{
  "name": "Police AI technologies",
  "generated": "2026-11-04T07:52:23.435Z",
  "count": 65,
  "technologies": [
    {
      "slug": "live-facial-recognition",
      "name": "Live Facial Recognition",
      "category": "Facial recognition",
      "summary": "Real-time scanning of every passing face against a police watchlist, operating at scale in the UK on force policy rather than statute, upheld by the High Court in April 2026 and now expanding to 50 vans.",
      "description": "Live facial recognition scans the faces of everyone passing a camera, converts each to a biometric template, and compares it in real time against a watchlist. Where the system alerts, an officer reviews the match and decides whether to intervene. Faces that do not match are, forces state, deleted immediately and automatically.\n\nIt is the most contested technology in this catalogue, and in the United Kingdom it now operates at scale with judicial approval and no dedicated statute.\n\nWHAT THE UK COURTS HAVE DECIDED\n\nTwo judgments frame current UK practice.\n\nBridges v Chief Constable of South Wales Police (Court of Appeal, 2020) found aspects of that force's deployment unlawful. It did not ban the technology. It required forces to tighten policies, address the public sector equality duty and engage with data protection law — while considering the broader context including potential racial bias and retention of innocent people's data.\n\nThompson v Metropolitan Police, decided 21 April 2026 by Lord Justice Holgate and Mrs Justice Farbey, is the first substantive test of what Bridges meant in practice. Shaun Thompson, a youth worker with an anti-knife-crime charity, was stopped outside London Bridge station in February 2024 after the system matched him — according to the court's media summary — against his brother, who was on a watchlist. Told he was \"a wanted man\", he asked what for and was told that was what officers were there to find out. His bank cards and passport did not resolve it; he refused to give fingerprints. He has described the experience as \"stop and search on steroids\".\n\nThompson and Big Brother Watch director Silkie Carlo, supported by the Equality and Human Rights Commission, did not argue that live facial recognition is unlawful in principle. Their argument was narrower and more structural: that the Met's policy leaves too much discretion to officers over where, why and against whom the technology is used.\n\nThe court dismissed the challenge, finding the Met's policy contains \"clear, interlocking and cumulative constraints\", and that the risk of racial discrimination was \"no more than faintly asserted\". Notably, it declined to consider hypotheticals, confining itself to the Met's actual policy. Thompson is appealing.\n\nTHE STRUCTURAL POINT THAT FOLLOWS\n\nWhat the court reviewed was a police force's own policy document. There is no UK statute governing live facial recognition. The lawfulness of the largest deployment in Europe currently rests on the adequacy of rules the deploying force wrote for itself, assessed by a court that expressly declined to consider circumstances other than those before it.\n\nThat is the single most important thing to understand about UK live facial recognition, and it holds regardless of one's view of the technology.\n\nTHE MET'S NUMBERS\n\nFigures come from different periods and are easy to conflate, so they are separated here.\n\nBy April 2026 the Met reported more than 2,100 arrests since the start of 2024, around a quarter involving violence against women and girls, and more than 100 sex offenders. Over the preceding year it reported more than three million faces scanned producing 12 false alerts, none of which led to an arrest, with every alert reviewed by a trained officer before action.\n\nFor September 2024 to September 2025 specifically, the force reported 962 arrests, and 10 false alerts. Of those ten people, eight were Black. Four were not stopped; the rest were spoken to for under five minutes. An earlier report recorded more than 1,400 arrests in total with over 1,000 charged or cautioned.\n\nA six-month static deployment in Croydon produced 173 arrests.\n\nCommissioner Sir Mark Rowley called the judgment \"a significant and important victory for public safety\", and has said the question is no longer whether to use the technology but why a force would choose not to. A Mayor's Office survey found 85% of respondents backed use to locate serious and violent criminals and those wanted by the courts.\n\nTHE CASE FOR IT\n\nThe arrest figures are substantial and the offence mix is not trivial: rape, strangulation, domestic abuse, child sexual offences, and offenders in breach of court-imposed conditions. These are people the Met says would otherwise have gone undetected.