Algorithm predicts crime a week in advance, but reveals bias in police response

predictive-policingalgorithmic-biascrime-predictionurban-mlfairness

Abstraction: ML model predicts urban crime 90% accurately while exposing police enforcement bias

Key points:

University of Chicago team built an event-level crime prediction model that forecasts violent and property crimes one week in advance with ~90% accuracy using public incident data
The model divides Chicago into ~1,000-foot spatial tiles (not traditional neighborhood boundaries) and detects time-space patterns; validated on 7 additional U.S. cities including Atlanta, LA, and Philadelphia
A second model analyzing arrest rates following crimes revealed that wealthier neighborhoods received more arrests per incident; disadvantaged neighborhoods saw arrest rates drop when overall crime stress increased, indicating resource reallocation bias
Crime in poor neighborhoods did not trigger more arrests, suggesting systematic enforcement bias rather than a neutral response to crime levels
Key methodological difference from prior "hotspot" approaches: isolates discrete event coordinates instead of epidemic/seismic spreading models, capturing city-specific social topology
Authors caution the tool should function as a simulation and policy analysis instrument — not as direct law enforcement targeting — and was funded by DARPA and the Neubauer Collegium

Connections: University Of Chicago · Darpa · Predictive Policing · Algorithmic Bias · Machine Learning