Algorithmic Bias and Predictive Policing in U.S. Cities
Predictive policing tools are often framed as neutral optimizers of limited public‑safety resources, but their inputs and operational rules reflect historical enforcement patterns and procurement choices that can reproduce and amplify inequities. This mixed‑methods study synthesizes deployment logs, geocoded incident and arrest records, and interviews with officers and community stakeholders across three mid‑sized U.S. cities to examine how algorithmic outputs are interpreted and acted upon. Quantitative analyses use correlation and difference‑in‑differences designs to measure alignment between model risk scores and historical enforcement, and to estimate changes in patrol hours and arrest rates after deployment. Qualitative interviews reveal operational heuristics—deference to scores, simplified hotspot tactics, and feedback loops from enforcement to training data—that explain how algorithmic signals become self‑reinforcing. The paper concludes that technical adjustments alone are insufficient; procurement transparency, independent audits, and community governance are necessary to mitigate harms while preserving legitimate public‑safety objectives.
Introduction
Municipalities increasingly adopt algorithmic systems to allocate patrols and prioritize investigations. While vendors emphasize efficiency gains, the data used to train these systems—historical arrests, calls for service, and complaint logs—reflect prior enforcement decisions and structural bias. This paper examines how model design, procurement constraints, and on‑the‑ground deployment practices shape outcomes and whether these systems reproduce or mitigate disparities.
Methods
We used a mixed‑methods approach. Quantitatively, we linked anonymized deployment logs and model outputs to geocoded arrest and call‑for‑service records (2014–2023) and estimated correlations and DiD models to measure changes in patrol hours and arrest rates. Sensitivity analyses tested alternative input weightings and counterfactual deployments. Qualitatively, 28 semi‑structured interviews with officers, supervisors, and community leaders were coded thematically to surface operational mechanisms and community impacts. Data access and human subjects protections were managed through institutional review and data‑use agreements.
Results
Model outputs were strongly correlated with historical arrest density (r ≈ 0.78). DiD estimates show that tracts in the top historical arrest quintile received significant increases in patrol hours post‑deployment relative to matched controls, with modest increases in arrest rates after controlling for reported incidents. Qualitative data reveal that officers often treat model outputs as authoritative guidance, supervisors apply simple hotspot heuristics, and increased enforcement generates additional data that feed back into retraining cycles—creating self‑reinforcing surveillance loops.
Discussion
Predictive policing systems can entrench historical enforcement patterns unless procurement, transparency, and governance are reformed. Technical fixes alone are insufficient; procurement clauses requiring vendor disclosure, independent fairness audits, operational limits on automated decision authority, and community oversight boards are necessary to align algorithmic tools with public‑safety and civil‑rights goals. Future work should evaluate the efficacy of these governance interventions in reducing disparities.
References
- Brayne S. Predict and Surveil. Oxford University Press; 2020.
- Benjamin R. Race After Technology. Polity; 2019.
- Ensign D, Friedler SA, Neville S, Scheidegger C. Runaway feedback loops in predictive policing. FAT* 2018.
- Lum K, Isaac W. To predict and serve? Significance. 2016.
- Barocas S, Selbst AD. Big data’s disparate impact. California Law Review. 2016.