Predictive policing

Predictive policing applies statistical or machine-learning models to historical crime data to forecast where future crime is more likely, place-based predictive policing, or, in a more contested variant, which individuals are more likely to be involved in future crime, person-based predictive policing. Place-based tools are the more common form still in active use, typically outputting a ranked list of small geographic areas, similar to hotspot analysis output, updated on a rolling basis for patrol deployment.

This differs from crime mapping, which shows where crime already happened; predictive policing adds a forward-looking model, often using time-of-day and seasonal patterns, to produce a forecast. Because the models train on historical crime and arrest data, they can reproduce and reinforce bias in that data: if a neighborhood was historically over-patrolled, it generates more recorded crime, which the model then reads as elevated risk, increasing future patrol there.

Documented concerns, raised by researchers, civil liberties groups, and some vendors themselves, center on this feedback loop, on algorithmic bias issues similar to those raised about risk assessment instrument tools, and on limited public accountability where predictions shape patrol allocation without independent audit of accuracy or disparate impact by neighborhood. Several cities have discontinued place-based programs after such reviews, while others continue using them; agencies that do generally pair predictions with human review, and the effect on clearance rate is disputed in the literature.