there is zero academic rigor to any of this. he simply either does not understand the issue or his brain is fully rotten by weird right-wing shit.

black faces appear to have low contrast against a dark background. this is a completely anodyne observation with technical consequences: particularly, that a low-contrast picture of a black face will have fewer landmarks for a model to key on.

in order to get even odds of recognition -- and this may not even be possible -- you have to look independently at the odds of recognition in demographic subgroups. this means adjusting your training set. you even want to do this if you are racist. you want your product to work.

similarly, your mathematical statement of your objective function does not always match a normative account of what you want to accomplish.

let's say you want to build a model which inputs a resume and rates how long that employee would be likely to be retained by the company.

you get a simple answer: men stay at the company way longer than women. you think: oh, sure, that must mean that the men are better employees.

the reason is that your company has a massive sexual harrassment problem and most women find the job intolerable beyond six months.

what you told the model is to make distribution A match distribution B. it did not tell you whether your hypothesis that this was a good idea is true; it just took distribution A and made it look like distribution B.

now, if you have fifty companies you're running this model for, and only one of them is Sexual Harassment Hell, then here's what your model will probably do: hire fewer women everywhere, even at the companies which don't have sexual harassment problems.

it is doing this because it is perversely optimizing an exception you didn't know about to the rule you thought was good. again, your model is not writing the rule "retention predicts desirability at time of hire." you wrote that. the model is doing what you tell it.

when we build a model, we are attempting to accomplish some human goal: we would like to hire people quickly, say, or more accurately. they do this by abstracting some pool of training data. when we embed these systems in human society, we often do not have good ground truth.

when we optimize for something like retention, it is not because retention is itself a good thing. it is because it is a proxy for "it was a good idea to hire them," the goal we want to accomplish. but we do not have access to information about whether it was a good idea.

in these cases, model fairness is not even clearly an attempt to achieve an equitable result: we are only specifying what it is we consider to be a good idea. in the case of the gender-unbalanced model, it introduces the model to the concept of following civil rights law.