Imagine asking a very confident person about a book they have never read.

An honest answer would be, “I don’t know.” But someone else might calmly describe its chapters, themes, and author—even though none of those details are true.

That is close to what we call an AI hallucination: an answer that sounds natural and convincing but is false or unsupported by reliable evidence.

What is an AI hallucination?

An AI hallucination occurs when a model generates information that is incorrect or unsupported by the available evidence, then presents it in fluent, plausible language.

It might invent a book title, attribute a quotation to the wrong person, provide a statistic with no source, or create an academic reference with a realistic-looking title, author, and publication date.

Not every hallucination is a complete fabrication. A model may combine two real facts incorrectly, change a date, misidentify a source, or add a detail that the source never contained. The problem is therefore not how polished the sentence looks, but whether its claims are grounded in evidence.

Why do AI models hallucinate?

Language models learn to predict the next text units from patterns in their training data. This makes them very good at producing coherent language, but it does not attach a guaranteed truth label to every sentence they generate.

During pretraining, a model sees enormous amounts of text without always receiving explicit labels that mark every claim as true or false. When a fact is rare, ambiguous, unavailable, or difficult to distinguish from similar information, the model may produce a plausible continuation instead of expressing uncertainty.

OpenAI research also argues that some evaluation methods can reinforce the problem. If a test rewards correct answers but does not sufficiently penalize confident guesses, guessing may score better than abstaining—even when it increases the risk of a false answer.

When are hallucinations more likely?

No simple rule predicts every hallucination, but the risk tends to rise in several situations.

Rare or local information: A model has fewer reliable patterns to use when you ask about an obscure person, a small local event, or a fact that appears rarely in its training material.

Exact citations and numbers: Models can imitate the structure of references and statistics. They may therefore generate a realistic-looking source, DOI, quotation, or percentage that does not exist.

Questions with a false premise: If you ask why someone made a statement they never made, a model may attempt to explain the statement instead of first challenging the assumption.

Recent or changing information: A model may not have current information, and a search tool may be unavailable or unused. Access to the web does not mean that every answer has been checked against current sources.

Missing or contradictory context: When instructions are vague or documents are incomplete, the model may fill the gaps with unsupported assumptions.

Can search and retrieval prevent hallucinations?

Search, retrieval-augmented generation, and access to selected documents can reduce hallucinations, but they cannot guarantee that every output is correct.

A model may misread a source, connect unrelated passages, cite a page that does not support its claim, or overlook a more authoritative source. Even advanced models can still produce confident errors.

A citation below an answer is therefore not proof by itself. Open the link, confirm that the page exists, evaluate the publisher, and check that the cited passage actually supports the claim.

How can you reduce the risk?

A useful rule is: the higher the cost of an error, the more verification the answer requires.

  • Ask the model to separate confirmed information from inference or uncertainty.
  • Request direct sources, then open and read the relevant passages yourself.
  • Compare important claims with an official source or multiple independent sources.
  • Break complex questions into steps and correct false premises before requesting an answer.
  • Do not rely on a model alone for medical, legal, financial, or safety-critical decisions.
  • When analysing a document, ask for the exact page or passage supporting each important claim.

Asking for sources is useful, but it does not prevent fabrication because a model can invent sources too. A citation becomes valuable only after it has been verified.

Is hallucination the same as lying?

Calling a hallucination a lie can be misleading because lying normally implies intent and awareness of the truth. A language model generates text from learned patterns, instructions, probabilities, and the context available to it.

NIST also notes that terms such as “hallucination” may anthropomorphize AI systems by suggesting human perception. Alternatives such as “confabulation” or “fabrication” are sometimes used, but “hallucination” remains the most familiar term.

In short

An AI hallucination is a false or unsupported output presented in convincing language. It can result from training data, probabilistic generation, missing information, incomplete context, or evaluation incentives that reward guessing.

Search, retrieval, better training, and improved evaluation can reduce the problem, but they do not turn model outputs into guaranteed facts. The most important question is not only, “Did the model answer?” but “Can I verify its answer with a reliable source?”