Definition
Plain language
When an AI makes up an answer out of nothing, rather than recalling something it once learned.
As stated in the literature
A hallucination subtype in which a model fabricates content with no training-data grounding; in drift work, its MLP retrieval dynamics look nearly identical to stale recall, which is why output confidence cannot separate the two.
Also called: confabulate, confabulated
Why it matters: Because a made-up answer can look just as confident as a remembered one, this failure is hard to catch and undermines trust in everything the model says.
For example, a model invents a plausible-sounding citation for a study that was never written, complete with authors and a title.
Heard on the show
“… Organic confabulation from an undefended stripped model is hazard-increasing in only a small sliver of draws, and the …”Episode 244 — The Open-Weight Defense That Feeds Attackers Confident, Falsified Answers