Algorithms can be reweighted to equality-neutral treatments that correct documented bias.

Evidence

Third, there are countless ways to employ legal technology that would not run afoul of nondiscrimination principles. The algorithm could simply treat everyone like the best-off group: that is, male defendants would be treated like female defendants and female defendants would be treated like female defendants. Alternatively, an average across the groups could be established and everyone could be treated according to this average: male defendants would be treated like male/female average defendants and female defendants would be treated like male/female defendants. As another alternative, the importance of the model could be decreased, and it could be used in a strictly advisory capacity. It would provide guidance, and prosecutors would be told that it had been reweighted to correct for previously documented bias. Regardless of which path is taken, it is certain that moving away from data and away from legal technology is not a solution. We think it would be a mistake, as it ensures that the status quo, where disparate outcomes caused by biased humans are the norm, will remain. — Joseph Avery, Jim Sidanius, Joel Cooper, Gregory Davis, Amanda N., Bias in the Law, 1209

Categorization

• DEFINE: Lays out concrete technical strategies for de-biased algorithmic treatment: treat-all-as-best-off, group averaging, or advisory-only reweighting.