Predictive policing

Predictive policing uses statistical models and machine learning to forecast where crime is likely, or who is likely to be involved. Place-based tools map hotspots; person-based tools score individuals and draw the strongest opposition.

Guide

What it is

  • Place-based tools use historical crime reports, calls for service and time patterns to map forecast hotspots, usually refreshed daily or each shift.
  • Person-based tools score individuals as being at elevated risk of committing, or becoming victims of, violent crime.
  • Custody risk scoring tools such as Durham Constabulary's HART classify suspects as low, medium or high risk to inform decisions such as bail.

How it is used

Outputs guide which areas get extra patrols, which people receive targeted intervention and which cases are prioritised. US departments adopted the technology most widely, with PredPol (now Geolitica) the best known. In the UK, Kent Police used PredPol for several years. The record tracks tools in use, those wound down, independent evaluations and the legal position.

Key tensions

The feedback loop

Models learn from recorded crime, which reflects where police have looked. More patrols bring more recorded crime, which seems to confirm the prediction. Chicago's inspector general found its person-based list was partly measuring its own enforcement.

False positives

Serious crime is rare even in high-risk areas, so a model that is 90 per cent accurate can still flag ten people who would not offend for every one it gets right.

Opacity and bias

Vendors often keep their algorithms secret. A review of Durham's tool found that postcode data introduced bias against people from poorer areas, and the UK's algorithmic transparency standard is still voluntary.

View all predictive policing records Latest coverage

In depth

The full guide, about 5 minutes to read.

What is predictive policing?

Predictive policing is a broad term for the use of statistical models and machine learning algorithms to try to forecast where crime is likely to occur, or in some more controversial applications, who is likely to be involved in it. The outputs of these systems are typically used to inform how police resources are deployed: which neighbourhoods get extra patrols, which individuals receive targeted outreach or intervention, which cases get prioritised for follow-up. The pitch from vendors is straightforward: use data to put officers where they are most needed, rather than relying on gut instinct or the inherited assumptions baked into traditional patrol patterns.

Place-based predictive tools are the most common variety. These ingest historical crime reports, calls for service, time and date patterns, and sometimes external data like weather, local events, or even social media activity to generate maps of forecast crime hotspots, typically refreshed daily or per shift. Person-based tools are a smaller but far more contentious category: rather than flagging locations, they score individuals, producing ranked lists of people assessed as being at elevated risk of committing or becoming victims of violent crime. It is the latter category that has attracted the most sustained criticism, and in several US cities, the most organised public opposition.

Where is predictive policing used?

The United States has been the most prolific adopter. PredPol, now rebranded as Geolitica, was for several years the best-known place-based system, deployed in dozens of departments from Los Angeles to Atlanta. ShotSpotter (which uses acoustic sensors and machine learning to detect and localise gunshots) has been rolled out across major US cities, though it has faced mounting questions about its accuracy and the disproportionate impact of its deployments on Black and Latino communities. The Chicago Police Department's Strategic Subject List, a person-based algorithm that assigned risk scores to individuals, was eventually decommissioned in 2020 following years of controversy and a damning report from the city's inspector general.

In the United Kingdom, adoption has been patchier and considerably more cautious in its public presentation, though several forces have quietly implemented tools that share the same basic logic. West Midlands Police developed and trialled its own Harm Assessment Risk Tool (HART) which used machine learning to classify suspects at the point of custody into low, medium, and high risk categories to inform decisions about pre-charge bail. Durham Constabulary used a similar tool called HART to inform custody decisions. These are risk-scoring tools rather than patrol-directing tools, but they rest on the same foundational idea: that an algorithm can usefully predict future behaviour from past data.

Kent Police was among the first UK forces to adopt place-based predictive software, working with the commercial product PredPol for a number of years. The system generated grid squares predicted to have an elevated risk of burglary, vehicle crime, and violence, and directed patrol resources towards them. Other forces have been interested observers without becoming committed adopters, partly due to cost, partly due to a lack of clear legal guidance, and partly (it is fair to say) because of reputational caution in a climate where the technology is publicly contested.

