It isn't dark skin, per se, that confuses the algorithms in the software. So, there isn't a line of code that says, in essence, "If subject's complexion is swarthy, then crash!"
Rather, as the excerpt pointed out, more White men than, e.g., dark-skinned women were used to "train" the software.
So, to balance out the software, they could simply use more dark-skinned and female test subjects.
Regards,
What exactly is the task?
Is it...
(a) determining the identity of someone based on their picture?
(b) determining the sex of someone based on their picture?
If it’s (a), I would think they would have the same amount of training data per subject, but maybe not...
if it’s (b), is a picture of someone not in the training set given as a test, or are the test pictures confined to the subjects?