“Because humans seem to prefer responses that are polite, helpful, encouraging, and that don’t contradict the user, the systems learn to be overly “sychophantic,”