Hallucination: Definition and Examples

In artificial intelligence, a hallucination occurs when a language model generates information that does not exist, is factually incorrect, or does not correspond to any reliable source. The term is borrowed from psychiatry, where it denotes a perception without a real object. In the AI context, it describes the tendency of models to 'invent' facts with an appearance of certainty.

Hallucinations arise because language models (LLMs) operate by statistical prediction of the next word. They do not 'know' in the human sense: they generate probable text based on learned patterns. When the model lacks information on a topic or the question is ambiguous, it may produce a response that is linguistically coherent but factually wrong — invented citations, fictitious statistics, events that never occurred.

The main danger of hallucinations lies in their appearance of credibility. Unlike a search engine that displays 'no results,' an LLM will almost always produce a response, even when it should admit ignorance. This can lead to the spread of misinformation, especially when the user does not verify sources.

Several techniques can reduce hallucinations: Retrieval-Augmented Generation (RAG) which anchors responses in verified documents, prompting with source requests, lowering generation temperature, or explicit instructions asking the model to say 'I don't know' rather than invent. Understanding this phenomenon is a central challenge of modern prompt engineering.