather than describing what it would do. Then match your review effort to the stakes. Fabrication is a minor annoyance in a first draft and a serious problem in anything filed, published, or shipped, and the tool cannot tell which one you are doing.

Frequently asked questions

Which tasks hallucinate most?

Specific citations and references, exact figures, library functions and API details, and anything after the training cutoff. Questions that assume something false are another, because the model usually plays along with the premise.

Does asking the model to be accurate help?

A little, and the more useful version is explicitly permitting uncertainty, telling it that not knowing is an acceptable answer. Asking whether it is sure after the fact helps very little, because the same process generates that answer too.

Will better models fix this?

They reduce the rate, sometimes sharply, and they do not remove the failure mode. Verification habits are still the safeguard, and the remaining errors are harder to spot precisely because the surrounding output is better.

Back to the glossary · Take the quiz · How we score tools