AI hallucination
Last updated: July 26, 2026
An AI hallucination is output from a language model that is stated confidently and fluently but is false or fabricated — a wrong fact, a made-up citation, or an invented detail. The model is not lying; it predicts plausible text and has no built-in check that what it produced is true.
What is an AI hallucination?
An AI hallucination is output from a language model that reads as confident and fluent but is false or fabricated. It looks exactly like a correct answer, which is what makes it dangerous.
The word covers several failure shapes: a stated fact that is wrong, a citation to a paper or case that does not exist or does not say what is claimed, an invented statistic, or a plausible detail with no basis. The common thread is that the model produced it with the same tone it uses for things it gets right.
A model does not know it has hallucinated. It generates the most probable next words given your prompt; when its training data is thin, ambiguous, or absent for your question, the most probable text can still be false. Nothing inside the model separately checks the claim against reality — which is exactly the gap a verifier fills.
What causes hallucination?
Hallucinations come from how language models work, not from a bug you can patch. The main causes are:
- Prediction, not lookup. A model predicts likely text. It has no database it consults and cites; facts are blended into weights.
- Gaps in training data. For rare, recent, or niche topics the model has little signal, so it fills the gap with something plausible — closely related to confabulation.
- A knowledge cutoff. Anything after the training date is answered from stale memory, confidently.
- Pressure to answer. Models are tuned to be helpful and rarely say 'I don't know', so they produce an answer even when the honest response is uncertainty.
Adding retrieval (RAG) reduces hallucination by giving the model real documents, but does not eliminate it — the model can still misread or overstate what a source says.
A worked example
Ask a model for support in a legal brief and it may return: "In Smith v. Jones, 512 U.S. 340 (1994), the court held that..." — a citation that looks real, in the right format, with a plausible year. The case may not exist, or may exist but say the opposite.
This is not hypothetical. A 2024 Stanford RegLab study, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (Magesh et al., later published in the Journal of Empirical Legal Studies in 2025), found that even purpose-built legal AI research tools from LexisNexis and Thomson Reuters hallucinated in a meaningful share of queries — the study reports each tool hallucinating between roughly 17% and 33% of the time, while general-purpose chatbots like GPT-4 fared worse. See the RegLab study.
A verifier catches this by treating the citation as a separate claim and grounding it against an independent source. If no such case is found, the verdict is contradicted or unverifiable — never a silent pass.
Frequently asked questions
Do all AI models hallucinate?
Yes. Every large language model can hallucinate because they all predict plausible text rather than looking facts up. Rates vary by model, task, and how obscure the topic is, but no current model is hallucination-free — which is why independent verification exists. See per-model pages like ChatGPT and Claude.
Can you stop AI hallucinations completely?
No technique eliminates them today. Retrieval (RAG), better prompting, and larger models reduce the rate but not to zero. The practical answer is to verify: check each claim against an independent source and flag what cannot be confirmed, rather than trusting the output. See how it works.
Is a hallucination the same as a lie?
No. A lie implies intent to deceive. A model has no intent and no awareness that its output is false — it is producing statistically likely text. That is why many researchers prefer confabulation, a term that captures the absence of deception.
Verify what your AI just told you.
Paste any AI answer and Mephistopheles checks each claim against independent sources — no sign-up to try.