Last updated:
🔮 Predictive Policing: What It Is, and Why It Keeps Getting Banned
Algorithms that forecast crime before it happens — sold as objective data science, documented as barely better than chance. How place-based and person-based systems work, why they amplify the bias in their own training data, and who has pulled the plug.
Key findings
Predictive policing uses historical data to forecast where crime will happen (place-based, like PredPol) or who will be involved (person-based, like Chicago's Strategic Subject List). The most rigorous public test found it barely works: The Markup examined 23,631 Geolitica predictions for Plainfield, NJ and found a success rate below half a percent. The bias critique is technical, not just political — feeding Oakland's drug-arrest records into a PredPol-style model would have concentrated patrols in Black neighborhoods by proxy, because arrest data reflects where police already went, creating a feedback loop that launders history as prediction. PredPol overpredicted crime in LA's Black neighborhoods by ~25%. At its peak the software sent 5.9 million predictions to 38 agencies, potentially touching 1 in 33 Americans. Santa Cruz — where PredPol was founded — became the first US city to ban it in 2020; the EU AI Act banned profiling-based prediction continent-wide in February 2025. PredPol rebranded to Geolitica, then sold part of itself to the company behind ShotSpotter.
Predictive policing is the analytical layer of the surveillance stack — the part that tries to draw conclusions from everything the cameras and databases collect. It arrived wrapped in the language of objectivity. The independent evidence, once it finally came in, told a different story. This page is what the data actually shows, with the honest counter-case included.
By Ned Walsch · Last updated July 2026
Two kinds, one promise
Both sell the same thing: objectivity. As Geolitica's own founders put it, the premise was "less bias, more transparency, and more accountability," starting from "objective, agreed-upon facts." The question that took a decade to answer properly was whether the facts going in were objective at all.
Does it work? The half-percent problem
The most rigorous public evaluation is devastating in its plainness. In 2023 The Markup obtained and analysed 23,631 predictions that Geolitica generated for the Plainfield, New Jersey police department across most of 2018. They checked each prediction against crimes later reported in that category and place.
It is not an outlier finding. LAPD's Inspector General concluded there was insufficient evidence PredPol reduced crime, and the department dropped it in 2020, having already shut its person-based LASER program in 2019. Early vendor-cited pilots (Atlanta's 8–9% drop in 2013, Shreveport's in-house model) were never controlled studies, and the enthusiasm did not survive independent scrutiny.
Why it amplifies bias — the feedback loop
This is the part that matters most, and it is a point about how the technology works, not merely how people feel about it. Place-based systems train on arrest and incident data. But that data does not record where crime happens — it records where police were sent. Those are not the same thing, and the gap between them is where the bias lives.
The measurable result: an analysis found PredPol overpredicted crime in Los Angeles's Black neighborhoods by roughly 25% relative to actual rates. Because the system excludes explicit personal characteristics and runs on location, vendors called it unbiased. Critics call it discrimination by proxy — the same outcome, one layer of abstraction removed.
How far it spread
Who pulled the plug
| Where | Action | When |
|---|---|---|
| Santa Cruz, CA | First US city to ban predictive policing — the city where PredPol was founded | 2020 |
| LAPD | Dropped PredPol (IG found no evidence it worked); LASER shut 2019 | 2020 |
| European Union | AI Act bans profiling-based prediction of individual criminal behavior — continent-wide | Feb 2025 |
| Oakland & others | Restricted or declined to renew | 2019–2024 |
| PredPol / Geolitica | Rebranded 2021; sold part of operations to SoundThinking (ShotSpotter) | 2023 |
Where it sits in the surveillance stack
Predictive policing rarely collects new data — it draws conclusions from what other systems gather. That is why it compounds. A model directs patrols to a neighborhood; that neighborhood's ALPR cameras log everyone who drives through; those faces run against a watchlist. Each tool is questionable in isolation. Chained together, they manufacture a self-reinforcing bubble of suspicion around exactly the places a flawed forecast already pointed at.
Frequently asked questions
What is predictive policing?
Predictive policing is the use of algorithms and historical data to forecast where crime is likely to occur or who is likely to be involved in it, so police can direct patrols or attention in advance. There are two broad kinds. Place-based systems, like the software once sold as PredPol, divide a city into small boxes — roughly 500 feet across — and predict which boxes will see crime on a given shift. Person-based systems, like Chicago's former Strategic Subject List, score individuals on their supposed likelihood of being involved in violence. Both borrow the language of objectivity and data science. Whether they deliver on it is the entire controversy, and the evidence is not kind.
