COMPAS — an AI tool sending or keeping people in Jail

Correctional Offender Management Profiling for Alternative Sanctions- abr. as COMPAS is a case management and decision support Algorithm. It was created and is now owned by Equivant, formerly Northpointe. This software is used by U.S. courts to evaluate the probability of a defendant to re-offend. Federal courts of New York, Wisconsin, California, Florida’s Broward County, and other jurisdictions are extensively using the COMPAS software.

Should you be AWARE or BEWARE on this? Look at this picture-

What are you guessing? Concepts like False Positive Vs. False Negative- right?

Try understanding the tool summarily then-

The COMPAS software employs an algorithm to evaluate the likelihood of future criminal behaviour, which we terms commonly as recidivism. Northpointe developed risk scales to assess the likelihood of General and Violent reoffending, as well as pretrial offending. The measures in the COMPAS Practitioner’s Guide were created based on behavioral and psychological factors that are very relevant to predicting the likelihood of reoffending and criminal trajectories in the particular’s history.

Pretrial risk assessment tool: Pretrial risk refers to the likelihood of a person not showing up for court proceedings and/or engaging in new criminal activities when they are released from custody. The study that influenced the development of the scale identified “current charges, pending charges, prior arrest history, previous pretrial failure, residential stability, employment status, community ties, and substance abuse” as the most influential factors in determining pretrial risk scores.

General Recidivism Scale: The General Recidivism Scale is specifically created to forecast the likelihood of committing fresh criminal acts following release, both before and after administering the COMPAS evaluation. The scale utilizes an individual’s criminal record and connections, engagement in drug-related activities, and signs of delinquent behaviour throughout their youth.

Violent Recidivism Scale:This scale measures the likelihood of committing violent offenses again after a previous conviction. The purpose of the violent recidivism score is to forecast instances of violent crimes that may occur after an individual’s release. The scale incorporates data or indications such as an individual’s “record of violent behaviour, record of non-compliance, vocational/educational difficulties, the individual’s age at intake, and the individual’s age at first arrest.”

Julia Angwin co-authored a ProPublica study of the algorithm in 2016. The research team discovered that blacks had almost double the likelihood of being identified as a greater risk without actually re-offending, but COMPAS tends to make the reverse error with whites: they are more prone to being labeled as lower-risk yet end up committing more crimes.

Furthermore, their findings revealed that a mere 20 percent of those who were projected to engage in violent criminal activities really followed through with such actions. However, Northpointe, in a written communication, expressed disapproval of ProPublica’s approach and explicitly indicated that they do not believe the analytical findings or the statements made based on those results are reliable or reflective of the model’s outcomes.

A further research has shown that COMPAS software exhibits a higher level of accuracy compared to people with little or no criminal justice experience, yet falls short of the accuracy achieved by such groups of persons. Their findings revealed that, on average, their accuracy rate was 63 percent. However, when their responses were combined, the group’s accuracy rate increased to 67 percent. On the other hand, COMPAS has a precision rate of 65 percent. A study conducted by researchers at the University of Houston revealed that COMPAS used for predicting criminal recidivism, does not meet the standards of group fairness and generates unequal results across different demographic groups based on sex and race.

So, what’s your take on such AI & ML driven software for Indian Court Room Judgements?