The feedback loop problem

The fundamental criticism of predictive policing is not that the maths is wrong, but that the data the models are trained on is not a neutral record of where crime happens. It is a record of where crime was previously detected and recorded, which is shaped as much by historical policing patterns as by underlying criminal activity. If a neighbourhood has historically been subjected to intensive stop-and-search, a high volume of arrests from that area will appear in the data. Feed that data into a model and it will predict more crime in that area. Send more officers there, make more stops and arrests, and the prediction appears to have been validated. The model has not found where crime is; it has found where policing has been.

This feedback loop is not merely a theoretical concern. The Chicago inspector general's report on the Strategic Subject List found that individuals who were assigned high risk scores were more likely to be stopped by police, which in turn increased their likelihood of appearing in the system's training data as people with police contact, which tended to increase their score further. The system, in effect, was partially measuring its own enforcement activity rather than actual criminal risk.

Critics also point to a deeper conceptual problem: most serious crime is comparatively rare, even in areas classified as high risk. A model that is accurate 90 per cent of the time at a population level can still generate enormous numbers of false positives when it is predicting something that affects a small minority. If one in a hundred people in a flagged area would actually go on to commit a violent offence, a 90 per cent accurate model will correctly identify nine of them, but it will also flag 90 people who would not. Every intervention based on the model's output involves a decision about what to do with those 90 false positives.

Bias, transparency, and accountability

Commercial predictive policing vendors have frequently resisted disclosing the details of their algorithms, citing intellectual property. This creates a significant accountability problem: it is very difficult to challenge a decision or assess its fairness when the logic behind it is opaque. Several US cities that adopted predictive tools struggled to obtain meaningful information about how they worked, even when subjected to public records requests. In the United Kingdom, the Algorithmic Transparency Recording Standard has been introduced as a voluntary framework for public bodies to publish information about algorithmic tools used in decision-making, but uptake among police forces has been slow and the standard is not yet mandatory.

Racial and socioeconomic bias in predictive policing outputs has been extensively documented in the US context, and there is meaningful evidence of similar patterns in UK deployments. A review of Durham Constabulary's HART tool found that postcode data used as a proxy for socioeconomic deprivation introduced bias against individuals from poorer areas, and raised questions about whether the tool was capable of complying with the Public Sector Equality Duty. The force modified the tool in response, but the episode illustrated how difficult it can be to build genuinely fair risk-scoring systems when the input data carries the imprint of structural inequality.

The current picture

Predictive policing is neither as transformative as its proponents claimed when the technology first emerged, nor as universally discredited as its harshest critics would suggest. The reality is messier and more contextual. Some place-based tools appear to produce modest reductions in certain crime types in certain conditions; others have shown no measurable effect when subjected to independent evaluation. Person-based tools have faced the most sustained opposition and several have been quietly wound down, though interest in them has not disappeared. It has shifted towards tools described as risk assessment or early intervention, which often share the same underlying logic under a less loaded name.

The legal framework governing predictive policing in England and Wales remains underdeveloped. There is no primary legislation specifically addressing algorithmic decision-making in policing, and while data protection law and equality duties provide some constraint, they were not designed with these tools in mind and their application is contested. This is an area where the law is moving considerably more slowly than the technology.

Records on the tracker

Open the full tracker

55 records on predictive policing, most recently checked first: 40 in use now, 10 withdrawn and 5 on the horizon. Each is one police force or public body using one system.

  1. United StatesDeveloped with Louisiana State University researchers (federal grant)2018

    Louisiana Department of Public Safety and Corrections — TIGER risk score for parole eligibility

    Louisiana

    OperationalChecked
  2. Spain

    Departament de Justícia i Qualitat Democràtica (Generalitat de Catalunya), prison services — RisCanvi prison risk assessment

    Catalonia

    OperationalChecked
  3. Spain

    Universidad de Málaga (ATREIDES project) — AI patrol-route planning from a digital twin of the city

    Málaga

    In developmentChecked
  4. United StatesDeveloped in-house by ICE2019

    US Immigration and Customs Enforcement — Hurricane Score (abscondment risk score for Alternatives to Detention)

    Nationwide (Enforcement and Removal Operations, Alternatives to Detention programme)

    OperationalChecked
  5. United States

    US Customs and Border Protection — Predictive Intelligence Program: licence plate reader networks with a travel-pattern algorithm

