“The ethics of a perfectly user-corrigible AI model are equivalent to the ethics of its user.”