“The Alignment problem asks: “How do we build machines with human values?” [but] the data we generate is overwhelmingly anti-social.