Dodging systematic human errors in scalable oversight — AI Alignment Forum