Critical Issues with the Application of Algorithms in Risk Assessment and Predictive Policing

This 24-page research paper examines the challenges of using artificial intelligence in risk assessment, with particular attention to algorithmic fairness, bias, transparency, accountability, and the competing interests of affected stakeholders. The project uses systems thinking and design thinking to examine how technical and institutional factors interact within criminal justice risk-assessment systems. The analysis identifies reinforcing and balancing feedback loops involving algorithmic bias, investment, data quality, public trust, and institutional responses. It also develops a stakeholder framework comprising four groups: Civic, Justice, Scientific, and Advocacy, each with distinct concerns related to fairness, transparency, accountability, accuracy, uncertainty, and oversight. The paper reviews and compares several approaches to addressing algorithmic fairness, including COMPAS, IBM’s AI Fairness 360, and the combination of optimal transport with conformal prediction. Each approach is evaluated in terms of its potential benefits, limitations, technical requirements, and ability to address the needs of different stakeholder groups. Based on this analysis, the project proposes three possible pathways: standardized data inventory and preprocessing, independent third-party auditing, and distributional equalization using optimal transport with conformal prediction. The research process involved literature review, systems and design-thinking exercises, stakeholder analysis, and expert discussions with Professor Richard Berk and Professor Alex Chohlas-Wood. These conversations contributed to successive revisions of the project’s central research question, moving the analysis toward the broader social, institutional, and policy trade-offs surrounding algorithmic fairness. The proposed solutions remain theoretical and were not implemented or empirically evaluated. The paper therefore presents a framework for analyzing fairness and risk in criminal justice AI systems rather than a validated intervention.