Training models to follow conversational principles of cooperation, like being truthful and informative—retreating to safer claims when uncertain.