Johns Hopkins researchers show how different "odds" can teach AI models to admit when they're not confident enough in an ...