What is AI hallucination

What is AI hallucination in plain words

Researchers describe generative AI hallucinations as smooth, convincing text that is factually wrong or entirely made up. Chatbots and AI-enhanced search engines both do it, especially on specialized questions.

The ai hallucination meaning is easy to miss because the answer reads well.

A model does not look facts up the way a person checks a source. It predicts likely words, so a plausible sentence can be false.

To define AI hallucination in one line: a confident answer that is false or invented.

AI hallucination examples and cases

• Fake citations

  Real-looking sources that do not exist.

• Wrong policy

  A promise your company did not make.

• Invented figures

  Numbers the source did not give.

• Made-up history

  Events told with a straight face.

Examples are easy to find once you know what to look for.

A model cites a court case that does not exist, quotes a study nobody wrote, or gives a policy detail your company does not have.

Some ai hallucination cases are funny, like a made-up historical event told with a straight face. In business they cost money, because a wrong answer reaches a customer or a contract.

How often AI hallucination happens

Studies on ai hallucination rates vary a lot by task. Stanford's AI Index reports that in one accuracy benchmark, hallucination rates across 26 top models ranged from 22% to 94%.

A 2026 study found advanced models still hallucinate on roughly 15 to 20% of factual citation tasks, and 35 to 55% on niche or recent topics.

Stanford's RegLab found general-purpose models were wrong on 69 to 88 percent of legal queries, as reported by a vendor summary.

Read these as ranges, not one score. The rate depends on the model, the question and how specialized the topic is.

Why it happens and how to reduce it

A model fills gaps with likely text instead of saying it does not know. Weak or missing source data makes this worse.

You can cut the risk in layers. Our retrieval augmented generation page shows how a model answers from your own documents.

Add checks and a person reviewing important outputs, as described in human in the loop.