    Border Patrol: southwest and northern borders and interior highways

    OperationalChecked
  6. United StatesSandia National Laboratories2024

    ICE Homeland Security Investigations — Investigative Prioritization Aggregator (scoring of entities in investigations)

    Nationwide

    OperationalChecked
  7. United StatesInvariant Corporation2025

    US Customs and Border Protection — Acoustic gunshot detection pilot (Invariant Corporation)

    Southwest border

    TrialChecked
  8. United States2013

    US Customs and Border Protection — Passenger Security Assessment Model (Automated Targeting System)

    Nationwide

    OperationalChecked
  9. United States2012

    US Customs and Border Protection — Traveler Entity Resolution models

    Nationwide

    OperationalChecked
  10. United StatesManTech2017

    US Customs and Border Protection — Passport Anomaly Model

    Nationwide

    OperationalChecked
  11. United StatesMaxar2022

    US Customs and Border Protection — KESTREL aircraft landing location predictor (Maxar)

    Air and maritime domain

    OperationalChecked
  12. United States

    Drug Enforcement Administration — Gunshot detection system

    Nationwide

    OperationalChecked
  13. United StatesSoundThinking

    Bureau of Alcohol, Tobacco, Firearms and Explosives — Access to ShotSpotter via external state and local systems

    Nationwide

    OperationalChecked
  14. United StatesSAS

    Federal Bureau of Prisons — BRAVO classification: predicting misconduct to assign security levels

    Nationwide

    OperationalChecked
  15. United States

    Federal Bureau of Prisons — Static-99 Sex Offender Data System

    Nationwide

    OperationalChecked
  16. United States

    Organized Crime Drug Enforcement Task Forces (US Department of Justice) — Predictive analytics for financial links among subjects of criminal investigations

    Nationwide

    ProposedChecked
  17. United StatesFlock Safety

    Richmond Police Department (Virginia) — Flock Safety gunshot detection

    Richmond, Virginia

    OperationalChecked
  18. United StatesSoundThinking (ShotSpotter)

    Columbia Police Department (South Carolina) — ShotSpotter acoustic gunshot detection

    Columbia, South Carolina

    OperationalChecked
  19. United StatesSoundThinking (ShotSpotter)2015

    New York City Police Department — ShotSpotter acoustic gunshot detection

    New York City

    OperationalChecked
  20. United StatesSoundThinking (ShotSpotter)2014

    Pittsburgh Bureau of Police — ShotSpotter acoustic gunshot detection

    Pittsburgh, Pennsylvania

    OperationalChecked
  21. United StatesSoundThinking (ShotSpotter)

    Cambridge Police Department — ShotSpotter acoustic gunshot detection

    Cambridge, Massachusetts

    OperationalChecked
  22. SwitzerlandLogObject AG

    Kantonspolizei Aargau — Precobs place-based burglary prediction

    Canton of Aargau

    OperationalChecked
  23. United Kingdom

    West Midlands Police — Place-based machine learning models (knife crime, vehicle theft, seasonality planner)

    West Midlands

    OperationalChecked
  24. United Kingdom

    West Midlands Police — Stalking and harassment harm escalation model

    West Midlands

    OperationalChecked
  25. United Kingdom

    West Yorkshire Police — Domestic Abuse Assessment Tool (machine learning risk prediction)

    West Yorkshire

    ProposedChecked
  26. United KingdomSolvarithm Ltd (contractor)

    Thames Valley Police — DARAT (Domestic Abuse Risk Assessment Tool, machine learning)

    Thames Valley

    In developmentChecked
  27. United Kingdom

    Metropolitan Police — Stalking Threat Assessment Centre machine learning triage

    London

    OperationalChecked
  28. United Kingdom2012

    Metropolitan Police — Gangs Violence Matrix (individual risk and harm scoring database)

    London

    DiscontinuedChecked
  29. Netherlands

    Reclassering Nederland — OxRec reoffending risk assessment

    PausedChecked
  30. SwitzerlandInstitut für musterbasierte Prognosetechnik (Precobs developer); LogObject AG listed by AlgorithmWatch CH2013

