Knowra Hallucination (artificial intelligence) Hallucination (artificial intelligence) An AI system’s production of information that sounds plausible but is unsupported or false. The term is especially associated with generative models that produce fluent responses without reliably grounding them in evidence.
Autoregressive language model : A model that generates a sequence one token at a time, conditioning each prediction on previous tokens. Next-token prediction can favor a plausible continuation without checking whether its claims are true.
Grounding (artificial intelligence) : The connection between an AI system’s representations or outputs and external data, observations, or actions. A response is less likely to hallucinate when its claims are anchored to relevant evidence.
Confabulation : The production of invented or distorted recollections presented as genuine memories, especially in human cognition. The term offers a human-cognition analogy, but AI systems do not necessarily possess memories or intentions.
Text summarization : The task of condensing a document or collection of documents into a shorter text preserving salient information. A summary hallucinates when it adds claims absent from the source or alters their meaning.
AI evaluation : The systematic measurement of an AI system’s capabilities, limitations, and behavior across defined tasks. Reliable hallucination evaluation must test evidence use, not just answer fluency.
Pretraining : The initial training phase in which a model learns patterns from a large dataset before task-specific adaptation. Learned statistical patterns supply fluent responses, but do not guarantee factual grounding.
Factuality : The degree to which statements correspond to facts about the world. Factuality is the central property a hallucinated claim fails to satisfy.
Misinformation : False or inaccurate information, regardless of whether it was produced or shared intentionally. A hallucination can become misinformation when circulated, but the terms describe different aspects.
Machine translation : The automatic conversion of text or speech from one language into another. A translation can invent, omit, or distort content while remaining grammatically fluent.
Explainable artificial intelligence : Methods for making an AI system’s predictions or behavior more understandable to people. Explanations may help investigate errors, but a persuasive explanation does not establish factual correctness.
Show all 26