AI in Law Enforcement: An Experimental Study with SFPD

• In Event: Integrating Artificial Intelligence (AI) and Randomized Controlled Trials (RCTs) in Police Work

Abstract

This article describes the results of a pilot study of artificial intelligence aided report writing by police officers in the San Francisco Police Department. The pilot study of an AI report drafting software includes 90 officers across two districts for a duration of 90 days. The software takes audio data captured by body worn cameras, transcribes and summarizes the information and generates the first draft of the incident report. Officers review and edit the draft before final submission. The present study evaluates the credibility of AI aided report against reports generated without assistance across two audiences, police administrators and the District Attorney's Office. Pilot success is defined along five metrics: draft report accuracy, report narrative quality, time savings, officer utility, and fiscal implication. Internal researchers conducted three analyses: quasi-experiment with officers treated officers in two districts and control group in two separate districts; comparative analysis by police administrators and the district attorney's officer; and cost-benefit analysis of department-wide implementation across different crime categories. Preliminary results will be discussed during the Annual Meeting.

Authors

• Carolyn M. Coles, San Francisco Police Department

• Nick Nikodinovic, San Francisco Police Department