Algorithms, Race, and Reentry: A Review of Sandra G. Mayson’s Bias In, Bias Out
In true Minority Report fashion, state actors are increasingly relying on algorithms to assess the risk a person will commit a future crime. Unlike Minority Report, these algorithms simply estimate the likelihood of rearrests; they do not offer the absolute answer to future criminal behavior that condemned the defendant, Tom Cruise, in the 2002 action film. Still, criminal justice actors are using many types of algorithmic risk assessments to inform their decisions in pre-trial investigations, bail recommendations and decisions, and post-trial sentencing and parole proceedings. Sandra G. Mayson’s article, Bias In, Bias Out, published this year in the Yale Law Journal, explains how these algorithms could reflect and project past and present racial bias in the criminal justice system and elsewhere.
At its core, an algorithm specifies individual traits that are correlated with crime commission. If the data show that people of color are arrested more frequently, then the algorithm will predict more arrests for people of color. In this sense, an accurate algorithm “holds a mirror to the past” by “distilling patterns in past data and projecting them into the future.” Mayson provides an in-depth, yet easy-to-follow explanation of why race neutrality is unattainable when the base rates of arrest differ across racial groups. These mirror-like algorithms give us the opportunity to clearly view the racial disparity in arrests and convictions. Is there something wrong with this image, and what should we do now that we’ve seen it?
Mayson’s analysis begins by focusing on the first of these questions—is there something wrong with the image we are seeing in our algorithmic mirror? If there are different offending rates across racial groups for a given underlying crime, then this image represents an accurate depiction of actual crime rates. If, however, there are not different offending rates across racial lines, with differential arrest rates merely the result of enforcement strategies, then the image depicts how people of color are targeted for arrest with greater frequency. Mayson does not dwell on this point, for good reason: unless we know true offending rates, it is impossible to know which of these scenarios is our reality.
The next question: what should we do now that we have seen this image? It may seem that the answer to this question depends on the answer to the first question, whether you think the image accurately reflects rates of crime commission. However, Mayson’s solution responds to either scenario, a difference in offending rates or a difference in enforcement practices. Before she explains how she would enact her “supportive” response, Mayson explains why the typical responses to the impossibility of a race-neutral algorithm are unproductive.
In response to the increasing use of algorithms, some scholars have called for “colorblind” algorithms (algorithms that do not take into account race or proxies for race) or “algorithmic affirmative action” (algorithms that are programmed to predict the same rate of adverse outcomes or same percentage of false-positives or false-negatives for each racial group). Both of these attempts at equality distort the algorithmic mirror, and thus corrupt the accuracy of the predictions. Mayson argues that such clouding of the data could harm communities of color. Forcing the algorithm to deliver these equal rates of either false-positives or false-negatives will likely lead to a higher error rate in predictions. Take bail decisions, for instance. Since there are more arrestees of color, more persons of color who are going to commit future offenses will be released (because of the compromised algorithm). This “wrongful” release will hurt communities of color because the majority of crime is intraracial. Moreover, “colorblind” algorithms could harm defendants of color because in some instances a person’s race could be a mitigating factor in her risk assessment. For example, Mayson recounts her experience as a public defender in New Orleans where multiple past arrests for a black man simply proved that he had been living in New Orleans, while multiple arrests for a white man likely showed something more ominous. An ideal algorithm would not be blind to race; it would consider the racial context of each factor, in each jurisdiction. This type of specialized decision-making would only be possible with a machine-learning program.
Rather than try to remedy the algorithms, some people have advocated ignoring the algorithm completely. Rejecting these algorithms would mean that criminal justice actors would be left to make their own subjective decisions informed by their own biases. Mayson argues that the algorithm’s predictions are more transparent than individual’s choices; therefore, the algorithm can more readily be held accountable for the racial disparity it predicts.
Mayson embraces the algorithm. She calls for the most competent algorithm possible, one that will deliver the most accurate predictions. Her solution ultimately responds to both scenarios—an underlying difference in offending rates and a difference in enforcement.
