AI Parole Risk Assessment Tools Showed Racial Bias Against Black Defendants Across Multiple US States

The National Institute of Justice released a comprehensive 2023 study examining algorithmic risk assessment tools deployed across more than 40 US states for criminal justice decisions. The research analyzed five major systems: the Public Safety Assessment (PSA), Ohio Risk Assessment System (ORAS), Level of Service Inventory-Revised (LSI-R), PATTERN, and COMPAS. These tools are used by judges, parole boards, and corrections officials to make high-stakes decisions about pre-trial detention, sentencing, and parole eligibility, affecting an estimated 500,000 individuals annually. The study revealed systematic racial disparities across all examined tools. Black defendants consistently received higher risk scores than white defendants with similar criminal histories and case characteristics. The PSA, used in over 35 jurisdictions including New Jersey and Kentucky, showed Black defendants were 31% more likely to be classified as high-risk for failure to appear and 39% more likely to be flagged for new criminal activity. The PATTERN system, used by federal prisons for recidivism assessment, exhibited a 15-point average score difference between Black and white defendants. Similar patterns emerged in state-specific tools like Ohio's ORAS system. The algorithmic bias traced to training data that incorporated decades of racially disparate policing, prosecuting, and sentencing practices. Variables like zip code, employment history, and family structure served as proxies for race, while the models learned patterns from historical data where Black defendants faced harsher treatment at every stage of the criminal justice process. The feedback loop effect meant that higher initial risk scores led to harsher treatment, generating data that reinforced the bias in future algorithmic updates. The disparate impact extended beyond individual cases to broader criminal justice outcomes. Counties using these tools showed increased racial gaps in pre-trial detention rates, with Black defendants held at 25% higher rates than comparable white defendants. Parole denial rates for Black inmates increased by 18% in jurisdictions implementing algorithmic risk assessment compared to traditional review processes. The bias also influenced judicial decision-making, with judges following algorithmic recommendations 85% of the time despite being instructed to use the scores as just one factor among many.