Police AI that was withdrawn or paused

Legal basis

NSW Police Force internal policy and guidelines rather than statute

Oversight

NSW Law Enforcement Conduct Commission (Operation Tepito); NSW Bureau of Crime Statistics and Research evaluation

Chimera collated police information sources weighted by perceived reliability and computed a Crime Severity Index score from charges and offending indicators, scoring each offence against the Australian Bureau of Statistics National Offence Index and typically covering the prior six months. A benchmark score limited the candidate list. Intelligence officers reviewed the shortlist alongside offending history and local crime patterns, could add people from local knowledge, documented a justification, and a senior officer approved selection. Those selected were subjected to increased overt and covert police attention. STMP was introduced in 2000, with STMP III and Chimera from November 2020. It was discontinued for children and young people in October 2023 and for adults from December 2023, following the Law Enforcement Conduct Commission's Operation Tepito investigation.

Source: NSW Law Enforcement Conduct Commission, Operation Tepito final report · last checked 2026-08-20

Legal basis

Municipal contracting; no dedicated statute

Oversight

LAPD Office of the Inspector General reviewed the deployment

The LAPD was the earliest adopter and co-developer, dropping the system in 2020 citing budget, after an inspector general report found it unclear whether the software affected crime trends. Geolitica has since ceased operations, with SoundThinking acquiring specific intellectual property and hiring the engineering team. A Gizmodo and Markup analysis of millions of predictions found on an unsecured server reported the software disproportionately and often relentlessly targeted low-income communities of colour.

Source: WIRED; Gizmodo and The Markup analysis · last checked 2026-08-20

Legal basis

Force procurement; no dedicated statute

Oversight

None identified beyond the force itself

The first UK force to test PredPol. The force cancelled the contract after saying it was challenging to demonstrate that the programme reduced crime. This is the same pattern seen in Los Angeles and in the Dutch Crime Anticipation System: the operating agency itself was unable to show the software worked.

Source: Liberty Investigates · last checked 2026-08-20

Legal basis

No dedicated statute; registered in the Dutch national algorithm register

Oversight

Dutch national algorithm register provides public transparency; no dedicated regulator identified

Divided cities into small grid cells and combined police crime records with national statistics including average neighbourhood income, benefit recipients, household composition and distance to the nearest known offender, outputting hotspots and hot times in four-hour blocks. Officers were sent to drive through flagged areas, often for only a few minutes. The model was closed, so officers could not see why a cell was flagged. An indicator counting non-Western immigrants was removed in 2017 after being found to add no predictive value. Piloted in 2015 in Enschede, Groningen-Noord, Hoefkade and Hoorn, then rolled out from 2017 across more than half of the then 167 police departments. Halted in January 2026 and discontinued after an internal report concluded its predictive capability was extremely limited and the force could never demonstrate its effect.

Source: AlgorithmWatch; Dutch national algorithm register · last checked 2026-08-20

Legal basis

Custody decision support

Oversight

Academic review; extensive published critique

An early and unusually well-documented custody risk-scoring tool, developed by Durham Constabulary with the University of Cambridge, using random forest models trained on historical criminal justice data to sort individuals into risk categories. Operational from 2016 until 2021, when Durham Constabulary discontinued it. The stated reason was the resources required to continually refine and refresh the model to satisfy ethical, legal and governance requirements. A Freedom of Information response obtained by Fair Trials showed the tool assessed 12,200 people a total of 22,265 times across that period. Its use of a commercial postcode-based variable attracted particular criticism. The variable was Experian's Mosaic socio-geodemographic segmentation product, a commercial consumer classification dataset rather than a policing one.

Source: Published academic evaluation and force statements · last checked 2026-08-20

Legal basis

Not established in available sources

Oversight

Reviewed by Social Finance, which could not retest the models

A database launched in 2016 holding records on close to half a million Bristol residents, with child sexual exploitation and child criminal exploitation risk models flagging children at risk; the child criminal exploitation model was introduced in 2019. Stopped around June 2023 after council staff judged the models not fit for operational use, with recent victims of sexual offences scoring below burglary perpetrators. No records of the decision were retained. Reviewers from Social Finance could not retest the models because the source code and model variables could not be found. That combination, a discontinued predictive system with no decision record and no recoverable model, is itself the finding worth recording.

