TL;DR: Predictive policing uses algorithms to forecast where crimes will occur or who will commit them. Research shows these systems amplify existing racial bias: when fed historical arrest data from Oakland, PredPol sent police to Black neighborhoods at twice the rate of White neighborhoods. Accuracy varies wildly: one study showed 90%, another found 0.6%. In February 2024, Chicago let its ShotSpotter contract expire after activists demonstrated the technology was unjust. In January 2024, seven Democratic members of Congress demanded the DOJ stop funding predictive policing programs entirely. These systems don't predict crime. They predict where police have historically been active, and create feedback loops that concentrate enforcement in the same communities.
What Is Predictive Policing?
Predictive policing uses data analysis to forecast criminal activity. There are two main types:
Place-Based Prediction
Algorithms identify geographic "hotspots" where crime is likely to occur. Police departments then concentrate patrols in those areas.
Major vendors:
- PredPol/Geolitica: Predicts 500x500 foot boxes where crimes are likely
- HunchLab: Incorporates weather, events, historical data
- ShotSpotter: Acoustic sensors that detect gunfire (now SoundThinking)
Person-Based Prediction
Algorithms create "risk scores" for individuals likely to commit crimes or become victims.
Examples:
- Strategic Subject List (Chicago): Scored individuals on likelihood of involvement in violence
- Crime Tracer (SoundThinking): Introduced in 2024, tracks individuals
- Palantir systems: Social network analysis to identify "at-risk" individuals
Both types claim to be objective: just math and data. Neither is.
The Fundamental Bias Problem
The Data Reflects History, Not Reality
Predictive policing algorithms are trained on historical crime data. But "crime data" really means "arrest data": a record of where police have been active, not where crime actually occurs.
If police historically patrol Black neighborhoods more heavily:
- They make more arrests in those neighborhoods
- The data shows more "crime" in those neighborhoods
- The algorithm predicts more crime there
- Police patrol even more heavily
- More arrests, more data, more prediction
This is a feedback loop that concentrates enforcement in already over-policed communities.
The Oakland Study
Researchers Kristian Lum and William Isaac tested whether PredPol would exhibit racial bias. They fed drug arrest data from Oakland into PredPol's publicly available algorithm.
Result: The algorithm would have sent police to Black neighborhoods at roughly twice the rate of White neighborhoods.
This didn't happen because the algorithm was programmed to be racist. It happened because the training data reflected decades of racially disparate policing. The algorithm learned that pattern and amplified it.
The Accuracy Illusion
When predictions send police to a neighborhood and they make arrests there, the prediction looks "accurate." But is it?
If officers are told a location is a predicted hotspot, they're primed to look for violations. They find what they're looking for. The prediction becomes self-fulfilling.
Meanwhile, crime in unpredicted areas goes undetected because police aren't there to observe it.
How Accurate Are These Systems?
Accuracy claims vary wildly:
- University of Chicago study: Claimed 90% accuracy in crime prediction
- Plainfield PD's Geolitica software: 0.6% accuracy (Cimphony, 2024)
That's not a typo. One study found 0.6% accuracy (less than one percent).
What "Accuracy" Means
Different studies measure different things:
- Did crime occur in the predicted area? (Easy to achieve with large enough areas)
- Did more crime occur in predicted vs. unpredicted areas?
- Did predictions improve on random chance?
- Did predictions lead to crime reduction?
The most meaningful question (does predictive policing actually reduce crime?) rarely has clear evidence supporting it.
ShotSpotter: The Case Study
ShotSpotter (now SoundThinking) deploys acoustic sensors to detect gunfire. When sensors detect a sound, algorithms determine if it's gunfire and alert police.
The Chicago Story
Chicago was one of ShotSpotter's largest customers. In February 2024, the city let its contract expire after years of controversy.
The problems:
- False alerts: Officers dispatched to fireworks, backfires, other sounds
- Concentration in minority neighborhoods: 90%+ of sensors in Black and Latino communities
- No evidence of effectiveness: Studies failed to show crime reduction
- Officer time wasted: Responding to alerts that weren't gunfire
PhD candidates and activist groups led the campaign, collecting research and raising public awareness about the technology's failings.
The Expansion Problem
In 2023, SoundThinking acquired Geolitica (formerly PredPol). It then integrated Geolitica's place-based prediction with ShotSpotter's acoustic detection.
In 2024, SoundThinking introduced Crime Tracer, person-based predictive policing software.
SoundThinking's integrated systems now operate in more than 250 policing jurisdictions. The company has overtaken Geolitica as the leader in the predictive policing market.
