Method prevents an AI model from being overconfident about wrong answers