Source: Liberty Investigates · last checked 2026-08-20

READ THE STATUS NOTE BELOW BEFORE APPLYING. The deployment is well established. Engineering News, carried by Polity on 28 July 2023, quoting Alderman JP Smith: acoustic triangulation and GPS pinpoint gunshots and relay the location to dispatch within seconds, with 'within a minute of a weapon being discharged, the police can be deployed' and information on shots fired and weapon type. Live in Hanover Park from December, Manenberg from February and Lavender Hill from March, with 35 firearms and 400 rounds confiscated and 50 arrests, and 10 million rand budgeted for a Nyanga expansion in 2023/24. Cape Argus, 18 June 2026, and IOL, 19 June 2026: the system began as a pilot in Hanover Park in 2016 and expanded to Manenberg and Lavender Hill from 2023; the City invested about 860 million rand over three years across CCTV, drones, body cameras, dashcams, aerial surveillance and gunshot detection; the City described the technology as useful for alerting staff to active shooting incidents and for identifying hotspots to guide deployment. Status note. The current status is genuinely unresolved and 'paused' is a deliberately cautious reading, not a confident one. What is quoted in June 2026 is Smith saying: 'The three-year gunshot detection technology tender has run its course and is currently under review.' That establishes the contract ended; it does not on its own establish that the programme is over. Separately, Police.AI's own archive carries an Inside Metros story of 5 August 2026 whose summary states that Cape Town's city council ended the use of the system in mid-2026, but that article could not be fetched in this run and the site's own summary is not treated here as a primary source. If you can open that article and it confirms the council ended the deployment, change this record to discontinued. Nothing read establishes that the system is still running, which is why it is not proposed as operational. Effectiveness is contested and the record does not take a side: the City calls the technology useful, while crime-focused organisations quoted in the same reporting question whether it achieved its objectives given continued high violence. Fields left out and why. 'legalBasis' and 'oversight' are omitted: nothing established. Cost is not given as a field because the 860 million rand figure covers a basket of technologies, not this system.

Source: Cape Argus · last checked 2026-09-09

Legal basis

No dedicated statutory basis. The EU AI Act, in force from July 2024, classified tools of this kind as high risk, which is widely reported as a factor in the decision to stop using it.

Oversight

Spain's Consejo de Transparencia y Buen Gobierno, which ordered the Interior Ministry to release information about the tool following requests by the transparency organisation Civio

VeriPol applied natural language analysis to written robbery reports to flag characteristics statistically associated with false reports. It analysed the text of a report a victim or witness gave rather than physical evidence or camera footage, which makes it a materially different form of police AI from the video and biometric systems that make up most of this tracker. Launched nationally in October 2018 after a 2017 pilot in Malaga and Murcia, and promoted at the time as over 90 percent accurate. Spain's Dirección General de la Policía confirmed to the transparency organisation Civio that the tool stopped being used on 21 October 2024, giving as its reason that it lacked validity in judicial proceedings. That date is three months after the EU AI Act came into force. The evidence base was thin from the outset and this is central to understanding the entry rather than incidental to it. The model was trained on 1,122 robbery reports from 2015, and rested on a starting assumption that 57 percent of violent robbery reports are false. A University of Valencia analysis identified serious design deficiencies. Usage fell sharply over time, from roughly 84,000 reports analysed up to October 2020 to 3,762 in 2022. The Interior Ministry has declined to release which police stations deployed it or how often it was used, and a transparency complaint on that point remains open.

Source: Civio, reporting the Directorate General of Police's confirmation · last checked 2026-08-27

Legal basis

Privacy Act (federal public sector); found in breach

Oversight

Office of the Privacy Commissioner of Canada, Special Report to Parliament

The RCMP's use of Clearview AI ended after the Office of the Privacy Commissioner found in 2021 that it violated the federal Privacy Act, since Clearview's underlying image collection was itself unlawful. Clearview ceased Canadian operations shortly afterward. In response, the RCMP established the National Technology Onboarding Program (NTOP) in 2021 to centralise assessment of new operational technologies. The RCMP's own 2026 transparency publication states it currently uses only 'face matching' (grouping and sorting lawfully obtained images and video, not searching to identify an unknown individual) rather than full facial identification, and that it does not use facial recognition on footage captured by its drone fleet. Separately, other Canadian police services have moved into live and image-search facial recognition; see the Edmonton, York/Peel and Toronto entries.

Source: Office of the Privacy Commissioner of Canada, Special Report to Parliament · last checked 2026-08-26

Legal basis

Privacy Act 1988; Australian Government Agencies Privacy Code

Oversight

Office of the Australian Information Commissioner (OAIC), joint investigation with UK ICO

This trial was run specifically by the Australian Federal Police's Australian Centre to Counter Child Exploitation (ACCCE), not the AFP generally. The AFP's own account, given in a Senate Estimates brief, states ACCCE personnel ran a limited pilot from November 2019 to January 2020, testing whether Clearview could match side-profile photographs of possible child exploitation victims against frontal reference images; temporary accounts were terminated in March 2020. The Australian Information Commissioner subsequently found the AFP had failed to comply with its privacy obligations, including a requirement to complete a privacy impact assessment before the trial. Separately, and consistent with the free-trial period documented for other Australian police services in this entry, Queensland, Victoria and South Australia police also had access during the same general window.

