Predictive Policing

Predictive policing is the deployment of automated decision-making systems in law enforcement contexts to allocate police resources, identify persons of interest, or flag locations for increased patrol based on statistical predictions of criminal activity. The central claim is that algorithmic prediction can be more accurate or less biased than officer discretion. The empirical record is more uncomfortable: predictive policing systems trained on historical arrest data reproduce the enforcement patterns that generated the data — patterns that reflect where police already patrol, not necessarily where crime occurs. This is feedback loop amplification masquerading as prediction. A system trained on arrests in heavily policed neighborhoods predicts more criminal activity in those neighborhoods, which increases patrol, which increases arrests, which trains the next version of the model. The system is not detecting crime. It is detecting its own prior deployments.