You may be under the impression that the sentence a defendant receives in court is based on legal guidelines which attach a certain type of crime to a certain amount of prison time. That’s not entirely true. A mathematical tool is also used to determine recidivism—that is, the likelihood that a convicted criminal will commit another offense within two years. The higher the score the recidivism algorithm produces, the longer the defendant’s sentence. The problem is: the math is biased against black people.
How can math be racist? Researchers for Science Advances recently sought to answer this question.
In the study, they examined a recidivism software called COMPAS, which has been widely used in courts throughout the United States since 2000. The software calculates 137 different features—none of them race—to determine the risk of recidivism. When researchers analyzed the results, they found that COMPAS labeled black defendants who did not later recidivate as “high risk” nearly twice as often as their white counterparts. Conversely, the software labeled white defendants who committed subsequent offenses as “low risk” nearly twice as often as their black counterparts.
In addition, they found that the overall predictive accuracy of the algorithm (i.e., whether the algorithm rightly predicted whether a defendant would reoffend) hovers at around 65 percent for both races. That leaves 35 percent of defendants who receive inappropriate sentences based on their risk of recidivism.
After discovering this unexplained racial bias in the math, researchers went on to examine the algorithm in greater detail. They found that by whittling down the predictive factors from 137 to just two (age and number of prior convictions), the results of the algorithm stayed about the same—still displaying bias toward black defendants.
The problem with the math, they discovered, is that the recidivism algorithm was not actually predicting whether a defendant would commit a crime later on, but rather whether a defendant would commit a crime, get caught, be arrested, be charged, be convicted and go to jail. The likelihood that these events would occur is higher for black defendants.
Recidivism software remains a controversial sentencing tool. Although it is used in courts around the country, the algorithm has not been vetted by the Department of Justice to ensure that it is free of bias.
What do you think? Should recidivism software be used in the sentencing process?