Artificial Intelligence and Recidivism: The Use of Risk Assessment as Prediction Instruments in Criminal Sentencings

The use of recidivism risk assessment instruments throughout different stages of the criminal justice process continues to grow. Much of the existing literature fails to focus on the algorithm's consistency in decision-making and implementation. This thesis analyzes the effectiveness and uniformity of COMPAS recidivism risk assessment instruments in the sentencing process. I use quantitative data from the Wisconsin Department of Corrections and qualitative interviews with experts to better analyze the efficacy and consistency of the algorithms from a practical standpoint. I find that, while the COMPAS risk assessment instrument validly determines individuals' risk of recidivism on a general level, COMPAS risk level may not correspond to one's sentence length. This finding indicates that judges deviate from COMPAS risk levels during the sentencing process. This discrepancy points to a need for standardized training for state-operated risk assessment instruments, rigorous external validation studies, and protocols encouraging judges to articulate their decision-making processes. These measures aim to enhance the instrument's practical efficacy and fairness in sentencing decisions.