“Predictive policing does not predict the future it predicts that the past will repeat itself, which guarantees that whoever was over-policed yesterday will be over-policed tomorrow.” Predictive policing refers to the use of AI and statistical models by law enforcement agencies to forecast the location and timing of future crimes (place-based prediction) or to identify individuals deemed at elevated risk of committing or being victims of crime (person-based prediction), with the goal of enabling pre-emptive intervention.
Executive Summary
Predictive policing has been deployed by law enforcement agencies in the United States (ShotSpotter, PredPol/Geolitica, LASER), United Kingdom, Netherlands, and across several emerging market jurisdictions. The technology’s promise more efficient crime prevention with limited resources has collided with documented evidence of racial bias amplification, civil liberties violations, and weak accuracy performance. Los Angeles discontinued its PredPol and CalGang systems in 2020 after an ACLU report documented systemic bias. Santa Cruz (2020), New Orleans (2022), and Colorado (2023) enacted bans. The EU AI Act classifies predictive policing systems in law enforcement as prohibited applications when they target specific individuals. The governance consensus is hardening: place-based prediction faces fewer objections than person-based prediction, where the civil rights and accuracy concerns are most acute.
The Strategic Mechanism
- Place-based prediction (hotspot policing): Statistical models analyze historical crime data to identify geographic areas with elevated predicted crime risk, directing patrol resources toward those locations. Less controversial than person-based prediction, with some evidence of efficacy when combined with community policing.
- Person-based prediction: Algorithms assign risk scores to specific individuals based on criminal history, social network associations, demographic characteristics, or behavioral indicators. The COMPAS recidivism tool and related systems have been extensively studied and found to exhibit racial bias.
- Network analysis: AI mapping of social networks among known offenders to identify individuals connected to high-risk actors, used by the Los Angeles LASER program (discontinued 2020) and similar systems.
- Feedback loop problem: Predictive systems direct police toward historically over-policed communities, generating more arrests in those communities, which train the next model iteration to predict more crime there amplifying rather than correcting initial bias.
- Pre-crime liability: Person-based prediction that results in police contact, surveillance, or intervention against individuals who have committed no offense creates constitutional and civil liability exposure that has driven legislative bans in multiple jurisdictions.
Market & Policy Impact
- ShotSpotter (now SoundThinking) acoustic gunshot detection system, deployed in 150+ US cities, faced ACLU litigation and city contract cancellations in Chicago, San Francisco, and Oakland after studies found high false-positive rates and discriminatory deployment patterns.
- The EU AI Act explicitly prohibits “AI systems used by or on behalf of law enforcement authorities for making individual risk assessments of natural persons in order to assess the risk of a natural person for offending or reoffending” in its list of unacceptable-risk applications.
- The RAND Corporation’s 2022 meta-analysis of place-based predictive policing found weak and inconsistent evidence of crime reduction effects, with most studies showing effects indistinguishable from conventional hotspot policing that doesn’t require AI systems.
- ProPublica’s 2016 analysis of COMPAS recidivism scores found Black defendants were nearly twice as likely as white defendants to be falsely flagged as high risk generating the most influential algorithmic accountability investigation in US journalism history.
- The Dutch National Police’s SyRI social welfare fraud prediction system was ruled illegal by the Hague District Court in February 2020 the first European court ruling prohibiting a government algorithmic decision system on human rights grounds, predating the EU AI Act.
Modern Case Study: Los Angeles LASER Discontinuation and the Accountability Turn, 2019-2020
The Los Angeles Police Department’s LASER (Law Enforcement Assisted Diversion) predictive program assigned point scores to individuals based on criminal history, gang associations, and police contact frequency, creating a list of individuals targeted for enhanced surveillance and intervention. An ACLU of Southern California investigation in 2019 found that Black and Latino residents were disproportionately flagged, that the system lacked meaningful oversight mechanisms, and that officers had limited training in interpreting or challenging algorithmic scores. Following the report and subsequent community organizing, LAPD Inspector General Mark Smith issued a review finding the program lacked transparency, evidence of effectiveness, and adequate civil liberties protections. LAPD discontinued LASER in April 2020. The case established the governance template for predictive policing accountability: independent audit, algorithmic transparency, and community consultation are minimum requirements for legitimacy standards that most deployed systems had not met and that drove the subsequent wave of legislative bans.