What If the Safest AI Is One That Doesn’t Always Agree With You?