Hallucination
Chapter 6 · page 96
Definition
A hallucination is the phenomenon by which a model generates false, invented, distorted or misleading information while presenting it with apparent coherence or confidence. Language models can thus produce answers that are credible, well written and convincing… but factually wrong. This follows from the very way a probabilistic model works: it does not check whether what it asserts is true; it simply predicts the most plausible continuation of words, given its training and the context of the conversation.
Example
Ask a chatbot to cite sources on a highly specialized subject and it may invent references that do not exist: “According to a study by Miller published in 2017 in the journal Health and Cognition…”, when neither the study nor even the journal cited exists. In the same way, an AI system can attribute a quotation to the wrong person, invent a court ruling or describe an imaginary feature of a piece of software. The more ambiguous, rare or hard to verify the request, the greater the risk of hallucination.
Why it matters
Hallucinations can be harmless in a humorous context, but troublesome in sensitive fields such as health, education or law. To limit them, current systems use various approaches: better training, access to up-to-date external sources, RAG-type architectures, and so on. These techniques can improve the reliability of answers, but they do not turn a probabilistic model into a fact-checking system. A convincing answer must therefore always be assessed critically.
How to cite this text
Vial, S. (2026). Hallucination. In A Living Lexicon of Artificial Intelligence (p. 96). Stéphane Vial, publisher, Montréal.
