Hallucination

The phenomenon in which an AI system, particularly a large language model, produces fluent, confident, but factually incorrect or entirely fabricated content, because the system generates statistically likely text rather than verifying truth against a fact database. Hallucination is close to an inherent property of how these systems currently generate output, not a rare malfunction, though techniques like retrieval-augmented generation can reduce its frequency.

In generative AI, an output that is fluent and confident but not supported by the source or by fact, for example a summary that states a medication the source note never mentions. The term is imprecise until defined for a specific task; a defensible metric replaces the word with an exact, gradable rule.

When an AI system states something false with full confidence, because it generates the most plausible output rather than the true one and cannot tell the difference. Not a bug awaiting a patch but a property of the technology, which is why high-stakes uses need a verification layer.

When an AI states something false with confidence, such as inventing a statistic, source, or date it does not actually have. Instructing the AI to admit uncertainty is a first-line guardrail against acting on hallucinated information.

When generative AI produces confident, coherent output that is factually incorrect; a fundamental property of statistical text generation.