It's not provocative, he says antifa once. It's meant as an example of a method of removing bias. He's saying you could train the AI entirely on an organizations strictly vetted and approved data, it could be antifa, the Catholic church or Coca Cola, it would still potentially say unhinged things even if you eliminate bias. Top many of these concepts are built into the language. Look at all the ways we use the word kill for a variety of topics, most being very benign, but it can be disturbing if the AI starts talking about killing things.