\n\nThe false alert rate, as the Met presents it, is genuinely low. Twelve false alerts against three million faces, none producing an arrest, is a better real-world error rate than most predictions about this technology assumed. Human review of every alert is a meaningful control and appears to be working as designed.\n\nPublic support, on the polling available, is high.\n\nTHE CASE AGAINST\n\nThe eight-of-ten figure deserves attention precisely because the Met published it. Ten false alerts is a small sample and no strong statistical inference can be drawn from it — but of the people wrongly flagged in that year, eight were Black, and that is the disparity the NPL threshold testing predicts at lower confidence settings. It sits uncomfortably with a court finding that discrimination risk was \"faintly asserted\".\n\nThe denominator argument cuts both ways. Three million faces scanned is presented as evidence of accuracy; it is also three million people whose biometrics were captured without suspicion, consent or any individualised justification. Counsel for Thompson argued that facial recognition data is comparable to a DNA profile, and that permanent installations would make it impossible for Londoners to move without routine biometric capture.\n\nThompson's own account illustrates what a low error rate means for the person inside it. He was compliant, carried identity documents, and still could not persuade officers the machine was wrong. That inversion — where documentary evidence is weighed against an algorithmic assertion — is the practical harm, and it is not captured by any accuracy statistic.\n\nNor is he isolated. A software engineer has announced plans to sue Thames Valley Police after being confused with a burglar.\n\nExpansion is proceeding regardless of the appeal. The Home Office announced 40 additional LFR vans, taking national deployment from around 10 to 50 across England and Wales.\n\nWHAT IS NOT ESTABLISHED\n\nWhether false alert figures are compiled consistently across forces, and whether a person spoken to but not arrested is counted, is not clear from published material.\n\nEthnicity data for false alerts is not routinely published by all forces, so the Met's disclosure cannot be compared nationally.\n\nNo published figure was found for how many of the 2,100 arrests resulted in charge, conviction, or a disposal that would not have occurred otherwise.\n\nThe outcome of Thompson's appeal is pending, and the Home Office consultation response on a facial recognition framework remained unpublished well past its due date as of August 2026.",
      "typical_vendor": "Typically NEC (NeoFace) in the UK context",
      "related_subject": "facial-recognition",
      "url": "https://www.policeainews.com/technologies/live-facial-recognition"
    },
    {
      "slug": "anpr",
      "name": "ANPR (Automatic Number Plate Recognition)",
      "category": "Vehicle tracking",
      "summary": "The UK's national number plate reading network — described by its own regulator as the largest non-military database in the country, capturing 75 to 80 million reads a day with no dedicated statutory framework.",
      "description": "Automatic number plate recognition reads vehicle registration plates from cameras, records the time and location, and stores the result. In the UK the reads are forwarded to a national database, giving police a searchable record of where vehicles have been.\n\nIts scale is the point, and it is routinely underappreciated because the technology is old and familiar.\n\nTHE SCALE\n\nIn October 2023 the outgoing Biometrics and Surveillance Camera Commissioner wrote to the Secretary of State for Transport setting out the position. His predecessor had described the system as having created the largest non-military database in the UK, with approximately 15,400 traffic lanes covered by cameras submitting between 75 and 80 million reads daily on a regular basis, and occasionally over 80 million. He noted the trajectory could reach 100 million reads a day.\n\nFor comparison, in 2010 the national data centre was taking data from around 10,500 cameras and recording 10 to 14 million plates a day.\n\nReads are held for up to twelve months.\n\nTHE GOVERNANCE GAP\n\nThe Commissioner's letter is unusually direct for an official document and is the most important source on this page. He wrote that he had highlighted in his final annual report that ANPR is now so vast and vital that the absence of its own regulatory and governance framework itself represents a significant risk; that his predecessor and others had raised these risks before his appointment; that he had seen no discernible mitigation over the course of his term; and that there was good reason to believe the risks had become more acute.\n\nHis annual report separately noted that the Independent Advisory Group for ANPR — comprising police, Home Office, academic and industry representatives — is neither a governance nor an oversight body, and contrasted this with the government's stated commitment to a strong legal framework for overt police surveillance.