Predictive policing

Based on Wikipedia: Predictive policing

The Algorithm That Knows Where Crime Will Strike

In 2011, the city of Santa Cruz, California tried something that sounded like science fiction. Instead of sending police officers on random patrols or waiting for 911 calls, they fed years of crime data into a computer program and asked it a simple question: where will the next burglary happen?

The results seemed remarkable. Burglaries dropped by nearly twenty percent in six months.

Nine years later, that same city became the first in America to ban the technology entirely.

This is the strange and contested story of predictive policing—a set of techniques that promises to forecast crime before it happens, and the growing realization that such predictions might be creating more problems than they solve.

What Predictive Policing Actually Does

At its core, predictive policing is mathematics applied to law enforcement. Police departments collect vast amounts of data: the times crimes occur, their locations, the types of offenses, weather conditions, local events, and countless other factors. Algorithms then sift through this information looking for patterns that human analysts might miss.

The RAND Corporation, a nonprofit research organization, identified four distinct flavors of these predictive methods. The first predicts crimes themselves—essentially generating maps showing where offenses are likely to occur next. The second attempts to predict offenders, flagging individuals whose behavior patterns suggest they might commit crimes. The third works backwards from crimes to identify perpetrators. And the fourth, perhaps most controversially, tries to predict who will become victims.

Think of it like weather forecasting for crime. Just as meteorologists look at pressure systems, humidity, and historical patterns to predict tomorrow's storm, predictive policing systems analyze past criminal activity to forecast future trouble spots.

The appeal is obvious. Police departments face a fundamental resource constraint: they cannot be everywhere at once. If an algorithm can tell you that car break-ins spike between 2 and 4 AM near a particular intersection on Thursdays, you can position officers there during those hours. In theory, this prevents crimes that would otherwise occur.

The Mechanics of Prediction

How does a computer actually predict where crime will happen?

The systems work by detecting signals and patterns in crime reports. They track when shootings occur, where cars get broken into, which blocks see the most robberies, and at what times. They factor in variables that might seem unrelated at first glance—concert schedules, paydays, even phases of the moon have been tried.

Some cities have gotten creative with their data sources. Chicago, for instance, combines population mapping with crime statistics to identify emerging patterns. Other departments have experimented with acoustic sensors that detect and locate gunfire in real time, feeding that information back into their predictive models.

The promise is speed and accuracy. A human analyst might spend weeks combing through reports to notice that a particular neighborhood sees more burglaries when a nearby factory changes shifts. An algorithm can identify such correlations in minutes and continuously update its predictions as new data arrives.

But here's the crucial part that often gets glossed over: a prediction is useless without a response. These systems must be coupled with prevention strategies—typically sending officers to the predicted time and place of a potential crime. The algorithm generates possibilities; humans must still decide what to do about them.

From Baghdad to Beijing

The intellectual roots of predictive policing reach into unexpected places.

One thread traces back to the streets of Iraq in 2003. After major combat operations ended, Improvised Explosive Devices—roadside bombs known as IEDs—became the insurgency's weapon of choice. American military analysts developed techniques to predict where these devices would appear, identifying what they called "actionable hot spots"—zones with high activity levels that were nonetheless too vast for comprehensive surveillance.

The challenge was familiar to any police strategist: too much territory, too few resources, and adversaries who adapted their behavior based on patrol patterns. The mathematical tools developed for counterinsurgency would eventually find their way into domestic law enforcement.

Another significant thread originates in China, where the concept has been woven into a much broader vision of social control. In 2016, Chinese leader Xi Jinping announced an agenda called "social governance"—the use of extensive information systems to promote what the government describes as a harmonious and prosperous society.

The most visible manifestation is China's social credit system, which uses massive data collection to digitize identities and quantify trustworthiness. Nothing comparable exists in Western democracies. The police component involves converting what officials call "intelligence-led policing"—effectively using information—into the "informatization" of policing, which means embedding information technology into every aspect of law enforcement.

