An Apple document reveals how AI digital assistant responses are rated for harmfulness, truthfulness, satisfaction, and more.