Large language models concept
Why AI hallucinates
ReferenceWhy does AI hallucinate? Because a language model generates likely text, and a plausible continuation can still be false. It can produce a confident answer even when it lacks a reliable basis for a detail, or when the question calls for information that is uncertain or highly specific. A fluent answer is not the same as a checked answer, so important claims need to be verified against sources.
Reviewed 2026-10-09
A likely answer is not necessarily a true one
An AI hallucination is an answer that presents false or misleading content as if it were true. The term can also cover text that conflicts with the prompt or contradicts something the system said earlier. A language model learns patterns in text and generates a continuation that fits those patterns. This can produce accurate answers, but pattern fit alone does not guarantee that each detail matches the world. The NIST Generative AI Profile calls these errors “confabulations” and notes that they are also commonly called hallucinations.
How a plausible mistake takes shape
First, the model receives a prompt and uses patterns learned from text to generate a continuation, one token at a time. Next, it selects text that fits the context; the result may sound like a familiar, coherent answer. But generating a likely continuation is not the same as checking each claim against a reliable source. This matters especially for open-ended answers, long responses, and topics that need detailed context or specialist knowledge. A 2025 paper by OpenAI researchers also argues that training and evaluation can reward guessing over saying when the model is unsure.
Why the mistakes can be hard to spot
A made-up detail may sit inside a mostly useful answer, so a reader can miss it. The answer may sound confident, but confidence in the wording is not evidence that the claim is true. Some questions ask for a rare or very specific fact; others are open-ended or need specialist context. Those are situations where a model's learned text patterns may not be enough to produce a correct detail. A response can also include a citation or a chain of reasoning that looks like support but is itself wrong. The NIST profile specifically warns that generated citations and reasoning may be confabulated.
Check the claim, not the confidence
Try this with any answer that matters: pick one specific, checkable claim, then look for a source that directly supports it. Open the cited source yourself and check that it exists, says what the answer says it does, and refers to the same person, date, or event. If there is no citation, search for a primary or otherwise trustworthy source and compare the wording. This is useful because a generated citation can be false or fail to support the answer, and a confident tone can make an unsupported claim seem persuasive. If you cannot confirm the claim, treat it as unverified rather than repeating it as fact.
Questions people ask
What are AI hallucinations, simply explained?
They are false or misleading details that a generative AI presents in its output. The answer may still sound plausible, and the term can also include output that conflicts with the prompt or earlier text.
Why do chatbots make things up?
A chatbot generates likely text from learned patterns; it does not automatically check every claim against a source. A 2025 analysis by OpenAI researchers argues that some training and evaluation practices can also reward guessing instead of admitting uncertainty.
Can a chatbot hallucinate a source or citation?
Yes. NIST notes that generated citations can be confabulated, so open the cited source and check that it exists and supports the specific claim.
How can I tell whether an AI answer is a hallucination?
Choose a specific claim and verify it against a source that directly supports it. A confident tone or a citation shown in the answer is not enough by itself.
Learn more
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
Explains NIST's term confabulation and how generated answers, reasoning, and citations can be false.
- Why language models hallucinate
Gives a research explanation of why plausible errors arise and how evaluation can encourage guessing.
- Why Language Models Hallucinate
Presents the researchers' argument that training and evaluation procedures can reward guessing over acknowledging uncertainty.
What changed
Sources
Part of Large language models: How a model learns and continues patternsWhere fluent output can fail