What Happened
Predictive policing systems, led by PredPol (later rebranded as Geolitica), used historical crime data to predict where crimes were likely to occur, directing police patrols to specific areas. Research published in Nature Machine Intelligence and other journals demonstrated that these systems created self-reinforcing feedback loops: more police in an area leads to more arrests in that area, which generates more data suggesting that area has higher crime, which directs more police there.
Because historical arrest data disproportionately reflected policing patterns in Black and minority communities rather than actual crime rates, the algorithms systematically directed more police resources to minority neighborhoods regardless of actual crime trends.
Timeline
Predictive policing tools were adopted by dozens of US cities starting around 2012. Academic research documenting the feedback loop problem accumulated throughout 2016-2020. Following the George Floyd protests in June 2020, scrutiny intensified dramatically. The LAPD ended its predictive policing program in April 2020. Santa Cruz banned predictive policing entirely in 2020. Multiple other cities followed with cancellations and moratoriums.
Impact
The predictive policing controversy became a defining case study in algorithmic amplification of systemic bias. It demonstrated that AI systems trained on biased data do not merely reproduce bias but can amplify it through feedback loops that were not present in the original human decision-making process.
The issue extended beyond any single product. It raised fundamental questions about whether historical criminal justice data can ever serve as a fair basis for predictive systems, given that decades of discriminatory policing practices are embedded in that data.
Response
PredPol denied that its system produced biased outcomes, arguing that it predicted crime locations rather than targeting demographics. However, the company rebranded to Geolitica and eventually shut down in 2023 as city after city cancelled contracts. Civil rights organizations including the ACLU campaigned successfully against predictive policing in multiple jurisdictions.
Lessons Learned
Predictive policing demonstrated that algorithmic systems can launder human bias through the appearance of mathematical objectivity. A human police captain directing more patrols to minority neighborhoods would face obvious scrutiny. An algorithm producing the same outcome could hide behind the veneer of data-driven neutrality.
The case established that feedback loops in AI systems — where outputs influence future inputs — require special scrutiny and that domains with historically biased data require fundamentally different approaches than simply training on available historical records.