A new onboarding technique from MIT can help humans collaborate more effectively with AI assistants. The process finds situations where the human either over-trusts or under-trusts the AI, describing these as natural language rules, and creates training exercises based on these rules to guide the user.