Скачать или смотреть When Conditions Mimic Real-World Criminal Justice, Algorithms Predict Reoffending Better than Humans

Statistical tools outmatch humans at predicting whether a person convicted of a crime will reoffend, a new study shows. While the study begins by replicating the findings of a 2018 Science Advances analysis that suggested humans were just as good as statistical tools in predicting recidivism, its finding that in data scenarios more akin to the real world, statistical tools are better, extends the findings in an important way. The results may inform policymakers’ efforts to safely contain prison growth – which involves identifying and releasing those least likely to reoffend. Risk Assessment Instruments (RAIs) are regularly used to help judges, correctional authorities, and parole boards make decisions about whether a defendant should be released or remain incarcerated, and about what kind of services he or she should receive. Although most research supports that these tools provide more accurate recidivism predictions than unaided human judgment, a 2018 study called their validity into question; in that work, researchers reported that commercial software widely used to predict recidivism, COMPAS, was no more accurate than untrained people at foreseeing recidivism. Their work involved conducting an online survey in which participants saw a brief description of a defendant and then predicted whether each defendant would commit another offense within two years of their most recent crime. Human predictions were approximately as accurate as COMPAS’s — correctly predicting whether or not the defendant would recidivate in roughly 65% of cases.

In this paper, involving 645 participants from Amazon’s Mechanical Turk, Zhiyuan Lin et al. sought to better understand the value of statistical tools in contexts that better approximate the criminal justice system. They found results similar to the previous study when their participants were given feedback about whether predictions were accurate and when they were given streamlined risk factor information. But, algorithms won out over people under conditions more closely resembling the real-life criminal justice system – when people didn’t get feedback about whether predictions were accurate, for example, or when they were presented with complex risk factor information from the RAI that went beyond factors like sex, age, current charge and number of prior adult and juvenile offenses. “These findings lend confidence to the basic principle that algorithms can help one obtain more accurate assessments of risk of re-offending,” say Sharad Goel and Jennifer Skeem, two authors on the study. “But, like any tools, risk assessment instruments must be coupled with sound policy choices to spur broader criminal justice reform. Additionally, these tools must be intentionally designed and regularly audited to ensure equitable outcomes.”