\n\nThe system operates under the National ANPR Standards for Policing and Law Enforcement, most recently updated in July 2026, which set technical and compliance requirements: reads must be located to within ten metres, fixed cameras must have recorded GPS coordinates, and all images must be linked to their read record and forwarded to national databases. It is subject to ICO regulation and to general data protection law, the Management of Police Information rules and the Surveillance Camera Code.\n\nThose are standards for how the system runs. They are not a statute authorising it.\n\nA COMPARISON WORTH DRAWING\n\nThis site records extensively the backlash against Flock Safety's plate reader network in the United States: over a hundred municipal cancellations, an FTC-adjacent controversy over misuse, and a company announcement in August 2026 introducing a seven-day default retention.\n\nThe UK national network retains reads for up to twelve months — roughly fifty times longer — across a far denser camera estate, and does so without a dedicated statute. Flock is not deployed in the UK, and British readers following that story may reasonably not realise the domestic position.\n\nTHE CASE FOR IT\n\nANPR is genuinely effective for the purposes it was built for. Uninsured and untaxed vehicles, stolen cars, vehicles linked to serious crime and vehicles of missing people are all found through it routinely, and the intercept rate for those categories is high.\n\nThe intrusion per read is also low in isolation. A plate is not a person, it is displayed publicly by legal requirement, and a single read records a vehicle in a public place at a moment in time.\n\nStandards exist and are published, which is more than can be said for several systems in this catalogue, and location accuracy requirements to within ten metres are a meaningful evidential control.\n\nTHE CASE AGAINST\n\nThe aggregate is a different proposition from the read. Eighty million reads a day, retained for a year, across 15,400 traffic lanes, produces a reconstructable movement history for most vehicles in the country — and therefore for most drivers, including everywhere they have worshipped, been treated, protested or slept.\n\nThe absence of a dedicated legal framework is the objection the regulator himself made, repeatedly and without effect. A system of this size operating on police policy rather than legislation means the rules governing it can change without parliamentary debate.\n\nThe oversight structure compounds it. The body that looks most like a governance mechanism is explicitly not one, by its own regulator's description.\n\nAnd retention is the sharpest practical issue. Twelve months is not an evidential requirement for the great majority of reads, which relate to vehicles of no interest to anyone; it is a decision to keep everything in case it becomes useful.\n\nWHAT IS NOT ESTABLISHED\n\nCurrent read volumes and camera counts are not published in real time; the figures above are from the Commissioner's 2023 correspondence and may have grown.\n\nHow many searches of the national database are conducted annually, by whom, and under what authorisation, is not published.\n\nWhether the risks the Commissioner identified have since been addressed by any government response is not established.\n\nNo published evaluation quantifies the contribution of the national network to case outcomes, as distinct from vehicle intercepts.",
      "typical_vendor": "National ANPR Service (UK); various in other jurisdictions",
      "related_subject": "anpr",
      "url": "https://www.policeainews.com/technologies/anpr"
    },
    {
      "slug": "hart",
      "name": "HART (Harm Assessment Risk Tool)",
      "category": "Predictive policing",
      "summary": "A machine learning tool used by Durham Constabulary to classify arrestees by reoffending risk and identify candidates for diversion from prosecution, discontinued in 2021.",
      "description": "The Harm Assessment Risk Tool was a machine learning model built by Durham Constabulary with the University of Cambridge, classifying arrestees as low, medium or high risk of reoffending. It is one of the few UK police AI systems to have been documented in academic literature by the people who built it, and it was discontinued in 2021.\n\nWhat it was actually for is frequently misdescribed, and the distinction matters more than almost any other detail on this page.\n\nWHAT IT WAS USED TO DECIDE\n\nHART was built as part of Checkpoint, a Durham programme offering an alternative to prosecution. Individuals classified as medium risk — likely to reoffend, but not in a seriously violent manner — qualified for a four-month intervention programme tailored to their needs, and were not charged if they completed it successfully.