Algorithmic crime maps versus human analysts: an actual randomized patrol comparator, not an agent trial

Sources: Mohler et al., JASA 2015, original randomized field trial; Brantingham, Valasik and Mohler, 2018, reanalysis of arrest patterns from the same trial (authors involved in building/testing the algorithm); US National Institute of Justice program assessment. Provides Dru’s demanded human comparator for a non-agentic earlier system, not evidence about mass deployed LLM investigators.

Design and effect

LAPD crime analysts and a short-term ETAS place-based statistical model made separate maps of twenty 150 × 150 m boxes per shift. After both maps were made, deployment of one set was randomized by day in each of three LA divisions during 2011–13 (510 experiment days in all); patrol officers did not know their map’s source. Both outputs sent human officers on patrol: algorithm did not observe people, join records or order arrests autonomously. Silent tests in LA and Kent found higher fraction of subsequent reported crimes inside algorithm-picked areas; in three LA deployed divisions, combined fraction was 4.2% algorithm versus 2.2% analyst for targeted crimes. Authors estimate 7.4% fewer reported crimes per 1,000 police minutes associated with algorithm-targeted patrol compared with nonsignificant analyst-patrol slope. This is a patrol-time response estimate under particular outcome and assumptions, not a citywide 7.4% crime decline.

Using those same LA trial days, authors found no significant change in the overall fraction of arrests across race/ethnicity (Cochran–Mantel–Haenszel p≈.70). Arrests inside the selected boxes roughly doubled under the algorithm but on more-crime boxes: arrests per reported crime were lower or not significantly different across three divisions; division-wide total arrests unchanged or marginally lower. This is not proof of no bias: data are from recorded crimes and arrests, not all stops/searches, complaints, wrongful targeting or comparisons with the distribution of unreported victimization; power for group-specific effects and generality beyond 3 divisions remain uncertain. Study authors include system designers, and human analyst methods/quality vary.

What the comparator teaches without overclaim

Do not take ‘automation magnifies state power’ to imply algorithmic output always performs worse than human practice on patrol targeting, nor take ‘no detectable extra race bias in this old case’ as a clean bill for current agentic surveillance. Police officer action and reporting data mediate both conditions; algorithmic redirection depends on data input and patrol budget. A proper agent trial would additionally randomize read permissions and outbound action authority, measure coverage, false accusation, appeal, biased human alternatives and longer-run feedback.