Constitutional Concerns
Fourth Amendment Issues
The Fourth Amendment protects against unreasonable searches and seizures. Courts have held that police need "reasonable suspicion" to stop and frisk someone.
Can an algorithmic prediction provide that suspicion?
If a system flags someone as "high risk," does that justify stopping them even without observing suspicious behavior? If a location is predicted as a hotspot, does that justify stopping anyone present?
SoundThinking's integrated systems raise particular concerns. Hotspot reports could potentially justify Terry stops (brief detentions for investigation) even without individualized suspicion.
Due Process Concerns
If you're on a "risk list," do you have a right to know? A right to contest the designation? A right to understand how the algorithm scored you?
In most jurisdictions, the answer is no. People can be flagged by opaque systems with no meaningful way to challenge the designation.
Political Pushback
January 2024 Congressional Letter
Seven Democratic members of Congress sent a public letter demanding the Department of Justice stop issuing grants to fund predictive policing projects unless they "can ensure that grant recipients will not use such systems in ways that have a discriminatory impact."
The letter stated:
"Mounting evidence indicates that predictive policing technologies do not reduce crime. Instead, they worsen the unequal treatment of Americans of color by law enforcement."
Academic Consensus
Researcher Rashida Richardson at Rutgers Law School noted:
"I think many predictive policing vendors like PredPol fundamentally do not understand how structural and social conditions bias or skew many forms of crime data."
The problem isn't that algorithms are malicious. It's that they inherit and amplify the biases embedded in historical data, biases that reflect decades of discriminatory policing.
The Lived Experience
One resident of Grahame Park in North London, where predictive policing is deployed, described the impact:
"It's labelled a crime hotspot. So, when the police enter the area, they're in the mindset of 'we're in a dangerous community – the people here are dangerous.'"
This is the human cost. Being labeled a "hotspot" changes how police interact with everyone in that area. Innocent people become suspects. Routine encounters become adversarial.
The algorithm didn't just predict where crime would occur. It predicted who would be treated as a criminal.
What Actually Works
Research suggests more effective approaches to public safety:
- Community investment: Addressing root causes: poverty, lack of opportunity, housing instability
- Violence intervention programs: Credible messengers who mediate conflicts before they escalate
- Mental health response: Trained responders for mental health crises instead of police
- Environmental design: Better lighting, sight lines, and public space management
- Focused deterrence: Targeted engagement with individuals at highest risk, with services offered alongside enforcement
These approaches address underlying causes rather than concentrating enforcement in already disadvantaged communities.
2025 and Beyond
The predictive policing industry is at a crossroads.
On one side: cities like Chicago are abandoning these systems, Congress is questioning federal funding, and research continues to demonstrate bias and ineffectiveness.
On the other: SoundThinking is expanding, integrating multiple surveillance technologies, and now offering person-based prediction. The systems are becoming more sophisticated, not less.
2025 will determine whether political leaders can ensure necessary reforms or whether a new generation of predictive policing technologies entrenches the problems of the old.
As one advocate put it: "Predictive policing has over-promised and under-delivered, and communities of color have carried the brunt of its failures."
The Bottom Line
Predictive policing systems claim to be objective: just algorithms following data. But the data reflects decades of racially disparate enforcement. Training algorithms on biased data produces biased predictions.
When tested, PredPol would have sent police to Black neighborhoods at twice the rate of White neighborhoods. Accuracy claims range from 90% to 0.6% depending on how you measure. Chicago let its ShotSpotter contract expire after activists demonstrated the technology didn't work and concentrated enforcement in minority communities.
These systems create feedback loops. Police go where the algorithm predicts crime. They make arrests there. The arrests become training data. The algorithm predicts more crime. The loop continues.
The question isn't whether the algorithms are racist. The question is whether we should deploy systems that amplify historical inequities and call it science. Seven members of Congress have said no. Cities are starting to agree. But the industry continues to expand.
Predictive policing doesn't predict the future. It encodes the past, and enforces it on the present.
References
- The Markup: Senators Demand DOJ Halt Funding to Predictive Policing
- MIT Technology Review: Training Data Meant to Make Predictive Policing Less Biased Is Still Racist
- TechPolicy.Press: Politicians Move to Limit Predictive Policing
- University of Michigan: Development of Predictive Policing (PDF)
- American University Journal: Future of AI in Predictive Policing
- JHULR: Algorithmic Justice or Bias
- MIT Technology Review: Predictive Policing Algorithms Are Racist