\n\nThe tool was therefore identifying candidates for diversion, not candidates for prosecution. A medium-risk classification led to an offer of support instead of a charge.\n\nTHE DELIBERATE ASYMMETRY\n\nHART was designed to err on the side of over-estimating risk. Its designers treated the two possible errors as unequal: classifying a genuinely high-risk person as low risk was considered more costly than the reverse, so the model was tuned accordingly.\n\nBig Brother Watch's criticism is that a model deliberately over-estimating risk — in effect, in its framing, the likelihood of guilt — is fundamentally incompatible with the rule of law and the presumption of innocence.\n\nThat argument is a serious one and is recorded here in full. But it sits awkwardly with what the classification was used for. If a medium-risk score routed someone to a diversion programme rather than a prosecution, then over-estimating risk meant offering more people an alternative to being charged, not charging more people. Whether an intervention programme is experienced as support or as a penalty imposed without conviction is a fair question — but it is a different question from the one the presumption-of-innocence framing implies, and this site records the tension rather than resolving it.\n\nTHE CASE FOR IT\n\nDurham did something unusual: it published. The model was the subject of a Cambridge thesis by a serving officer and of a peer-reviewed paper co-authored by academics and constabulary staff, examining accuracy and proposing a framework for proportionality in experimental police algorithms. Most forces in this catalogue have disclosed nothing comparable about systems still running.\n\nThe purpose was also, on its face, the least punitive use of risk scoring documented on this site. A tool whose output is an offer of diversion is a different proposition from one whose output is a bail refusal or a sentence.\n\nTHE CASE AGAINST\n\nMachine learning on police data carries the feedback problem documented elsewhere on this site: a model trained on prior police contact learns patterns of enforcement alongside patterns of offending, and the individuals it flags reflect who has been policed before.\n\nThe deliberate over-estimation, whatever its purpose, means the model was calibrated to be wrong in a particular direction, and a person classified medium risk on that basis has had a machine judgement made about their future conduct on the basis of an acknowledged skew.\n\nTransparency of the model does not equal transparency to the individual. Publishing a paper about a model is not the same as an arrested person understanding why they were classified as they were, or being able to contest it.\n\nWHAT IS NOT ESTABLISHED\n\nThe reason for discontinuation in 2021 is not established by the sources reviewed. This site records the date rather than the rationale, and a force explanation would be worth obtaining.\n\nWhether Checkpoint continued after HART ended, and if so what replaced the classification step, is not documented here.\n\nNo independent evaluation of HART's accuracy in operation, as distinct from the designers' own published assessment, has been identified.\n\nHow many people were classified by HART over its operational life, and how many were diverted as a result, is not published in the sources reviewed.",
      "typical_vendor": "Durham Constabulary, developed with the University of Cambridge",
      "related_subject": "predictive-policing",
      "url": "https://www.policeainews.com/technologies/hart"
    },
    {
      "slug": "microsoft-copilot",
      "name": "Microsoft 365 Copilot",
      "category": "Generative AI",
      "summary": "General-purpose generative AI embedded in the office software police forces already use, adopted by at least 21 English forces — the least policing-specific technology in this catalogue and, for that reason, among the least examined.",
      "description": "Microsoft 365 Copilot is not police technology. It is generative AI built into Word, Outlook, Teams and the rest of the Microsoft office suite: drafting documents, summarising meetings, searching across an organisation's own files and email.\n\nIt appears in this catalogue because at least 21 English police forces have deployed it, and because a general-purpose tool inside a police force operates on police material.\n\nWHY A GENERAL TOOL RAISES SPECIFIC QUESTIONS\n\nThe documents in a police force's Microsoft environment are not ordinary corporate documents. They include intelligence reports, victim and witness statements, case files, custody records and safeguarding referrals.\n\nA tool that summarises meetings is summarising strategy discussions about live investigations. A tool that searches across an organisation's files is searching material subject to disclosure obligations, data protection restrictions and, in some cases, legal privilege. A tool that drafts documents is drafting text that may end up in a case file.