On the lethality of biased human reward ratings — AI Alignment Forum