The answer was confident. Detailed. Completely wrong. The model did not know it was wrong.
Literal meaning: An AI hallucination is a response in which a language model produces information that is factually incorrect, fabricated, or entirely invented — stated with the same fluency and confidence as accurate information.
Origin: The term was borrowed from psychology and neuroscience, where hallucination describes perception without external stimulus. Applied to AI systems from around 2020, it became standard terminology in machine learning research. The borrowing was not neutral: it imports a vocabulary of subjective experience into a description of a statistical process. Models do not hallucinate in any meaningful psychological sense — they generate plausible-sounding text based on patterns in training data, without any mechanism for checking whether it is true.
A language model output that is confidently and fluently wrong — indistinguishable in form from a correct answer.
The Appeal: The term is vivid and communicates the phenomenon quickly. It has become useful shorthand in public discourse about AI limitations, and has helped non-technical audiences grasp that AI systems can be wrong in ways that are hard to detect.
The Friction: The terminology does real work. Calling it a hallucination implies a mind that perceived incorrectly — which subtly obscures the actual mechanism: statistical text generation that has no access to ground truth. Sycophancy (AI) — the tendency of AI systems to confirm what users want to hear, baked into training — compounds the problem: a model that agrees with false premises and generates false elaborations is not malfunctioning. It is working as designed. AI Literacy — understanding how AI works, not just how to use it — is exactly what the hallucination metaphor can undermine: it makes the system sound like it experienced something, rather than computed something.
Why This Matters: Once you know how hallucination actually works — pattern completion without truth-checking — “the AI said so” becomes a different kind of sentence. The fluency is not evidence of accuracy. It never was.
Related terms: Sycophancy (AI) · AI Literacy · AI Dependency · Cognitive Offloading · Deskilling
Read more:
- Survey of Hallucination in Natural Language Generation — Ji et al. (2022)
- On the Dangers of Stochastic Parrots — Bender et al. (2021), ACM FAccT