AI-Powered Forensic Risk Assessment Tools and Bias Debates

A new report prepared by the Center for Democracy & Technology (CDT) examined the integration of artificial intelligence and machine learning-based algorithmic risk assessment tools used in policing, pre-trial release, and recidivism predictions, drawing attention to issues of racial bias and accountability.

Courts and prison authorities have relied on simple statistical scorecard systems for risk scoring for decades; however, the federal government's artificial intelligence inventory shows that machine learning is now used in high-risk release decisions. In policing, algorithmic risk assessment encompasses statistical tools used to predict the likelihood of an individual failing to appear in court or committing future crimes, which affects pre-trial detention and prison sentences.

Pre-trial risk assessment is applied to individuals who have been arrested but not yet convicted, helping judges set bail or impose conditions of release. Recidivism prediction, on the other hand, evaluates individuals after conviction to determine supervision levels and release dates. Systems such as PATTERN, Public Safety Assessment, and COMPAS utilize inputs such as age, employment, and criminal history in this process. The Department of Justice's 2025 AI inventory classified PATTERN as a high-impact tool utilizing agent-based AI to predict recidivism.

Historically, these tools have been criticized for moderate to poor predictive accuracy and racial disparities. ProPublica's 2016 and the Department of Justice's 2021 reviews revealed that tools such as COMPAS and PATTERN over-predicted risk for Black, Hispanic, and Asian individuals. Critics argue that the introduction of machine learning does not solve these underlying data and bias problems.

Recommendations on Transparency and Civil Liberties

Machine learning models lead to losses in transparency and accountability, making it difficult for defense attorneys to audit scores due to complex structures and unstructured case notes. While the continuous retraining of AI models complicates validation studies, future real-time biometric integrations could move risk scoring outside of courtrooms. To protect civil liberties, it is recommended to disclose system names, versions, and validation dates, as well as to eliminate trade secret defenses so that defendants can examine score calculations.