Why is it hard for AI to detect human bias?