The algorithmic assessment of risk is just that, an assessment of risk, not an assessment of blame. Because of this important distinction, Mayson suggests an approach that is supportive rather than coercive or punitive in bail decision making. She discusses the usual coercive responses of custodial detention and GPS monitoring that both are costly and may not work effectively to prevent future crime. She suggests instead that the courts take a “supportive, needs-oriented response,” one that would help eliminate the barriers that arrested and convicted persons face in their quest to lead lawful lives. Instead of the default response being one of incapacitation, the court system should strive to help the defendant succeed and avoid future crime commission. Mayson appears to model this solution on a program for which she was a volunteer as a law clerk. The Supervision to Aid Reentry (STAR) program, run by Magistrate Judge Rice of the U.S. District Court for the Eastern District of Pennsylvania and Judge Restrepo of the U.S. Court of Appeals for the Third Circuit, focused on individuals who were at risk for committing future violent crimes. The STAR reentry team worked on obtaining housing, education, counseling, and substance-abuse treatment for defendants. Meanwhile, law students enrolled in a clinic helped defendants navigate the court system to restore their driver’s licenses and pay fines. A university psychology department developed a cognitive-behavioral therapy option for defendants. Overall, Mayson and others have reported, the program was successful.
This type of supportive response would aspire to mitigate a possible difference in enforcement. Although criminal justice actors would still target people of color, this targeted response would offer help rather than arrest and detention. In addition, Mayson advocates responding to a possible difference in underlying offending rates by using algorithms to diagnose areas of need. She believes that any difference in offending rates is the result of years of inequality and disadvantage. An algorithm could first diagnose the areas that are in need by showing us which areas possess the most at-risk individuals for future crime commission. Then, public or private funding could target these areas and establish community-support programs, and greater investments in schools and housing. A supportive response would hope to counteract some of the structural disadvantage faced by these communities. Thus, Mayson’s proposed approaches responds to either idea of the image in the algorithmic mirror. On the one hand, Mayson proposes tackling institutional inequality in response to the image of actual difference in offending rates. On the other hand, she seeks to mitigate the effect of a difference in enforcement by shifting from a punitive response to a supportive one.
The primary problems I see are not with Mayson’s evaluation of the risk-assessment tools, but with her proposed “solutions.” The approaches she advocated would be costly, and, realistically, would be small-scale and jurisdiction-dependent. The algorithms themselves, at least in Mayson’s view, would need to be calibrated to a particular region and its racial context in order to most accurately predict crime commission. On the other end, the response to the risk assessments would be individualized and intensive. Who is going to pay for all of this? One can certainly expect that many officials and judges would prefer to continue with custodial detention and GPS monitoring because the infrastructure for such programs already exists.
Programs like STAR are not only costly, but also hard to reproduce on a large-scale. These programs seem to require the help of either a full-time staff, or an army of university volunteers (or free labor sources like over-eager law students). Such intensive support is not going to be possible in less urban areas that do not have access to university resources. Moreover, the support imagined is individualized for the defendant, and thus tailored to helping the defendant succeed in that region. Some larger, less homogenous states would have difficulty enacting state-wide policy guidelines or restrictions because the help a defendant needs could vary depending by locality in the state. Moreover, counties or municipalities themselves would need to devise these programs with state funds. That may be politically tricky, and without strict state guidelines and oversight, localities could devise drastically different programs—such that similarly situated defendants in the same state would not receive the same treatment. And what do we do with these algorithmic risk assessments in jurisdictions that reject Mayson’s solution because it is too costly, or otherwise not feasible for their region—perhaps simply due to lack of political will?
Mayson mentions that jurisdictions are increasingly turning to algorithmic tools at every stage of the criminal justice process, from investigation to sentencing. If we accept Mayson’s belief that these tools are reproducing racial inequality, is there a way to address that problem short of a complete overhaul of extant criminal justice procedures and responses? Essentially, we need a Plan B.
Finally, even if all of Mayson’s ideal approaches were adopted, there are downsides. People of color would be faced with greater government interference in their lives. Mayson accepts an algorithm that produces a higher rate of false-positives for people of color because she calculates that outcomes will improve the lives of the larger communities. That may well not pan out. Some states following Mayson’s logic may implement mandatory education programs, mandatory mental health treatment, or mandatory substance-abuse treatment. State power, in the form of “supportive” programs, would still infringe, at a higher rate, on the liberty of people of color. There could, and likely should be a discussion about whether programs like STAR are productive or beneficial for communities of color. But we should keep in mind that at the end of the day, such programs would still be state policing of black lives.
 Sandra G. Mayson, Bias in, Bias Out, 128 Yale L.J. 2218 (2019).