She Predicted a Murder: The Rise and Critique of Crime Prediction Systems

Not long ago in Chicago, a police predictive analytics system determined that Robert McDaniel would soon be involved in a shooting. Police began monitoring him: officers would stop by the store where he worked and visit his home. Eventually, a crime did occur—local gang members attempted to kill Robert, believing that the police attention meant he was working with law enforcement. While we can’t be sure if events unfolded exactly as McDaniel described—journalists only know the story from his account—predictive analytics systems have become a staple for American law enforcement. Police in other countries are also looking to follow the U.S. example. Dmitry Serebrennikov, a junior researcher at the Institute for the Rule of Law at the European University at St. Petersburg, explains how these systems work and why experts are raising concerns about “black box” justice.

The History of Predictive Policing

Predictive policing programs collect and analyze data on past crimes to identify potential future offenders or predict where crimes might occur. The idea is that this allows police to prevent incidents or implement preventive measures. Academics began developing such tools as early as the first half of the 20th century. For example, in 1928, Ernest Burgess, a founder of the Chicago School of Sociology, created a statistical model to predict the likelihood of recidivism after parole. However, these studies remained largely theoretical for decades.

Preventing crime and ensuring the “prosperity” of city life was a central goal of European police forces as early as the 17th century. By the 20th century, especially in English-speaking countries, this principle shifted to a “reactive policing strategy,” focusing on rapid response to law violations.

The demand for crime analysis—and thus crime prediction—emerged in the 1960s, leading to the creation of crime concentration maps and closer collaboration between law enforcement and social scientists (notably in the U.S. and UK). In the 1990s, this collaboration intensified with the rise of New Public Management, a government approach that applies private sector management models to the public sector and emphasizes data-driven decision-making. For more on this doctrine, see articles on the CompStat crime control system, which also emerged during this period.

Types of Predictive Policing Systems

Modern programs fall into two main categories, both popular in the U.S. and Western Europe:

• Person-based: Focused on identifying individuals likely to commit crimes.
• Place-based: Focused on predicting locations where crimes are likely to occur.

There are many “person-based” tools, including predicting recidivism after parole, analyzing social networks and open data to find potential offenders, and calculating bail amounts and sentences based on predicted risk. These systems emerged as police accumulated large volumes of citizen data, which became easier to collect and analyze with digital record-keeping.

Each technology sparks debate, often centered on the opacity of algorithms and concerns about privacy and data protection laws. In some countries (like Germany), the use of such algorithms is restricted.

Place-based systems are more universal, and debates about them focus on the same technical solutions. The arguments for and against these systems often apply to person-based tools as well, so we’ll focus on place-based systems below.

How Place-Based Systems Work

Developers of these systems are familiar with criminological theories like “broken windows,” routine activity, and situational crime prevention. However, in practice, police tools use models from seismology—originally designed to predict cascading earthquakes. The logic is that a sudden spike in certain crimes in an area is likely to continue for a while, so police need to respond quickly.

The main task becomes monitoring “hotspots”—areas with disproportionately high crime rates. As one early study found, 50% of police calls come from just 3% of city locations. The effectiveness of predictive tools is measured by how accurately they identify these hotspots or safe areas.

Two Industry Giants: PredPol and HunchLab

Crime prediction software evolved from flexible but complex tools to simpler, more user-friendly ones—much like the evolution of smartphone interfaces. This is illustrated by the stories of two industry giants: PredPol and HunchLab.

HunchLab, launched in 2008, was designed as a multifunctional assistant for police, offering highly customizable analysis of different crime types using various statistical models and data (like current crime rates, weather, and socioeconomic indicators). However, its first version flopped: only two out of 60 targeted police departments signed contracts.

PredPol took a radically different approach, requiring only crime incident data to generate maps highlighting areas with high predicted crime risk. Officers simply had to decide how to patrol the highlighted spots. This user-friendly interface was a hit: after launching in 2012, PredPol quickly spread across U.S. police departments.

Interestingly, one of PredPol’s creators, Jeffrey Brantingham, is a UCLA anthropology professor who studied hunter-gatherer adaptation in northern Tibet. He later applied his field knowledge to studying Los Angeles crime neighborhoods, drawing parallels between hunters searching for game and criminals seeking easy targets.

