Confabulation (LLMs)
Term increasingly preferred over "hallucination" for fluent but factually invented LLM outputs, on the grounds that the underlying mechanism is gap-filling generation, not perceptual misfire.
Confabulation, borrowed from clinical neuroscience, describes the production of coherent narrative content that fills gaps in memory without intent to deceive. Applied to language model behavior, it names outputs that are syntactically fluent, stylistically appropriate, and factually invented — generated citations, nonexistent API calls, plausible-but-false historical claims. Researchers including Beren Millidge and clinicians publishing in PLOS Digital Health have argued that "confabulation" is mechanistically more accurate than the more common term "hallucination." Hallucination implies a perceptual error: seeing something that is not there. LLMs have no perceptual channel. What they do is sample from a learned distribution over token sequences; when the distribution does not concentrate near the truth, the model emits the next most plausible continuation. That is gap-filling, the defining feature of confabulation. The distinction has practical consequences. Treating the failure as perceptual suggests fixes like better "grounding in reality." Treating it as confabulation points at the training objective itself: a system optimized for plausibility will confabulate whenever plausible and true diverge. Mitigations follow from that framing — retrieval, verification, abstention training, and uncertainty estimation — rather than from exhortations to the model to be more truthful.