Built-in safety features that cause a model to decline responding to certain types of requests, such as those involving harmful, illegal, or unethical content.