[abdullahkhalids]

> If you don't feed race in as a feature, however, the outputs are hardly racist.

The article addresses this

> When “race-neutral” approaches are employed in model development, prediction will tend to be poorer for racial minority populations.... Two explanations for differentially poorer model performance can be addressed by collecting more data: too few observations of members of racial minority groups and unrepresentative sampling that can differentially limit generalizability. However, an additional cause of algorithmic bias is not well appreciated and cannot be overcome simply by adding more of the same kind of data to a learner....

[ralusek]

That is a completely different argument, and has nothing to do with structural racism. That is literally just saying that minorities are less likely to have made up a sizable portion of the data sets trained on, because they're minorities, and the model is potentially less well suited to deal with issues specifically related to that minority. If the primary point of the article was that we should overcorrect for this by making including disproportionately high representation of minority data, then that's a potentially reasonable case, so long as it doesn't break the model. In the case of facial recognition not working as well on non-whites, for example, I think it's an entirely reasonable case to make to include a disproportionately higher amount of training data on those areas where the model fails to perform its function.

But you also have to realize that this is always going to be somewhat arbitrary.