Hallucination

In short

A hallucination is when an AI tool confidently states something false or made up, such as a fake fact, quote or source. It happens because chatbots predict likely-sounding text rather than look facts up.

Also called: AI hallucination, Confabulation, Fabrication
Topic: The basics
Level: Beginner
First described: The term was used in machine translation and image research for years before chatbots made it mainstream. Ji and colleagues' 2023 survey is the standard reference³.

Chatbots write by predicting the next most likely word. Most of the time that produces true sentences, because true sentences are common in their training data. But a fluent, plausible answer is not the same as a checked one.

The US standards body NIST prefers the word "confabulation" for this, describing it as AI systems that "confidently present erroneous or false content", and notes it is "colloquially" called hallucination¹. NIST also points out that some people object to "hallucination" because it makes software sound human¹.

A 2025 OpenAI paper argues one cause is how models are trained and tested: scoring rewards a confident guess over saying "I don't know"². A widely cited survey of the problem groups hallucinations into ones that contradict the source given and ones that add things the source never said³.

Tools that search the web or quote documents (see retrieval-augmented generation) make hallucinations less likely, but do not remove them.

Example

In 2025 the High Court in London warned lawyers after fictitious case citations, which the court found had likely been produced with AI, were put before it. The judgment states that freely available AI tools "are not capable of conducting reliable legal research"⁴.

Why it matters

It is the main reason not to trust an AI answer on anything important without checking. A made-up source looks exactly like a real one.

Questions people ask

Why do AI chatbots make things up?

They generate text by predicting likely words, not by looking up facts. A 2025 OpenAI paper adds that training and testing reward confident guessing over admitting uncertainty².

Can you stop an AI from hallucinating?

Not completely. Asking it to cite sources, giving it the documents to work from, and using tools that search the web all reduce the risk. Checking important claims yourself is still necessary.

Is hallucination the same as lying?

No. Lying means knowingly saying something false. A model has no intent; it produces a confident answer that happens to be wrong. That is why NIST prefers the word confabulation¹.

Related terms

Tools that use it

Sources

Back to the glossary