Confabulation

Confabulation is a term borrowed from clinical neurology, describing the unintentional habit of filling memory gaps with confident, plausible-sounding inventions. AI safety and interpretability literature has increasingly preferred this label over Hallucination for LLM fabrication, because the model isn't 'seeing' anything — it is simply completing missing information with what feels likely. The framing helps explain why models confabulate especially around topics underrepresented in their training data, and why mitigations like RAG, Grounding and explicit 'say I don't know' instructions matter so much. The terminology shift may sound minor, but framing the failure mode correctly is a prerequisite to designing the right defences.