Source: Office of the Australian Information Commissioner, Commissioner-initiated investigation determination · last checked 2026-08-26

Legal basis

Hellenic DPA ruled the procurement, operation and use unlawful in Decision 45/2025 (31 Dec 2025)

Oversight

Hellenic Data Protection Authority

1,000 handheld devices for Hellenic Police performing facial recognition, fingerprint ID, and document/plate scanning during street stops, matching against national, EU and international databases including SIS II, VIS, EURODAC and Interpol-linked systems. Contract signed 2019 with Intracom Telecom, €4m, 75% funded by the EU's Internal Security Fund. Devices were delivered fully functional. Following a complaint by Homo Digitalis (Mar 2020), the Hellenic DPA ruled the whole programme unlawful in Decision 45/2025, dated 31 Dec 2025 -- an EU-funded, fully procured and delivered biometric system found unlawful after the money was spent. Reported as targeted at stops of undocumented migrants, placing it in borders/immigration as much as ordinary policing. Current operational status unclear following the ruling -- recorded as discontinued pending confirmation. Also in Greece: the Ministry of Asylum and Migration was fined a record €175,000 by the Hellenic DPA over the KENTAUROS and HYPERION border/camp management systems.

Source: EDRi; Homo Digitalis; Hellenic DPA Decision 45/2025 · last checked 2026-08-26

Legal basis

Ley 5688 (Sistema Integral de Seguridad Pública de la Ciudad), under which the system was permitted to search only for persons listed in the national fugitives register CONARC or subject to a judicial order. Judicially suspended on 21 April 2022; its implementation was declared unconstitutional on 7 September 2022 and that ruling was upheld on appeal on 29 April 2023.

Oversight

The city legislature approved the system subject to control by an ad hoc commission of legislators and by the Defensoría del Pueblo. CELS reports that as of February 2024 the court required, before any resumption, an effective audit by the control authorities, a budgetary framework for it, access to the supplier's software information, and an evaluation of possible discriminatory bias.

Operated in the Buenos Aires underground and public space from 2019, matching faces against CONARC. Judge Roberto Gallardo suspended it on 21 April 2022 after an amparo brought by the Observatorio de Derecho Informático Argentino (ODIA); raids on city offices found more than 9 million queries against the national identity register when CONARC held only some 35,000 to 40,000 records, including searches on the President, the Vice-President, judges, journalists, activists and business figures, and on more than 15,000 people not on the CONARC list. The city said these were identity validations for permits, pandemic procedures and emergency services. On 7 September 2022 Judge Elena Liberatori declared the implementation unconstitutional, finding the system had been launched and run outside the existing regulatory framework and without control mechanisms; the Cámara en lo Contencioso Administrativo rejected the city's appeal on 29 April 2023. Status is 'paused' rather than 'discontinued' because CELS, Fundación Vía Libre and ODIA all report the system as suspended, with conditions specified for resumption, rather than terminated. vendor is left empty: the AlSur regional report of June 2025 names Danaide as the supplier, but that is a single secondary source and no primary document read here confirms it. Sources in this record are in Spanish. Note that the city's prosecution service separately acquired Clearview AI in June 2025 — a distinct deployment with its own record.

Source: CELS · last checked 2026-09-10

Legal basis

None in place at the time; this absence is what the regulatory reprimand below addresses

Oversight

Finnish Data Protection Ombudsman, who reprimanded the National Police Board

Finland's National Bureau of Investigation ran approximately 120 Clearview AI searches in early 2020 while trying to identify possible victims in child sexual abuse material, producing one hit, and separately tested Arachnid for the same purpose; police concluded Clearview was unsuitable for the task. The Finnish Data Protection Ombudsman subsequently reprimanded the National Police Board because the trial had taken place without appropriate controller approval or supervision.

Source: Office of the Data Protection Ombudsman (Finland), statutory reprimand · last checked 2026-08-26

Legal basis

City council authorisation

Oversight

City council vote following public disclosure

Police Chief Mike Canfield disclosed to the city council in February 2026 that federal agencies had gained access to the city’s Flock camera data without the city’s knowledge. The council voted unanimously to terminate the contract and the department discontinued use. Included as a specific, well-documented instance of a US agency withdrawing a deployment after an oversight failure, distinct from the wider Flock Safety entry tracked separately.

Source: City council disclosure, reported by State of Surveillance and the ACLU · last checked 2026-08-26

Legal basis

Municipal contracting

Oversight

None identified

A self-driving camera-equipped patrol robot deployed in Times Square subway station, accompanied by police officers, as a trial of crime prevention effectiveness. Leased at 9 dollars an hour under a seven-month contract worth 12,250 dollars. The NYPD retired the robot in January after four months deployed and confirmed the pilot was complete. Huntington Park Police Department in California renewed a Knightscope contract for a sixth year in August 2024, though that detail is vendor-sourced and carries lower confidence.