    Stadtpolizei Zürich — Precobs place-based burglary prediction

    City of Zürich

    OperationalChecked
  31. SwitzerlandLogObject AG

    Kantonspolizei Basel-Landschaft — Precobs place-based burglary prediction

    Canton of Basel-Landschaft

    DiscontinuedChecked
  32. United KingdomIn-house, with Cambridge University2016

    Durham Constabulary — HART (Harm Assessment Risk Tool)

    England

    DiscontinuedChecked
  33. SpainVioGen2007

    Secretariat of State for Security (Ministry of the Interior) — Algorithmic Risk Assessment

    Spain, national

    OperationalChecked
  34. United StatesPredPol, renamed Geolitica in 20212011

    Los Angeles Police Department — Place-Based Predictive Policing

    California and multiple US jurisdictions

    DiscontinuedChecked
  35. United KingdomPredPol2013

    Kent Police — Place-Based Predictive Policing

    Kent

    DiscontinuedChecked
  36. NetherlandsBuilt in-house by the Dutch National Police2015

    Dutch National Police (Politie) — Crime Anticipation System (CAS)

    Netherlands, national

    DiscontinuedChecked
  37. United KingdomBuilt in-house2017

    Avon and Somerset Police — Predictive Analytics Programme

    Avon and Somerset

    OperationalChecked
  38. United KingdomNot specified2016

    Bristol City Council — Think Family Database and child exploitation risk models

    Bristol

    DiscontinuedChecked
  39. United KingdomUK Ministry of Justice; technical support from Capita2001

    HM Prison and Probation Service — OASys offender risk assessment

    England and Wales

    OperationalChecked
  40. United KingdomDeveloped in-house (Ministry of Justice data science team)2023

    Ministry of Justice — Individual Risk Scoring

    England and Wales (research phase)

    ProposedChecked
  41. South AfricaSoundThinking (formerly ShotSpotter)2016

    City of Cape Town — ShotSpotter acoustic gunshot detection

    Hanover Park, Manenberg and Lavender Hill, Cape Town

    DiscontinuedChecked
  42. United StatesSoundThinking (ShotSpotter)2015

    Denver Police Department — AI Gunshot Detection

    Colorado

    OperationalChecked
  43. United StatesSoundThinking (ShotSpotter)2020

    Albuquerque Police Department — AI Gunshot Detection

    New Mexico

    OperationalChecked
  44. United StatesSoundThinking (ShotSpotter)2014

    Worcester Police Department — AI Gunshot Detection

    Massachusetts

    OperationalChecked
  45. United StatesSoundThinking (ShotSpotter)2022

    Richland County Sheriff's Department — AI Gunshot Detection

    South Carolina

    OperationalChecked
  46. United StatesSoundThinking (ShotSpotter)2022

    Macon-Bibb County Sheriff's Office — AI Gunshot Detection

    Georgia

    OperationalChecked
  47. United StatesSoundThinking (ShotSpotter)2013

    Peoria Police Department — AI Gunshot Detection

    Illinois

    OperationalChecked
  48. United StatesSoundThinking (ShotSpotter)

    City of Fayetteville (North Carolina) — AI Gunshot Detection

    North Carolina

    OperationalChecked
  49. United Kingdom

    Essex Police — Knife Crime and Violence Model (Fearless Futures)

    Essex

    TrialChecked
  50. United KingdomTrilateral Research (CESIUM)2023

    Lincolnshire Police — CESIUM child exploitation risk identification

    Lincolnshire

    OperationalChecked
  51. United KingdomTrilateral Research2024

    Nottinghamshire Police — CESIUM child exploitation risk identification

    Nottinghamshire

    TrialChecked
  52. United StatesSoundThinking (ShotSpotter)

    Chicago Police Department — AI Gunshot Detection

    Illinois

    DiscontinuedChecked
  53. ColombiaHigh Tech Software S.A.S.2025

    Policía Nacional de Colombia — IAPol crime prediction and patrol allocation

    OperationalChecked
  54. United StatesEquivant, formerly Northpointe2000

    US state courts and corrections agencies — COMPAS recidivism and pretrial risk scoring

    Multiple US states

    OperationalChecked
  55. AustraliaBuilt in-house by NSW Police Force2000

    New South Wales Police Force — Suspect Target Management Plan (STMP III / Chimera)

    New South Wales

    DiscontinuedChecked

Latest coverage