Abstract

Crime risk prediction and predictive policing can lead to safer communities, by focusing on crime hotspots. Yet predictive tools should be reliable, and their outputs should be valid, especially across diverse cultures. Machine learning methods in policing systems are topical as they seem to be causing unintended consequences that exacerbate social injustice. Research into machine learning algorithm bias is prevalent, but bias, as it relates to predictive policing, is limited. In this paper, we summarise the findings of nascent scholarship on the topic of bias in predictive policing. The unique contribution of this paper is in the use of a typical police prediction modelling process to unpack how and why such bias can creep into algorithms that have high predictive accuracy. Our research finds that especially when resources are limited, trust in machine learning outputs is elevated; systemic bias of preceding assumptions may replicate. Recommendations include a call for human oversight in machine learning methods with sensitive applications such as automated crime prediction methods. Routine reviews of prediction outputs can ensure unwarranted community
targeting is not magnified.

Keywords:
predictive policing, ethnic identity, racial profiling, crime risk prediction, algorithm bias, demography.