Models tend to default to their training data even when they lack sufficient context, they've never been trained to recognize their own uncertainty, so they hallucinate confidently instead.