I think it's indirectly relevant in a disturbing way; if you had a perfect "algorithm"[1] to predict (say) recidivism (let's not quibble over how it works or exactly what risk factors it would assess, let's just posit it is very accurate) then that algorithm would be racist because its predictions would be correlated with the race of the prisoner.

So we can either have fair, colour blind algorithms or accurate algorithms, but we can't have both.

[1] the term "algorithm" is used in the article but it sounds more like a risk assessment framework.