\n\nNone of that makes deployment wrong. It does mean the assurance questions differ from those a private company would ask, and that the absence of policing-specific scrutiny is itself notable — this site's tracker records at least one force having paused its deployment.\n\nTHE CASE FOR IT\n\nPolice forces run enormous administrative overheads, and officers consistently report paperwork as the largest drain on time available for policing. Generic productivity gains in drafting and summarisation are real and apply as much to a force as to any other large organisation.\n\nIt is also enterprise software from an established supplier with contractual data-handling commitments, deployed inside an existing environment. That is a different risk profile from bespoke police AI bought from a startup with no track record, which is much of what this catalogue contains.\n\nAnd unlike almost everything else here, it is not directed at the public. Its subject is the force's own paperwork.\n\nTHE CASE AGAINST\n\nThe absence of specificity cuts both ways. Because it is not a policing tool, it may not have been assessed as one — no equality impact assessment, no evaluation against policing use cases, no consideration of what happens when generated text enters an evidential document.\n\nGenerative AI produces fluent text regardless of accuracy, and this site's page on AI report writing documents what that means when the output is a police document. Copilot is not marketed for report writing, but nothing prevents an officer from using it that way, and the disclosure requirements now legislated in some US states for AI-drafted reports would not obviously be triggered by a general-purpose tool.\n\nAuditability is the practical problem. A dedicated report-writing tool logs that AI was used. A general assistant embedded in Word does not necessarily leave that trace, which makes it impossible to establish afterwards which documents were machine-drafted.\n\nSummarisation carries its own risk: a summary that omits a material detail from a long document is a failure mode that looks like success.\n\nWHAT IS NOT ESTABLISHED\n\nWhich forces have deployed it, and with what restrictions on the material it may access, is not systematically published.\n\nWhether any force has conducted a policing-specific impact assessment is not established.\n\nThe reason for the paused deployment recorded on this site's tracker is not documented.\n\nWhether outputs are logged as AI-assisted in any force is not established.",
      "typical_vendor": "Microsoft",
      "related_subject": "generative-ai",
      "url": "https://www.policeainews.com/technologies/microsoft-copilot"
    },
    {
      "slug": "facewatch",
      "name": "Facewatch facial recognition",
      "category": "Facial recognition",
      "summary": "Live facial recognition operated by shops rather than police, matching customers against retailer-compiled watchlists, and the clearest documented case of the human review safeguard failing alongside the algorithm.",
      "description": "Facewatch supplies live facial recognition to UK retailers. Cameras in a shop scan every customer entering, compare each face against a watchlist compiled from incidents reported by participating businesses, and alert staff when the system flags a match. Staff then decide whether to approach.\n\nIt belongs in a catalogue of policing AI despite not being police technology, because it performs a policing function — identifying suspected offenders in public-facing spaces and excluding them — without any of the statutory framework, oversight or appeal route that would attach if a police force did it.\n\nTHE NETWORK\n\nCustomers include B&M, Home Bargains, Sports Direct, Farmfoods, Spar, Southern Co-op, Costcutter, Budgens and Iceland. Sainsbury's expansion illustrates the pace: two shops in September 2025, six by February, more than 55 by early July, with around 150 more planned.\n\nVolume is substantial. Facewatch issued 297,433 alerts to British retailers between January and June, roughly 1,643 a day.\n\nWHAT THE COMPANY SAYS\n\nFacewatch states an accuracy rate of 99.98%. Its account of the process emphasises brevity and deletion: it takes less than a second to create a biometric image of a face, assess it against the bespoke watchlist, and automatically delete it where there is no match. The company maintains the technology deters crime and protects staff.\n\nRetailers report results. Sainsbury's cites a 46% drop in theft, aggression and antisocial behaviour where the cameras operate, and says more than 90% of people identified do not return.\n\nThose figures deserve one caution. Shoplifting offences recorded by police in England and Wales fell 4% to 507,086 in the year to March 2026 — the first fall in years — but police-recorded crime and incidents logged inside a retailer's own stores are different measures, and neither establishes what caused the other.