Predictive algorithms cannot cleanly label predictors legitimate or biased proxies

Evidence

When the legal technology employed makes use of predictive analytics, there are numerous methods for addressing bias in the historical record (see Corbett-Davis et al., 2017; Hardt, Price, & Srebro, 2016; Kamiran, Zliobaite, & Calders, 2013). Many of these methods rely on determining which predictors are legitimate (e.g., which ones are valid in determining predicted recidivism) and which are illegitimate (e.g., race or other features—such as zip code—that might function as mere proxies for race). While we just gave a few examples, doing so elided that it is difficult to determine which predictors are legitimate. Harcourt (2015), for instance, argued that criminal records are proxies for race. Is this true? An argument can be made that criminal convictions are legitimate predictors but criminal arrests are not, since the latter might reflect policing bias. With such an approach, in subsequent generations those with clean records will be treated equally; but this approach acknowledges that not all past injustice can be corrected with one algorithm. It should be clear that decisions like this one—how to handle criminal records—require deep knowledge of the area of expertise (Ho, 2018). — Joseph Avery, Jim Sidanius, Joel Cooper, Gregory Davis, Amanda N., Bias in the Law, 1184

Categorization

• DEFINE: Clarifies the legitimate/illegitimate predictor distinction and why it is intrinsically contested.
• XREF: connects to: Algorithmic filtering amplifies preexisting biases rather than causing them — Sander van der Linden, Foolproof: Why Misinformation Infects…, 414