COMPAS (software)

• 1. The Algorithmic Gavel

  ›

  In 1998, Northpointe released a software system called COMPAS to help U.S. courts decide how to treat defendants before trial. The company later changed its name to Equivant, but the tool remained a staple in states like Wisconsin and California. Judges used this program to predict whether someone would show up for court or commit new crimes while waiting for their day in court. The software did not just guess; it calculated risk scores based on specific behavioral factors. These numbers appeared on screens during sentencing hearings across the country. A defendant might see a score that suggested they were high risk for re-offending. That number could influence whether they got bail or how long they stayed behind bars.

  15

• 2. Scales of Behavior And History

  ›

  Northpointe designed three distinct scales within the COMPAS algorithm to measure different types of future criminal behavior. The pretrial release scale looked at current charges, pending charges, prior arrest history, and residential stability. It also weighed employment status, community ties, and substance abuse issues against each other. The general recidivism scale focused on predicting new offenses after release by analyzing an individual's criminal history and drug involvement. This scale included indications of juvenile delinquency as a key factor. The violent recidivism scale predicted violent offenses following release using data points like age-at-intake and age-at-first-arrest. It incorporated history of violence, history of non-compliance, and vocational education problems into its final calculation. Northpointe stated these constructs had very high relevance to recriminal careers according to their internal research.

  6

• 3. The Wisconsin Supreme Court Ruling

  ›

  In July 2016, the Wisconsin Supreme Court issued a ruling regarding the use of COMPAS risk scores during sentencing proceedings. The court decided judges could consider these scores but must provide warnings about the tool's limitations and cautions. This decision came after years of debate over whether proprietary algorithms violated due process rights. Critics argued that because the code was a trade secret, defendants could not examine how the software reached its conclusions. They claimed this lack of transparency prevented fair legal challenges. Proponents countered that such tools mitigated human biases like the hungry judge effect where leniency increased after meals. Despite the ruling, questions remained about whether the public could truly understand the logic behind the numbers presented in courtrooms.

  7

• 4. Disparate Outcomes And Fairness Metrics

  ›

  A 2016 investigation by Julia Angwin at ProPublica found stark racial disparities in how COMPAS labeled defendants. The team discovered blacks were almost twice as likely as whites to be labeled higher risk without actually re-offending. Conversely, white defendants were much more likely than black defendants to be labeled lower risk yet go on to commit other crimes. Northpointe responded with a letter criticizing the methodology and stating they did not agree with the results or claims made based on that analysis. A separate group called Community Resources for Justice published a rebuttal concluding the ProPublica findings contradicted existing studies showing actuarial risk could be predicted free of bias. Later research from Williams College in 2024 showed that while use of COMPAS reduced confinement rates overall, it exacerbated differences between racial groups in Broward County.

  21012

• 5. Human Experts Versus Machine Scores

  ›

  Studies comparing COMPAS accuracy against human judgment revealed mixed results regarding predictive power. One study found individuals with little criminal justice expertise got the right answer 63 percent of the time. When answers from such individuals were pooled together, their collective accuracy rose to 67 percent. COMPAS itself achieved an accuracy rate of 65 percent in similar tests. Researchers from the University of Houston noted the software did not conform to group fairness criteria across sex-based demographic groups. Another study indicated COMPAS was somewhat more accurate than single untrained people but less accurate than groups of those same people. The data suggested machines performed better than isolated humans but struggled when compared to collaborative human decision-making processes.

  1316