I find Russell's assumption that machines will learn good behavior from humans to be incredibly naive. History is a nightmare from which we cannot escape, as Joyce said. A robot exposed to news and general human behavior today would learn to lie and kill just as well as it ties knots. I hope he has a better plan than that.

I think the idea is that the machines are supposed to learn what humans want from human behavior, and then help humans get what they want their own way, rather than imitating human behavior.

An AI would have to have a good parent, like everybody else. Of course.

Was just discussing with my wife (teaches user-centered design) that some in the HCI profession will be migrating to this.

Anyone have any good Arxiv research papers on this inverse DRL?