hallucination is a perception — a sight, sound, smell, or touch — that seems real but arises without any external stimulus, and, by extension, a plausible but false output produced by an artificial intelligence system.
In its original, perceptual sense, what defines a hallucination is the missing stimulus: hearing a voice when no one has spoken, seeing a figure in an empty room. Hallucinations can involve any sense and are associated with neurological and psychiatric conditions, high fever, and extreme sleep or sensory deprivation. The word is often confused with its neighbors. An 📝illusion is a misperception of something actually present, such as a stick that looks bent in water; a delusion is a fixed false belief rather than a false perception. Some neuroscientists push the boundary further, describing ordinary perception itself as a kind of controlled hallucination — the brain's best guess about the world, held in check by the senses — which is the question the lesson below takes up.
In artificial intelligence, hallucination names output from a model, typically a 📝Large Language Model (LLM), that is fluent and confident but fabricated or unsupported: invented citations, events that never happened, facts with no source. The usage surged with 📝generative AI, and Cambridge Dictionary chose hallucinate as its Word of the Year for 2023 for that new sense. Some researchers prefer confabulation, arguing that a model perceives nothing and so cannot hallucinate in the clinical sense. Hallucination rate, the share of unsupported claims in a model's output, is now a standard evaluation measure; it is a recurring concern in 📝AI Agent Evaluation and in debates over whether 📝Chain of Thought Reasoning reflects genuine reasoning.
The noun entered English in the 1640s, meaning the seeing or hearing of something that is not there, from Latin alucinari (later hallucinari), "to wander in the mind, dream, ramble in thought," likely from Greek alyein, "to wander in mind, be at a loss."
I saved the TED-Ed lesson below in 2019, years before the word moved into AI. Now it runs through my AI notes as a reliability risk to design around, which is why the 📝Anti-Hallucination Protocol (prompt) lives in my library.