China's Police Geographical Information System, or PGIS, began in the 1970s as a tool for urban planning. By the mid-1990s, it had migrated into public security work. Today it handles spatial queries and hot spot mapping, with crime trajectory analysis and prediction still in development. The system is being upgraded to cloud-based architecture with plans for integrating multiple data sources into real-time visualization.

Between 2015 and 2018, several Chinese regions launched predictive policing initiatives. Zhejiang and Guangdong provinces built systems to predict and prevent telecommunications fraud through real-time surveillance of suspicious online activity, partnering with private companies like Alibaba to identify potential suspects. In 2018, police in Zhongshan made over nine thousand warning calls to potential fraud victims and intercepted more than thirteen thousand suspicious phone calls.

More troubling applications have emerged. Chinese authorities have used predictive policing to identify and target people for detention in Xinjiang internment camps, part of a broader campaign against the Uyghur Muslim minority. The Integrated Joint Operations Platform, operated by the Central Political and Legal Affairs Commission, reportedly flags individuals for "re-education" based on algorithmic assessments of their behavior.

The American Experiment

In the United States, predictive policing has spread across states including California, Washington, South Carolina, Alabama, Arizona, Tennessee, New York, and Illinois. The approaches vary considerably.

New York City's police department implemented a program called Patternizr, designed to help officers identify commonalities in crimes committed by the same offenders. When a detective is investigating a robbery, the system generates possible "patterns"—other crimes that might be connected. The officer manually reviews these suggestions to determine if they warrant deeper investigation. It's less about predicting future crime than connecting past ones.

The Santa Cruz experiment with PredPol software represented a different approach: pure location-based prediction. The system would generate daily forecasts identifying five-hundred-foot-by-five-hundred-foot boxes where property crimes were most likely to occur. Officers would then spend extra time in those zones.

The early results looked promising. That twenty percent reduction in burglaries seemed to validate the entire concept.

But as the years passed, concerns accumulated.

The Mathematics of Bias

Here's the problem that haunts every predictive policing system: the algorithm can only learn from the data it's fed. And the data comes from past policing.

If a neighborhood has historically been subjected to more police scrutiny—more stops, more searches, more arrests—it will appear more criminal in the data, even if underlying crime rates are similar to less-policed areas. The algorithm then recommends more policing of that neighborhood, generating more data showing it as criminal. It's a feedback loop that can amplify existing biases rather than correct them.

Critics argue this isn't a bug but an inherent feature. Communities of color and low-income neighborhoods have historically experienced more intensive policing. When you train a system on that history, you're teaching it to perpetuate those patterns.

There's another, more mathematically subtle problem. Even with perfectly unbiased data—even if you could somehow strip away every historical inequity—differential surveillance rates create their own distortions.

Here's why. Every prediction has some probability of being wrong. If you're monitoring one neighborhood four times as intensively as another, you're not just four times as likely to catch crimes there—you're many times more likely to generate false alerts. The math shows that such neighborhoods can see over twenty times more false positives, not because of higher crime rates, but because of how probabilities compound at scale.

Worse, these systems appear to hit critical thresholds beyond which false alerts become essentially certain in heavily-monitored areas. This suggests that discriminatory impact on minority communities might be structurally inevitable—a mathematical reality that cannot be fixed by better algorithms or cleaner data.

The Backlash

In 2020, following the murder of George Floyd and subsequent protests against police brutality, something remarkable happened in the academic world. A group of mathematicians published an open letter in the Notices of the American Mathematical Society—a respected professional journal—urging their colleagues to stop working on predictive policing entirely.

More than fifteen hundred mathematicians signed the letter.

This was notable because mathematicians, as a profession, tend toward political quietism. They work on abstract problems; many would say their equations are neutral tools. The letter represented a collective recognition that neutrality is itself a choice, and that some applications of mathematics deserve professional refusal.

The practical backlash has been equally significant. Cities across the United States have begun enacting legislation to restrict predictive policing technologies and other intelligence-gathering techniques deemed "invasive." Santa Cruz's 2020 ban was just the most dramatic example.

