That was a good read. So in essence, what has been happening so far in many cases is that the AI didn't remove the bias, but rather automated it
On a more serious note, when it comes to facial recognition, it seems like a combination of the quality (or lack thereof) of algorithm training data and physical properties of the universe. I think it's safe to assume that facial recognition is likely to be more accurate on people with lighter skin tones, because more reflected light means more data allowing the software to make a match more accurately, whereas it is more error-prone on people with darker skin?
Also regarding Amazon, it's not surprising their recruitment AI tool turned out to be crap. I mean, even their regular internal search algorithm is fucking horrible when you are trying to find some product on their website!