Source: Fox News; Knightscope contract announcement · last checked 2026-08-26

Legal basis

New Orleans city ordinance restricting police facial recognition; the ACLU alleges the department's use breached it

Oversight

New Orleans City Council; ACLU of Louisiana investigation

Project NOLA is a private non-profit founded in 2009 by a former police officer, based on the University of New Orleans campus, acting as a hub for live video from thousands of cameras on private volunteer properties whose owners pay annual connectivity fees. Around 200 of the most sophisticated cameras run live facial recognition, continuously matching passers-by against hot lists of roughly 250 individuals. An automated voice alerts staff on a match; staff review and relay information to police once an evidentiary threshold is judged met, including unsolicited tips. From early 2023 until April 2025 the police department received real-time facial recognition alerts from a network of over 5,000 cameras, more than 200 facial-recognition equipped, despite city restrictions. The department suspended the real-time alert partnership pending a legal compliance review; the camera network was later reopened to police while facial recognition alerts stayed suspended. The City Council has moved towards an ordinance that would make New Orleans the first US city to formally authorise this form of facial recognition policing. Recorded as paused to reflect that the camera access continues while the facial recognition alerting does not.

Source: NPR; ACLU of Louisiana · last checked 2026-08-20

Legal basis

City anti-violence initiative; commuters and civil rights groups argued mass subway scanning is impractical and unconstitutional

Oversight

Evolv has faced federal investigations into its marketing by the Federal Trade Commission and the Securities and Exchange Commission

A month-long pilot across 20 subway stations, beginning at Fulton Street on 26 July, reporting 2,749 scans. It found zero firearms, produced false positives and detected some knives. The department did not disclose how many passengers opted out or how many officers were needed to run the devices, and no agreement with Evolv had been finalised while the pilot outcomes were being assessed. Recorded as discontinued on the basis that the pilot completed with no continuing deployment established. LA Metro has run a comparable pilot at Union Station.

Source: Associated Press · last checked 2026-08-26

Legal basis

Internal force policy; no dedicated statutory framework for generative AI use

Oversight

House of Commons Home Affairs Committee inquiry

West Midlands Police switched off force-wide access to Microsoft Copilot in January 2026 after the tool generated false intelligence about a West Ham versus Maccabi Tel Aviv fixture, and associated disorder, that had never taken place. The fabricated information reached briefing material and shaped a risk assessment. Recorded as paused rather than discontinued. Acting Chief Constable Scott Green described the shutdown as temporary, in place until further notice, both in January and again in March 2026. No evidence of reinstatement had emerged as of August 2026, but a force describing a suspension as temporary is not the same as a decision to stop using a technology permanently, and this tracker records the difference. The force initially denied that AI had generated the erroneous intelligence; the Chief Constable subsequently corrected that evidence. The Home Affairs Committee reported in February 2026 that West Midlands Police failed to apply proper due diligence to false information generated by AI, and criticised what it called a remarkable lack of professional curiosity. This pause did not extend nationally. A Sky News survey published in February 2026 found at least 21 English forces still using Microsoft Copilot.

Source: House of Commons Home Affairs Committee correspondence, reported by ITV News · last checked 2026-08-26

Oversight

Reviewed by the Independent Police Conduct Authority, which reported on 30 November 2020.

Established by the Independent Police Conduct Authority's published outcome of 30 November 2020, an official oversight body's own record. The Authority found that New Zealand Police trialled Clearview AI facial recognition software between February and March 2020; that the trial was approved by an internal governance group but undertaken without approval from the Police Executive; and that there was no consultation with the Privacy Commissioner or the Government Chief Data Steward. The software 'was deemed not operationally viable'. Police subsequently established a central governance body for emerging technologies and released a policy in September 2020 governing how emergent technologies may be tested or trialled. The Authority's broader assessment was that Police use of surveillance technologies was generally conservative, well-considered and broadly compliant with privacy and data security principles. Status is discontinued rather than paused: the Authority records that the software was found not operationally viable and that the response was a new governance regime, not a suspension pending resumption. Nothing read in this run suggests any later Clearview AI use by New Zealand Police. This record is included because the withdrawal is the substance: it is the event that produced New Zealand Police's emergent technologies governance framework, the Expert Panel and the later facial recognition policy, all of which the two other New Zealand records in this changeset depend on. Fields left out and why. 'legalBasis' is omitted - the point of the finding is that the trial proceeded without the approvals it should have had. The number of images searched is not stated in the Authority's published outcome and is not asserted anywhere in this record.

Source: Independent Police Conduct Authority · last checked 2026-09-09