In one line: language models predict the most plausible next words, not the true ones. When they don't know, they don't stop — they generate something that sounds right. That's a hallucination.
What it is
A hallucination is confident, fluent output that is factually false or entirely invented — a fabricated citation, a made-up statistic, a non-existent product feature, a court case that never happened. The danger is that it reads exactly like a correct answer.
Why it happens
A language model is a next-token predictor. It produces the most statistically likely continuation of the text so far. When training data is rich, that continuation is usually true. When data is thin, ambiguous or absent, the model still produces its best guess — and a confident guess is indistinguishable, in its output, from knowledge.
Why it can't be fully fixed
A 2025 result showed hallucination is mathematically inevitable for models built this way. It is a structural property, not a bug. Any claim of "zero hallucinations" is marketing. The realistic goal is to minimise and detect, covered in how to reduce hallucinations.
The confidence paradox
Counter-intuitively, models built to reason harder often hallucinate more on factual questions. OpenAI's o3 hit ~33% on PersonQA — double its predecessor o1. More elaborate reasoning produces more elaborate, more confident fabrication. Smarter is not the same as more honest. See which AI hallucinates most.
The "confident liar" problem
A model has no internal sense of certainty. It can present a fabrication with exactly the same confidence as a fact. A model that says "I'm not sure" is far safer than one that bluffs — which is why we weight confident-wrong answers more harshly in the Truth Score.
It depends heavily on the task
| Task | Typical rate |
|---|---|
| Summarising supplied text | <1.5% |
| General factual Q&A | 15–33% |
| Medical summaries | 43–64% |
| Legal queries | 58–88% |
More detail by sector in hallucination by industry.
Want to design around it? Read how to reduce hallucinations and what AI can't do.