The bias comes from the human-generated training data in the first place; the machine isn't introducing its own. For instance, the machine has no inherent concept of disparaging someone's language because it's from an identifiable inner city dialect. If it picks up that bias, at least it will apply it consistently. When we investigate the machine, the machine will not know that it's being investigated and will not try to conceal its bias from us.
On the other hand, eliminating bias from humans basically means this: producing a new litter of small humans and teaching them better than their predecessors.
AI certainly has a lot of potential for bias, but claims that AI bias is somehow worse than good old human bias always seem shoddily supported (Note I'm not claiming it's untrue. Just that it's never been shown to my satisfaction, which is not surprising given how quickly AI is changing. It may well be true.)
Well, AI bias can combine sampling bias with human bias. Like say we train the AI with the output of only 10 human paper graders, all chosen from the same school district.
Due to the sampling bias, that data could create markedly more (or less!) bias than the entire population of human paper-graders.
The resulting AI will ideally mimic those 10 humans, though; it shouldn't show more bias than that group. If those 10 are flaring racists, and grade accordingly, the AI will be the same. (In fact, we hope that it will be the same, if the algorithm actually works in mimicking human grading.)