\n\nTHE REGULATORY HISTORY\n\nThe Information Commissioner's Office investigated and concluded in March 2023 that Facewatch's data processing had breached data protection law across a number of principles, including lawfulness, fairness and transparency, purpose limitation, storage limitation, the processing of special category data, the processing of data relating to criminal offences, and the rights of children.\n\nFollowing private correspondence with the regulator, the company overhauled its practices. The ICO subsequently closed its formal inspection without requiring further action, satisfied that crime prevention is a legitimate purpose. It held that retailers may place individuals on a watchlist only where they are serious or repeated offenders.\n\nBig Brother Watch has noted that the Policing Minister at the time, Chris Philp, held private meetings with the company while it was under investigation and stated he would do everything possible to promote its widespread use.\n\nTHE FINDING THAT MATTERS MOST\n\nIn 2024 a nineteen-year-old shopping in a Home Bargains in Manchester was flagged as a suspected shoplifter. She was confronted by staff, subjected to a bag search, publicly removed from the store, and told she was barred from shops across the country using the same system. She had never stolen anything.\n\nIn subsequent correspondence, Facewatch admitted that both its technology and its human \"super-recogniser\" reviewer had produced the error.\n\nThat second part is the most consequential sentence on this page. Every live facial recognition deployment documented on this site — including police deployments upheld by the High Court — relies on human review of an alert as the safeguard against algorithmic error. Here is a documented case where the algorithm was wrong and the trained human reviewer confirmed the mistake. Human review is a mitigation, not a guarantee, and this is the clearest evidence of that available.\n\nShe is not the only case. Another woman was wrongly added to the shoplifter watchlist; a 2026 Guardian investigation spoke to multiple people wrongly identified, who reported struggling to understand why they had been approached, how their data was being used, or how to challenge it. A Sainsbury's store suspended its system after an error attributed to human mistake.\n\nTHE CASE FOR IT\n\nRetail crime is real, rising in most measures until recently, and frequently involves the same individuals repeatedly. Staff assaults are a genuine and underreported problem, and a system that warns staff that a person who previously threatened them has entered the shop is not a trivial benefit.\n\nNon-matching faces are deleted immediately, which is a narrower retention position than many systems here. And the ICO's requirement that watchlists be limited to serious or repeated offenders is, if followed, a meaningful constraint.\n\nTHE CASE AGAINST\n\nThe due process objection is the strongest. A person can be added to a watchlist on the say-so of shop staff, without being told, without a hearing, and without any finding by a court — and because reports travel across the network, the consequence is exclusion from many shops rather than one. There is no equivalent of a criminal standard of proof, no disclosure, and no obvious route of appeal.\n\nThe private-to-private nature is what makes this possible. A police watchlist attracts judicial review, data protection scrutiny and parliamentary attention. A retailer's watchlist has attracted one regulatory investigation, resolved through private correspondence.\n\nFor essential shops — supermarkets and food retailers — exclusion is not a trivial sanction. Being barred from B&M, Home Bargains, Iceland, Spar and Sainsbury's simultaneously is a significant restriction imposed without any judicial process.\n\nAnd the 99.98% figure needs reading at scale: at roughly 1,643 alerts a day, even a very low error rate produces wrongly accused people regularly.\n\nWHAT IS NOT ESTABLISHED\n\nHow many people are on Facewatch watchlists nationally, and how many have been added without prosecution, is not published.\n\nWhether the ICO's serious-or-repeated-offender restriction is complied with in practice is not independently audited.\n\nThe basis for the 99.98% accuracy claim — what was measured, on what population — is not published.\n\nNo figures were found for how many alerts result in confirmed identifications versus misidentifications, which is the ratio that would let anyone judge the system.",
      "typical_vendor": "Facewatch (United Kingdom)",
      "related_subject": "facial-recognition",
      "url": "https://www.policeainews.com/technologies/facewatch"
    },
    {
      "slug": "oifr",
      "name": "Operator-Initiated Facial Recognition (handheld)",
      "category": "Facial recognition",
