A new wave of artificial intelligence tools designed to forecast crime hotspots and identify potential offenders is drawing sharp criticism from criminologists, civil liberties advocates, and computational sociologists who warn that the technology is repackaging long-discredited predictive policing methods under a more sophisticated veneer. Recent academic and policy debates have intensified through late 2024 and into 2025, as several U.S. and European police departments expand pilot programs using machine-learning systems to allocate patrols, flag individuals, and map “risk terrain.”
The renewed controversy follows the publication of multiple peer-reviewed studies and investigative reports showing that algorithmic forecasting tools continue to disproportionately target Black, Latino, and low-income neighborhoods — even when designers claim to have stripped racial variables from training data. Researchers argue that historical arrest records, the foundation of most predictive datasets, encode decades of biased enforcement, meaning the algorithms simply automate inequity at scale.
From CompStat to Machine Learning
Predictive policing is not new. Its lineage traces back to the CompStat era of the 1990s, when the New York Police Department began using statistical mapping to direct resources. By the 2010s, vendors like PredPol (now Geolitica) and Palantir were marketing place-based and person-based forecasting to dozens of departments. A 2023 investigation by The Markup found that Geolitica’s predictions were accurate less than half a percent of the time — a finding that prompted multiple cities, including Los Angeles, to abandon the systems.
Yet the industry has not retreated. Instead, it has rebranded. New entrants now market “risk terrain modeling,” “AI-augmented situational awareness,” and “behavioral threat assessment,” often integrating data from license plate readers, social media scraping, gunshot detection, and even utility records. Critics call this iteration “predictive policing 2.0” — the same logic dressed in transformer-based language and sleeker dashboards.
What the New Research Says
Criminologists publishing in journals such as Criminology & Public Policy and the Journal of Quantitative Criminology have documented that the feedback loops inherent in these systems create what scholars call “runaway selection.” When officers are repeatedly directed to the same blocks, more arrests are recorded there, which the algorithm interprets as confirmation of risk, prompting still more patrols. The result is a self-fulfilling prophecy that has little to do with actual crime distribution.
Sociologists working at the intersection of computational methods and criminal justice — a field sometimes called computational criminology — have been especially vocal. They point out that crime is dramatically underreported in many communities, and that property crime, white-collar offenses, and intimate partner violence rarely appear in the data pipelines that feed predictive systems. The American Civil Liberties Union has called for outright moratoriums, arguing that no audit framework currently in use is sufficient to catch downstream discrimination.
Departments and Lawmakers Respond
The pushback is reaching policymakers. The European Union’s AI Act, which entered staged enforcement in 2025, classifies predictive policing tools targeting individuals as “unacceptable risk” and bans them outright. In the United States, regulation remains patchwork, but Minneapolis, Oakland, and Santa Cruz have all passed local ordinances restricting algorithmic policing. Federal interest is also growing: the National Institute of Justice has funded several independent audits, and a bipartisan group of lawmakers introduced legislation last year requiring transparency for any federally funded predictive system.
Defenders of the technology, including some police chiefs and data scientists, argue that abandoning algorithmic tools entirely would discard genuine analytical gains. They contend that, properly constrained and paired with community oversight, machine learning can help departments deploy scarce resources and identify environmental risk factors — broken streetlights, abandoned lots — rather than people.
What to Watch Next
The next twelve months will likely be decisive. Expect more municipal audits, additional academic studies measuring real-world impact, and a growing push from sociologists to embed ethnographic and qualitative methods alongside computational ones. Whether “predictive policing 2.0” survives this scrutiny — or joins its predecessor in the technological graveyard — depends largely on whether departments accept independent oversight and whether legislators move from concern to enforceable rules.
For more reporting and analysis on the social sciences, technology, and the forces shaping public life, visit science.wide-ranging.com for related coverage and deeper dives.
