Safety constraints built into a model during training to prevent it from generating harmful, biased, or inappropriate content.