Built-in guardrails in a model that prevent it from generating harmful, illegal, or unethical content by refusing certain requests.