Unmasking Algorithmic Bias in Predictive Policing: A GAN-Based Simulation Framework with Multi-City Temporal Analysis

Predictive policing systems amplify racial bias dramatically: The research found extreme disparities in how algorithmic systems direct police resources, with Baltimore's detected crime mode showing a Disparate Impact Ratio as high as 15.7 in 2019—meaning Black residents were detected at rates up to 15 times higher than expected based on actual crime occurrence.

Bias persists across multiple cities and metrics: Analysis of over 378,000 crime records from Baltimore and Chicago revealed consistent inequality measures (Gini coefficients of 0.43-0.62), indicating the problem is not isolated to one city or time period but represents a systemic issue in how these algorithms operate.

Technical fixes alone cannot solve the problem: The study tested a debiasing approach using advanced machine learning (Conditional Tabular GAN), which only partially reduced disparities. The authors conclude that algorithmic adjustments cannot eliminate structural bias without accompanying policy changes, suggesting that technical solutions must be paired with institutional reform.

Neighborhood racial composition strongly predicts police contact: The research found a strong correlation (0.83) between the percentage of Black residents in a neighborhood and detection likelihood, demonstrating that algorithmic bias is deeply tied to historical patterns of over-policing in communities of color rather than actual differences in crime rates.