Sixteen police AI systems that were switched off, and why

Almost everything written about AI in policing is about adoption: a new system, a new force, a new capability. The other half of the story gets far less attention. This site's tracker currently records sixteen deployments that were discontinued or paused, and the reasons form a pattern worth setting out on its own.

None of these are guesses about why a contract quietly lapsed. Each is documented: a council vote, a regulator's ruling, an inspector general's report, or a police force's own admission. Read together, they group into four recognisable failure modes.

Oversight failures: the technology worked, the paperwork did not

Some of the clearest shutdowns had nothing to do with whether a system worked. Greece's Hellenic Police spent €4 million, 75% funded by the EU's Internal Security Fund, on 1,000 handheld biometric devices from Intracom Telecom. They were delivered fully functional in 2019. After a complaint from the civil liberties group Homo Digitalis, the Hellenic Data Protection Authority ruled the entire programme unlawful in December 2025, six years after the money was spent and the devices deployed.

Mountain View, California, discontinued its Flock Safety licence plate cameras after Police Chief Mike Canfield disclosed to the city council that federal agencies had gained access to the city's camera data without the city's knowledge. The council voted unanimously to end the contract.

New Orleans presents a variant worth noting precisely. The police department received real-time facial recognition alerts from a network of over 5,000 privately owned cameras, more than 200 of them facial-recognition equipped, from early 2023 until April 2025, despite the city's own restrictions on the technology. Camera access was later reopened; facial recognition alerting stayed suspended. That is recorded here as paused rather than discontinued, because the two halves of the arrangement ended differently.

Regulatory rulings: courts and commissioners found the legal basis missing

Clearview AI accounts for three of the sixteen shutdowns on its own. Canada's Office of the Privacy Commissioner found in 2021 that the RCMP's use of Clearview violated the federal Privacy Act, because the underlying image collection was itself unlawful. Australia's information commissioner reached a comparable finding about the Australian Federal Police's Centre to Counter Child Exploitation. Finland's National Bureau of Investigation ran roughly 120 Clearview searches in early 2020 while trying to identify victims in child sexual abuse material, producing one hit, before its data protection ombudsman found the processing unlawful.

Three different countries, three different regulators, the same underlying finding: a facial recognition search is only as lawful as the database behind it, and Clearview's database was built by scraping images without consent.

The evidence did not hold up when someone actually checked

Predictive policing produced the most consistent pattern of any technology in the record. The Los Angeles Police Department, an early co-developer of PredPol, dropped the system in 2020 after its own inspector general found it unclear whether the software had any effect on crime trends. Kent Police, the first UK force to trial PredPol, cancelled its contract after concluding it could not demonstrate the programme reduced crime. The Dutch national police's Crime Anticipation System, which divided cities into grid cells combining crime records with neighbourhood income and benefits data, met the same fate. Three police forces, three countries, the same conclusion reached independently: the predictions could not be shown to change outcomes.

VeriPol, deployed by Spain's National Police to flag robbery reports as potentially false using language analysis, was discontinued in 2024 after use had already fallen from roughly 84,000 reports analysed in 2020 to under 4,000 by 2022.

The most striking case is Bristol. A predictive system flagging children at risk of sexual and criminal exploitation, holding records on close to half a million residents, was stopped in 2023 after council staff found recent victims of sexual offences scoring below burglary perpetrators. When reviewers from Social Finance later tried to examine what had gone wrong, they could not: the source code and model variables could not be found. No record of the decision to stop was retained either. A predictive system with no recoverable model and no decision record is not simply a shutdown. It is a demonstration of what accountability looks like when it fails.

Pilots that ended because the pilot was the point

Several entries are less dramatic: a trial that ran its course and was not renewed. New York's Knightscope patrol robot in Times Square subway station was retired after four months, with the department confirming the pilot was complete. A month-long AI weapons detection trial across twenty New York subway stations, using Evolv scanners, reported 2,749 scans, zero firearms found, some false positives, and no agreement finalised with the vendor afterwards.

These are not failures in the way the regulatory and evidence-based cases are. They read more like a system doing exactly what a pilot is supposed to do: generating enough information to make a considered decision not to continue.

The one still unresolved

West Midlands Police switched off force-wide access to Microsoft Copilot in January 2026 after the tool generated false intelligence about a football fixture and associated disorder that had never happened. The fabricated information reached briefing material and shaped a risk assessment. The force initially denied that AI was responsible; the Chief Constable later corrected that account. The Home Affairs Committee found the force had failed to apply due diligence to the AI-generated information, and criticised what it called a remarkable lack of professional curiosity.

The shutdown has been described as temporary twice, in January and again in March, with no evidence of reinstatement as of this writing. That is recorded here as paused, not discontinued, because a force calling a suspension temporary is a different thing from a decision to stop. It is also worth noting this pause did not extend nationally: a survey published in February 2026 found at least 21 other English forces still using the same tool.

What the pattern shows

Four different failure modes, and four different fixes. Oversight failures are addressed by following the process that already existed. Regulatory rulings are addressed by not building a database on an unlawful footing in the first place. Evidence failures are addressed by measuring effect before committing years to a programme. Pilots ending on schedule need no fixing at all.

Only one of the sixteen was reversed by improving the technology, and even that is arguable: the West Midlands pause was a human response to an obvious factual error, not a model update. Governance, not better algorithms, is what ended most of the deployments on this list. The full record of all sixteen, including sourcing for each, is on the withdrawn deployments page.