AI Hallucination: What It Is and How to Avoid It

Level	Beginner

A simple definition

A hallucination is a convincing mistake. The AI does not flag any doubt: it states false information in the same confident tone as true information. The term is figurative, since the model does not “see” anything. It simply produces plausible text that does not match reality.

Concrete examples

You ask for the biography of a little-known person: the AI invents a birth date or a degree.
You ask for references for a term paper: it cites articles with believable titles that do not exist.
You ask about a specific law: it gives a section number and wording that are wrong.
You share a document: it “quotes” a sentence that is not in it.

Why it happens

A large language model predicts what comes next in a text. When the information is missing from its data, rare or ambiguous, it still fills in with whatever “sounds right”. In their paper “Why Language Models Hallucinate” (arXiv, 2025), OpenAI researchers explain that common evaluation methods favor models that venture an answer over those that admit they do not know, much like a student guessing on a multiple-choice test with no penalty for wrong answers.

Then there is the knowledge cutoff: without web search, the model is unaware of anything after its training ended and may fill that gap.

Why it matters to you

An unnoticed hallucination can be costly: an email quoting a wrong price to a customer, an assignment with fake citations, a decision based on a made-up figure. Good habits:

Check anything specific: names, dates, numbers, quotes, links.
Give it the source yourself: an AI working from your document makes fewer mistakes. This is the idea behind grounding and RAG.
Ask “where does this come from?” and then actually open the links.

Common myths

“If the AI cites a source, it’s reliable.” The source itself may be invented.
“Paid models don’t hallucinate.” They may err less often, but none is immune.

To understand where the model data shown on this site comes from, see the methodology page.

Related terms

Grounding · Making an AI base its answer on supplied sources (web results, documents, databases) rather than only on what it memorized in training.

RAG (retrieval-augmented generation) · A technique that first retrieves relevant passages from your documents, then hands them to the model so its answer is based on them.

AI search · A search where an AI reads web pages for you and writes a summarized answer with its sources, instead of a list of links.

Knowledge cutoff · The date after which an AI model received no more training data, so it knows nothing about later events.

Large language model (LLM) · A program trained on huge amounts of text to predict what comes next in a text; it is the engine behind AI assistants.

Temperature · A setting that controls how much randomness goes into a model's answers: low keeps it predictable, high makes it more varied.

Frequently asked questions

Why does AI make things up?

A language model produces the most plausible continuation of a text, not the most truthful one. A paper published in September 2025 by OpenAI researchers adds that training and evaluation methods often reward guessing over admitting uncertainty.

Will hallucinations go away?

They are becoming less frequent with some models and techniques, but no provider guarantees they are gone. It is safest to assume any answer may contain an error, especially on precise details.

How can I reduce the risk?

Provide the reference documents yourself, ask for sources and check them, allow the AI to say « I don't know », and use web search for recent information. For medical, legal or financial topics, check an official source or ask a professional.

Hallucination: when AI makes things up with confidence

Key figures