Why would AI companies use human-level AI to do alignment research? — LessWrong