Optimal predictive legal models deliberately include race as a feature
Evidence
Machine learning involves using feature frequency information to statistically map correspondence between features and outcomes. For the present purposes, these are case features (arrest charges, arrest report, etc.) and case outcomes (in most instances, the plea bargain reached). Of course, case features also include defendant characteristics, such as age, gender, and race. Thus, if one seeks to build an optimally predictive model, race should be included in the model (see Mayson, 2019; see also Bartlett, Morse, Stanton, & Wallace, 2018). Indeed, this approach, in which bias is a feature, is the usual one when it comes to predictive analytics in the law. — Joseph Avery, Jim Sidanius, Joel Cooper, Gregory Davis, Amanda N., Bias in the Law, 1218
Categorization
Category: Contrarian or Surprising Claims
- XREF: Contrasts with the fair-ML debate (COMPAS, equalized odds); connects to the reader’s prior chapters on COMPAS risk scores encoding bias.
- DEFINE: Clarifies the standard ML approach where bias is a predictive feature versus a debiasing approach.
Related
- Standard legal predictive models deliberately include race to maximize accuracy. (similarity 0.96)
- A suggestive model removes race to predict race-neutral case outcomes (similarity 0.83)
- A bias-removing model should predict counterfactual race-neutral outcomes, not likely ones. (similarity 0.82)
- Predictive algorithms cannot cleanly label predictors legitimate or biased proxies (similarity 0.80)
- Legal AI can be deployed to identify and correct existing bias in law (similarity 0.80)
- Predictive models can suggest case outcomes under counterfactual race-neutral processing. (similarity 0.80)
- Race-blind prediction models can encode bias; active debiasing improves fairness. (similarity 0.78)