Over time, these systems took on more administrative functions, such as route planning for patrols and tracking whether officers followed their assigned routes. Police performance began to be evaluated through these systems, raising new challenges. For example, HunchLab offers four ways to set performance metrics:

• The police department sets its own criteria.
• Metrics are based on the material damage of crimes detected by officers.
• Criteria are based on average court sentences for different crimes.
• Metrics are set through public discussion.

Each approach has flaws: administrative logic can be opaque, economic calculations may ignore legal principles, court-based metrics depend on shifting criminal policy, and public opinion is highly variable. These issues are common to any system that relies on statistical evaluation.

Despite these controversies, PredPol and similar systems quickly gained popularity among law enforcement in the U.S. and abroad, thanks to claims and studies suggesting they reduce crime. For example, a randomized controlled trial in Los Angeles and Kent (UK) found that predictive models forecasted twice as many crimes as expert criminologists and reduced some crime rates by up to 7.4%. Supporters argue that these tools are more cost-effective, can reduce racial bias (since the algorithms are “color-blind”), and improve public safety and quality of life.

Criticism of Predictive Policing

Critics argue the picture isn’t so rosy. All studies confirming PredPol’s effectiveness were conducted by its developers or affiliated researchers, raising concerns about bias—similar to pharmaceutical companies funding studies on opioid safety. Other independent studies found no significant effect from new police technologies, though they often analyzed lesser-known tools.

Corporate secrecy makes predictive algorithms opaque even to the police using them, making independent evaluation nearly impossible. This opacity can also benefit police leadership, who can manipulate statistics to justify budget increases by claiming crime reductions thanks to algorithms.

For example, Memphis police reported that their system reduced crime, but later it was revealed they compared data to a year with an unusual crime spike. A five-year review showed a more complex reality: while overall crime dropped 8% (with unclear credit to predictive analytics), violent crime actually rose 14%.

Another concern is that algorithms trained on historical data may inherit human biases—especially against minority groups (notably African Americans in the U.S.). This can create feedback loops: more data from high-risk areas leads to more police patrols, which in turn generates more crime reports (since officers need to meet quotas), perpetuating the cycle. MIT professor Gary Marx described this as “categorical suspicion”—officers treating all residents of flagged neighborhoods as suspects.

Similar issues exist in person-based systems. A 2018 study compared predictions from the COMPAS system with crowdsourced judgments about whether inmates would reoffend. Untrained people, with less data than the algorithm, performed comparably—and also showed racial bias.

Accuracy: percentage of correct recidivism predictions; False positive: percentage of false positives; False negative: percentage of false negatives (recidivism occurred but was not predicted). Human: human predictions; COMPAS: program predictions. White bars: white suspects. Black bars: African American suspects.

Testing for Bias

One of the most famous demonstrations of algorithmic bias was conducted by Kristian Lum and William Isaac, who simulated police work after implementing predictive policing with racially biased initial data. Using Oakland, California drug arrest data—where drug activity is evenly distributed but arrests are concentrated in African American neighborhoods—they recreated PredPol’s algorithms. The result: the system disproportionately flagged nonwhite neighborhoods as high-risk, recommending police patrol these areas twice as often as others.

PredPol’s creators responded to criticism, acknowledging the system’s imperfections. They analyzed their own experiment, alternating between algorithm- and expert-generated hotspot data for police, and found no significant difference in time spent in hotspots or number of arrests. However, this doesn’t address whether predictive systems perpetuate the racial biases of the humans they’re meant to replace.

PredPol’s developers admit that police bias can influence technology. Co-author Jeffrey Brantingham used simulations to test how much initial data distortion (bias against minorities) would significantly affect predictions, concluding that a 5% bias threshold was critical. The extent of real-world data distortion remains unknown.

The Future of Predictive Policing

It’s hard to predict how public consensus on predictive analytics will evolve. Police remain enthusiastic, and developers continue to refine both evaluation methods and system design. Interest in predictive analytics is growing worldwide—not just in the U.S. and China (known for its algorithmic governance experiments), but also in the UK, Germany, Poland, and other countries. Recently, Russia’s Ministry of Internal Affairs commissioned a study on using machine learning to identify serial crimes and offenses. While predictive technologies are still emerging in Russia, there is a clear trend toward their adoption by law enforcement.

Finally, it’s worth noting that discussions about predictive analytics in policing often overlook how officers themselves perceive these systems and whether they actually influence decision-making. Police may treat expensive analytics systems like weather forecasts—useful but not critical. But that’s a topic for another study.