The issue is that the pattern matching reinforces the problem based on race when race is not the cause. There are lots of historical reasons someone could be be a bad credit risk for example, but it isn’t because they’re black or some other race.

You want the algorithm to account for cause not correlated patterns that reinforce racial divisions and then perpetuate the inequality.

It’s not because of the color of someone’s skin. It’s not even clear where racial lines are drawn and previously distinct “races” have disappeared into the general “Caucasian” label depending on time period.

Your comments are why it’s a problem - people are unable to understand the difference and overfit to patterns in the data. Our learning shares a similar problem to ML.

When you don’t correct for this it hurts people of that race and helps to perpetuate the imbalance - reinenforcing the historical data and increasing inequality.

https://www.cell.com/ajhg/pdf/S0002-9297%2818%2930363-X.pdf

"Genetics demonstrates that humans cannot be divided into biologically distinct subcategories."

"Most human genetic variation is distributed as a gradient, so distinct boundaries between population groups cannot be accurately assigned."

etc.etc.

You can only make that claim by examining a single locus. When you examine multiple loci, human populations are clearly separated into distinct groups. The fact that this oft repeated myth is so commonly addressed that it is a named fallacy should be an indication that it warrants a little investigation into its accuracy. Here's a handy graph to help visualize how Lewontin's claim is dishonest and false: https://ars.els-cdn.com/content/image/1-s2.0-S03069877090053...

If you want to argue about the science, please research the subject first. There's no way anyone could do even a cursory bit of research and not come across this as it has been beaten to death, resurrected, and re-beaten to death.