What Works in American Policing Part 6—Predictive Policing: Promise, Pitfalls, and the Limits of Algorithmic Forecasting

This is the sixth in a seven-part series examining the policing strategies that shape public safety in America, including what the evidence supports, what has fallen short, and what policymakers and practitioners should prioritize going forward.

Predictive policing is among the most debated, and most frequently misunderstood, strategies in this series. Crime is not randomly distributed but concentrated in particular places and among a small share of people, so data can help police anticipate where harm is most likely and focus their efforts before it occurs. Predictive policing is the application of analytical techniques to identify likely targets for police intervention—places, times, or people. In its most disciplined and transparent forms, predictive policing is a natural extension of hot spots policing, which carries one of the strongest evidence bases in criminology, and it offers real potential to help agencies use scarce resources more efficiently. In other applications, the results have been more mixed, and the technology often used in predictive policing has raised legitimate questions about bias, transparency, and civil liberties. The most balanced view falls between the optimism of its advocates and the caution of its critics. Predictive policing holds promise, but realizing it depends on using these tools carefully, confirming that they actually work, and pairing them with the right safeguards.

Where Predictive Policing Came From

The term gained traction in the late 2000s, when then-Los Angeles Police Department (LAPD) Chief William Bratton began working with federal officials on a data-driven approach to forecasting crime. In 2009, the National Institute of Justice funded seven agencies to develop predictive models, and in 2011 it backed implementations in Chicago and Shreveport, Louisiana, using the RAND Corporation to evaluate them. A 2013 RAND reference guide became the field’s foundational text, sorting predictive methods into four categories: predicting where and when crimes will occur, predicting who will offend, predicting who will be victimized, and identifying who committed crimes that have already occurred. That last one is less a forecast than an investigative aid, using analytics to point toward a likely suspect in an unsolved case. In practice, two broad families emerged—place-based prediction, which forecasts locations, and person-based prediction, which forecasts individuals. The appeal of both was rooted in the hope that an algorithm could uncover patterns that human analysts might miss and allocate resources more objectively than tradition or instinct.

Place-Based Prediction: Promise and Mixed Results

Place-based prediction builds on the same idea behind hot spot policing, which is that crime clusters in a small number of places. Its most transparent and theory-driven form is risk terrain modeling (RTM), which forecasts risk by mapping the environmental features, such as bus stops, liquor stores, and vacant properties, that tend to breed crime. Unlike ordinary hot spot mapping, which plots where past crimes occurred, RTM forecasts risk from these underlying environmental conditions so it can flag a location before a crime is recorded there. A systematic review found that RTM reliably identifies high-risk places across a range of crime types, and evaluations across several cities, including Chicago, IL, Colorado Springs, CO, Glendale, AZ, Kansas City, MO, and Newark, NJ, found that directing tailored interventions to RTM-identified locations reduced crime. Because the method is open and explainable, it sidesteps much of the criticism aimed at opaque commercial tools.

The better-known commercial product was PredPol, which grew out of an LAPD–University of California, Los Angeles (UCLA) collaboration around 2010 and was later rebranded as Geolitica. It borrowed a model originally built to forecast earthquake aftershocks, applying it to predict property crime within small boxes on a map, each roughly 500 feet across, and it became the most widely used predictive policing software in the country. Early adopters reported encouraging numbers: After the Atlanta Police Department piloted PredPol in two zones in 2013, aggregate crime fell 8 and 9 percent in the treated areas while rising in the zones without it, and the department expanded it citywide. Other departments reported similar before-and-after declines. The most favorable finding came from a 2015 study reporting that the software outperformed the hot spot maps produced by human crime analysts and modestly reduced crime in field trials with the LAPD and Kent (UK) police—though several of that study’s authors held a financial stake in PredPol, a conflict that tempers the result.