In Europe, pushback has been even stronger, with resistance emerging at both national and European Union levels. The Danish POL-INTEL project, operational since 2017 and based on technology from the controversial American firm Palantir, has faced persistent criticism. The same Palantir system has been used by German state police and Europol, Europe's law enforcement agency, drawing concern from privacy advocates.

The Deeper Questions

Predictive policing forces us to confront uncomfortable questions about the nature of crime prevention itself.

The first is philosophical. What does it mean to prevent a crime that hasn't happened yet? The concept has an ancestor in "pre-crime," the dystopian idea from Philip K. Dick's 1956 story "The Minority Report" (later a Steven Spielberg film) where psychics predict murders before they occur and perpetrators are arrested for crimes they haven't committed. Predictive policing doesn't go that far—it predicts locations and times, not specific individuals—but it operates in the same conceptual territory.

The second question is practical. Does it actually work? The evidence is surprisingly mixed. Early studies showed promising results, but more rigorous evaluations have been less conclusive. Some critics argue that the apparent successes could be explained by officers simply paying more attention to their patrol strategies, regardless of algorithmic guidance. Others note that crime rates have been falling for decades due to factors that have nothing to do with prediction.

The third question is political. Who benefits from predictive policing, and who bears its costs? The technology tends to be deployed in neighborhoods that are already over-policed. Even if it reduces some crimes, it may increase other harms: more stops, more searches, more opportunities for confrontation between police and residents. The communities subjected to algorithmic prediction rarely get a say in whether they want this form of protection.

An Alternative Vision

Some researchers have proposed a different approach entirely. Rather than predicting crime to enable enforcement, what if we used the same analytical tools to prevent crime by addressing its causes?

The "AI Ethics of Care" framework recognizes that some locations have higher crime rates because of negative environmental conditions—abandoned buildings, poor lighting, lack of economic opportunity, inadequate social services. Artificial intelligence could identify these conditions and direct resources toward addressing them.

This inverts the standard logic. Instead of asking "where should we send police?" the question becomes "where should we send investment?" Instead of predicting which individuals might commit crimes, we predict which communities might benefit from intervention.

It's a more optimistic vision, though also more expensive and politically complicated. Installing streetlights and funding youth programs doesn't generate the dramatic statistics—arrests made, crimes "prevented"—that police departments use to justify their budgets.

The Indian Experiment

While Western democracies debate the ethics of predictive policing, other nations are racing ahead.

India has emerged as an enthusiastic adopter. Various state police forces have implemented AI technologies to enhance their capabilities. Maharashtra Police launched MARVEL—the Maharashtra Advanced Research and Vigilance for Enhanced Law Enforcement system—billing it as the country's first state-level police AI system for crime prediction and detection.

Uttar Pradesh Police use an AI-powered mobile application called Trinetra for facial recognition and criminal tracking. The name comes from the Hindu concept of a third eye that sees what ordinary vision cannot.

These systems are being deployed in a country with significant concerns about civil liberties, religious tensions, and police accountability. The same tools that might help solve crimes could also facilitate surveillance of political opponents, religious minorities, or simply anyone who attracts official attention.

Where This Leads

Predictive policing is not going away. The underlying technologies—machine learning, big data analytics, geographic information systems—will only become more powerful and cheaper to deploy. The question is not whether law enforcement will use prediction, but how.

The Santa Cruz story offers a cautionary tale. The city that pioneered predictive policing in America became the first to ban it, not because the technology failed in any simple sense, but because the community decided its costs outweighed its benefits. That decision required political courage and public engagement—qualities that algorithmic governance tends to discourage.

The deeper lesson may be that prediction itself is not neutral. When we build systems that forecast where crime will occur, we embed assumptions about what crime is, who commits it, and how it should be prevented. Those assumptions deserve democratic scrutiny, not just technical optimization.

The mathematicians who signed the boycott letter understood this. They recognized that their tools could be used to reinforce injustice as easily as to reduce crime. And they decided that some problems require human judgment, not algorithmic solutions.

Whether the rest of society reaches the same conclusion remains to be seen. In the meantime, the algorithms keep running, the predictions keep flowing, and police officers keep showing up at places where crimes haven't happened yet—but might.