Computerized assessment algorithms are gaining traction within the criminal justice system, being used for such tasks as setting bail conditions and determining criminal sentences. Often, researchers will develop these algorithms by determining what factors might predict an offender's future behaviors, such as the risk that he or she will be violent or commit another crime. These predictors may include demographic factors such as age, sex and race; historical factors such as the age of criminal behavior onset and the nature of prior arrests; and social factors including housing stability and social support.
Take, for instance, the case of Eric Loomis who was arrested in 2013 for alleged involvement in a drive-by shooting. After his arrest, Loomis was taken into custody and administered the Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) risk-assessment tool. This information was fed into a computer and identified Loomis as "high risk" in all three COMPAS recidivism risk scales: pretrial risk, general risk and violent risk. These scores were heavily considered when sentencing Loomis to six years in prison.
These largely automated assessments appeal to courts because they are efficient in determining the risk of a future offense and avoid expending the resources required for a full forensic evaluation. As such, legal decision-makers can drastically increase the expediency of their decision-making within an often slow and overburdened system. Also, these algorithms promise to yield consistent and reliable assessments, helping to minimize human bias and encourage fair and equitable application of the law.
But there are risks to using these algorithms, as illustrated in Loomis's appeal to the Wisconsin Supreme Court. After his sentencing, Loomis filed a motion for post-conviction relief arguing that the court's reliance on an opaque algorithm violated his due process rights. Although the court affirmed the sentence, it held that presentence investigation reports relying on the COMPAS must include five warnings regarding its limitations: 1. Its proprietary nature means that the exact risk factors used are unknown; 2. It identifies high risk groups but cannot meaningfully inform about specific individuals; 3. Some data suggest it may be racially biased; 4. It has not been validated or normed on a Wisconsin sample; and 5. It was not developed for use at sentencing (State v. Loomis, 2016).
Researchers have raised concerns that the use of algorithms could discriminate against certain groups, such as those who live in particular zip codes and those with longer arrest histories. In addition, some have argued that inputting historically biased data into these models contributes to a perpetuation of "bad data" because the algorithm uses each new case as a reference for the next.
As the legal, ethical and statistical debate concerning the use of computerized risk algorithms in judicial decision-making continues, so does the refinement and integration of the technology itself. One possible direction the field could take would be incorporating professional judgment into these models to address the concerns of those who oppose a strict actuarial/statistical model of judicial decision-making. In the meantime, educating judicial decision-makers about the strengths (e.g., efficiency, consistency) and limitations (e.g., due process concerns, discrimination) of these tools is essential. Finally, research that addresses outcomes associated with computer risk algorithms would add important empirical evidence for or against the continued use of such tools.
"Judicial Notebook" is a project of APA Div. 9 (Society for the Psychological Study of Social Issues).