Does predictive policing actually work?
The most rigorous public test suggests it barely does. In 2023 the nonprofit newsroom The Markup examined 23,631 predictions generated by Geolitica — the company formerly called PredPol — for Plainfield, New Jersey in 2018. Fewer than 100 of those predictions lined up with a crime that was later reported in the predicted category and location. That is a success rate below half a percent. LAPD's own Inspector General separately concluded there was insufficient data to show PredPol reduced crime, and the department dropped it in 2020. Some early pilots reported crime reductions — Atlanta cited 8 to 9 percent drops in treated zones in 2013 — but these were not controlled studies, and the vendor's later renaming and partial sale to the company behind ShotSpotter tells its own story about how the product fared.
Why is predictive policing accused of racial bias?
Because it learns from arrest data, and arrest data reflects where police have already been sent, not where crime actually happens. This is the core of the critique and it is a technical point, not just a political one. In an influential simulation, researchers fed Oakland's drug-arrest records into a PredPol-style algorithm and found it would have concentrated patrols in predominantly Black neighborhoods — not because more drug use occurred there, surveys suggest it did not, but because that is where past enforcement was focused. The algorithm then sends officers back to those areas, producing more arrests, which feed back into the model as fresh confirmation. It is a feedback loop that launders historical bias as a neutral prediction. A separate analysis found PredPol overpredicted crime in Black neighborhoods in Los Angeles by around 25 percent relative to actual rates.
Which cities and countries have banned it?
The bans started in the place the technology was born. Santa Cruz, California — where PredPol was founded — became the first US city to ban predictive policing in 2020. Oakland and a number of other jurisdictions have restricted or dropped it. The most sweeping action is in Europe: the EU's Artificial Intelligence Act, which took effect in February 2025, bans predictive policing systems that forecast an individual's criminal behavior through profiling. That is a continent-wide prohibition on the person-based variety. In the US there is no federal ban; the picture is a patchwork of individual departments quietly dropping contracts after the results and the lawsuits came in.
What happened to PredPol?
It rebranded and receded. PredPol grew out of an LAPD and UCLA collaboration around 2010, using a model originally built to forecast earthquake aftershocks and applying it to property crime. It became the most widely used predictive policing product in the country — by 2020, more than 50 departments used it, and analyses found over 5.9 million predictions were sent to 38 agencies between 2018 and 2021, potentially touching more than one in 33 Americans. In 2021 the company renamed itself Geolitica, and in 2023 it reportedly sold part of its operations to SoundThinking, the firm previously known as ShotSpotter. The name most associated with predictive policing effectively retired after the independent analyses landed.
How is this different from ALPR or facial recognition?
They are different layers of the same surveillance stack, and they increasingly feed each other. Automated license plate readers record where specific vehicles go. Facial recognition identifies specific faces. Predictive policing is the analytical layer that sits on top — it does not collect a new kind of data so much as try to draw conclusions from data other systems collect, plus historical records. The concern with combining them is compounding: a predictive system that directs patrols to a neighborhood, whose ALPR cameras then log everyone who drives through, whose faces are then run against a watchlist, produces a self-reinforcing bubble of scrutiny around exactly the places the flawed model already pointed at. Each tool is questionable alone; together they are a machine for manufacturing suspicion.
Can I find out if my city uses predictive policing?
Sometimes, with effort. Unlike ALPR cameras, predictive policing software leaves little physical trace, so the paper trail is procurement records and public-records requests rather than a map. The EFF's Atlas of Surveillance catalogues which US agencies have used which surveillance technologies, including predictive tools, and is the best starting point. Beyond that, city council budget documents, contract disclosures, and the work of local investigative reporters are where these programs surface. Because vendors rebrand and departments rarely announce cancellations, current information is genuinely hard to pin down — a records request asking specifically which crime-forecasting or risk-assessment software a department has licensed, and when, is the most reliable route.
Is predictive policing the same as the movie Minority Report?
It is the comparison everyone reaches for, and it is both apt and misleading. Apt, because the premise — using prediction to act against crime before it happens — is exactly the same, and the film's unease about punishing people for forecasts rather than acts is precisely the civil-liberties objection. Misleading, because Minority Report imagines prediction that is essentially perfect, three psychics who are almost never wrong. Real predictive policing is the opposite: the documented problem is not that it works too well and traps the innocent through infallible foresight, but that it works barely at all, while lending a veneer of scientific certainty to patrol patterns that mostly reproduce where police already went. The dystopia is not omniscience. It is a coin-flip dressed as a crystal ball.