Hallucination
The model produces fluent output that is wrong. It isn't an occasional glitch but an inherent property of this class of model — the objective is plausibility, not truth.
It made that up.
What share of the output have we actually spot-checked?
Why this one gets stretched
Calling it a hallucination makes it sound like an intermittent illness that will eventually be cured. As long as the generation mechanism is the same, it doesn't disappear — it just gets harder to catch, because fluency keeps improving and fluency is exactly what makes people stop checking.
Usually heard: Right after someone takes model output into a meeting for the first time.
Related terms
The literal definitions aim to be accurate. The only column joking on this page is the meeting-room one.