Algorithms fail to improve predictions

Research MethodsIn the United States, algorithms are commonly used to predict the likelihood that a criminal defendant will commit a crime, and these predictions influence pretrial, parole, and sentencing decisions. Commercial software, such as the widely used COMPAS system, promises to make these predictions more accurate than human judgments. Dressel and Farid show that COMPASu0027s impressive-sounding 137-feature black box is nearly equivalent to a trivial linear classifier using two features, and both approaches are no more accurate or fair than predictions made by people with little or no criminal justice expertise.Sci. Adv. 10.1126/sciadv.aao5580 (2018).