The Problem with 'Friendly' Artificial Intelligence