But the more rigorous the test, the more modest the findings. RAND’s randomized evaluation in Shreveport, Louisiana—one of the first controlled trials of a place-based predictive model—found no meaningful drop in property crime compared with similar areas that did not use it. (Shreveport’s program, called Predictive Intelligence Led Operational Targeting (PILOT), used a statistical model the department built in-house, neither RTM nor a commercial product.) LAPD’s inspector general likewise concluded there was insufficient data to show that PredPol reduced crime, and the department dropped it in 2020; its person-focused LASER prediction program had also been shut down in 2019. A 2023 analysis of roughly 23,600 predictions from Geolitica, PredPol’s renamed successor, for Plainfield, New Jersey, found they almost never matched reported crime. Smaller agencies, such as Rio Rancho, New Mexico, concluded the software told them little they did not already know. A 2024 systematic review of place-based prediction screened 161 studies but found that only six met the strongest evidence standard—randomized, field-tested, and accounting for crime displacement. Despite enormous interest, rigorous independent proof of effectiveness remains scarce across place-based prediction as a whole, not just commercial products.  

Person-Based Prediction: The Targeting Logic, Done Well and Done Badly

Person-based prediction rests on the premise that a small share of people drive a disproportionate amount of serious violence—but its track record relies almost entirely on how agencies act on the prediction. Done well, the logic is the foundation of focused deterrence, which identifies the highest-risk individuals, pairing a direct warning with offers of social services and community support. That approach has strong evidence behind it as a systematic review of two dozen evaluations found that most focused deterrence programs produced clear reductions in violence, often substantial ones.

Done badly, the same logic creates a secret watchlist. Chicago’s Strategic Subject List (SSL)—also known as, the “heat list”—launched in 2012 and assigned residents a score from 0 to 500 reflecting their predicted likelihood of being a “party to violence,” either as a shooter or a victim. By 2016, more than 280,000 Chicagoans had received a score high enough to put them on the department’s radar. RAND’s evaluation of the first version found that the list did not reduce gun violence. Being on it was associated with a higher likelihood of arrest, but not with a lower likelihood of being shot. Cases like that of Robert McDaniel—a man with a thin record who was visited and warned by police because an algorithm had flagged him—crystallized the due process and stigma concerns. Chicago decommissioned the program in 2019 because the city’s inspector general cited unreliable scores, inadequately trained personnel, weak access controls, and interventions that attached consequences to arrests that never resulted in conviction.

Florida offered an even starker warning. Beginning in 2011, the Pasco County Sheriff’s Office built an “intelligence-led policing” program that used an algorithm to flag residents—including minors—as likely future offenders, then dispatched deputies to make repeated, often nighttime, visits to those individuals and their families. An investigation documented thousands of such “prolific offender” checks, with former deputies describing an internal directive from the sheriff’s office leadership to pressure targets until they moved away or sued.

In 2024, the office settled a federal civil-rights suit, agreeing to pay $105,000 to four targeted residents and to permanently end the program. In the settlement, the office acknowledged that the program had violated residents’ First, Fourth, and Fourteenth Amendment rights, and it is now barred from running anything similar. The contrast is instructive as the difference between focused deterrence and the Pasco County program was not the underlying data but transparency, due process, and whether the intervention offered help or simply applied pressure.

What the Evidence Shows

Taken together, the record supports a measured conclusion. The approaches with the strongest evidence—risk terrain modeling on the place side, focused deterrence on the person side—are also the most transparent. The weakest results come from opaque commercial algorithms and enforcement-only watchlists. As participants in a 2025 National Academies workshop emphasized, effective predictive policing requires both accurate predictions and well-implemented responses; rigorous, independent studies are still needed to confirm that the commercial tools work. Notably, where place-based prediction has worked, it has largely reproduced the logic of hot spots policing—putting officers where crime concentrates—making it difficult to isolate what the algorithm adds beyond a well-drawn pin map. That is a real limitation, but it is not the whole story. The transparent forms that do work succeed precisely because they identify the environmental conditions driving risk, which lets agencies design targeted interventions rather than simply flood a hot spot with patrols. The promise, then, lies less in out-predicting a skilled crime analyst than in explaining why a place is dangerous and what to do about it—and for the proprietary tools, that added value remains unproven.

The Data and Bias Problem

The central limitation is that predictive systems learn from historical police data, and that data records where police have looked, not where crime objectively is. An analysis documented how data produced during periods of flawed, racially skewed, or unlawful policing becomes “dirty data” that yields flawed predictions. The danger is a self-reinforcing cycle because an algorithm directs officers to a neighborhood, their presence produces more stops and arrests there, that new activity flows back into the model, and the location stays “hot” regardless of the level of underlying crime.

