Built-in constraints that prevent a model from generating harmful, offensive, or inappropriate content in its responses.