In an influential simulation, researchers showed that feeding Oakland’s drug-arrest records into a PredPol-style algorithm would have concentrated patrols in predominantly Black neighborhoods, not necessarily because more crime occurred there, but because that is where past enforcement had been focused. None of this is a reason to abandon the technology, but it is a reason to insist on transparency and regular auditing—safeguards that open, explainable tools like RTM have and that closed, proprietary ones generally do not.

Constitutional Boundaries and Accountability

Predictive policing also raises distinct Fourth Amendment concerns. Legal scholars have warned that algorithmic prediction can transform reasonable suspicion from a protection against unreasonable stops into a justification for them. For example, when a computer-generated flag is treated as grounds to detain someone absent any observed wrongdoing, it stretches the definition of reasonable suspicion and erodes the specific basis the Fourth Amendment requires. Courts are only beginning to grapple with these questions, as the U.S. Court of Appeals for the Fourth Circuit did when it confronted a prediction-driven stop in United States v. Curry. Transparency deficits compound the problem: In Los Angeles, basic facts about the programs surfaced only after years of public-records litigation, and Chicago declined to disclose the variables driving its scores.

Policymakers have started to respond. Santa Cruz, California, the city where PredPol was founded, became the first in the nation to ban predictive policing in 2020. The European Union’s AI Act, which took effect in February 2025, bans predictive policing systems that forecast criminal behavior by profiling individuals. And in 2024, some members of Congress urged the Department of Justice (DOJ) to halt grant funding for these tools until recipients could demonstrate they would not be used in discriminatory ways. The accountability guardrails that should govern any such system—data transparency, independent bias audits, documented criteria for who and where gets flagged, retention limits, and external oversight—need to be built in before deployment, not retrofitted after harm occurs.

Where It Stands Now

Standalone predictive policing has largely receded, but the underlying impulse has migrated into broader artificial intelligence (AI)-driven police technology. Some of that evolution points toward more disciplined uses. For example, the New York Police Department’s “Patternizr” is an in-house tool that helps analysts link related burglaries, robberies, and grand larcenies across precincts, and it is confined to connecting reported crimes rather than forecasting who will offend. But the broader trend is toward consolidation and opacity. SoundThinking, the company formerly known as ShotSpotter, absorbed Geolitica’s clients and now folds prediction into a wider suite of products, while real-time crime centers, AI-equipped body cameras, and facial recognition increasingly blend place-and person-based inference. The branding has shifted from “predicting crime” to “data fusion” and “real-time intelligence,” but the civil liberties questions remain.

Assessment

Predictive policing is promising, but not yet proven as a standalone investment. Its premise is sound, and its most transparent forms(risk terrain modeling and focused deterrence) have credible evidence behind them and fit naturally with the proven, place-based work at the heart of modern policing. For understaffed agencies trying to stretch limited patrol time, a tool that helps target scarce resources has obvious appeal. But the closed, proprietary products that dominate the commercial market have a more mixed independent track record, and both place-and person-based tools carry real constitutional risks when a computer’s predictions take the place of an officer’s own judgement, or when enforcement piles up in communities the data already overrepresents.

Predictive policing belongs in the “adopt with caution” category. Agencies should deploy it only with independent validation, transparency, bias auditing, and constitutional guardrails favoring open, explainable models over proprietary ones. Pairing predictions with well-designed interventions is superior to enforcement alone, and treating technology as a complement to proven strategies rather than a substitute ensures it is constitutional. Above all, agencies should keep a human in the loop, treating a prediction as a tool to inform an officer’s judgment, never a replacement for it. The fiscal logic is the same one that runs through this series: taxpayers should not be asked to fund these tools without independent evidence that they work, and the agencies that can meet that burden of proof transparently are the ones most likely to see real returns for the communities they serve.

bool(true)

bool(true)

string(3) "100"

Next in this series: Part 7—Conclusion: No Single Strategy Makes Public Safety Work