      "summary": "Operator-initiated facial recognition: an officer photographs a person in front of them on a handheld device and checks the image against a police database, the fastest-spreading UK facial recognition mode and the least debated.",
      "description": "Operator-initiated facial recognition allows an officer to photograph someone they are already speaking to and check that image against a police database, usually to establish identity where the person cannot or will not provide it. It sits between the two better-known modes: unlike live facial recognition it does not scan crowds, and unlike retrospective search it happens in the moment, at the roadside or on the street.\n\nIt is spreading faster than either, and has attracted the least public argument.\n\nHOW IT WORKS AND WHO USES IT\n\nAn officer uses a phone or handheld device to capture an image, which is compared against a police image database. South Wales Police has used the mode since 2024. The Metropolitan Police announced its own rollout from December 2025.\n\nThe stated use case is narrow: confirming identity during an encounter that is already taking place, where a person is suspected of an offence, is unable to identify themselves, or is believed to be vulnerable or missing.\n\nTHE ACCURACY EVIDENCE\n\nThis mode has the strongest evidence base of the three, and it is worth stating plainly because it cuts against the general direction of criticism.\n\nThe National Physical Laboratory's operational evaluation of the NEC algorithm used by UK forces found, for operator-initiated and retrospective use, a true positive identification rate of 100% with no significant difference by gender or ethnicity. Even organisations opposed to police facial recognition have acknowledged this squarely, while arguing that accuracy is not the only question.\n\nThe reason is straightforward: a photograph taken deliberately, at close range, of a stationary and usually cooperative person is a far better input than a frame grabbed from a crowd at distance. Most of what degrades facial recognition accuracy is absent.\n\nTHE CASE FOR IT\n\nEstablishing identity is a routine and legitimate policing task, and the alternative where someone gives no name is often arrest for the purpose of identification. A handheld check that resolves the question in seconds can therefore mean fewer arrests, not more, and short-circuits a genuinely coercive process.\n\nThe accuracy evidence supports proportionate use, and unlike live facial recognition it does not process the biometrics of people who are not part of any encounter.\n\nTHE CASE AGAINST\n\nThe absence of debate is itself the concern. Live facial recognition has been litigated to the High Court; this mode arrived largely without argument, despite involving a biometric check on a specific named individual during a coercive encounter.\n\nThe governing framework is thinner than for live deployment. There is no statute, and the position rests on common law powers and force policy — the same structural problem this site's live facial recognition page describes, with less scrutiny attached.\n\nDiscretion is the practical issue. Whether to photograph someone is an individual officer's decision at the roadside, and the distributional question — who gets asked, and who does not — is the same one that attaches to stop and search, with no published data to answer it.\n\nThere is also a threshold question. Because the check is quick and easy, it can be used where identity is not genuinely in doubt, and a technology that lowers the cost of an action tends to increase how often it is taken.\n\nWHAT IS NOT ESTABLISHED\n\nNo published data was found on the number of operator-initiated checks conducted by any UK force, nor on the demographics of those checked.\n\nWhether an image captured during a check is retained, and for how long, is not consistently documented across forces.\n\nWhat authorisation, if any, is required before an officer runs a check is not published in the material reviewed.\n\nWhether the NPL's 100% finding, based on the evaluated algorithm version, holds for the versions and devices now in field use has not been re-established independently.",
      "typical_vendor": "Home Office-supported deployment across forces",
      "related_subject": "facial-recognition",
      "url": "https://www.policeainews.com/technologies/oifr"
    },
    {
      "slug": "retrospective-fr-pnd",
      "name": "Retrospective Facial Recognition (PND search)",
      "category": "Facial recognition",
      "summary": "Retrospective searching of the Police National Database's custody image collection, running over 25,000 searches a month against more than 19 million images, with a regulator-confirmed bias finding and a replacement algorithm planned.",
      "description": "Retrospective facial recognition against the Police National Database takes an image of an unknown person — from CCTV, a phone, a doorbell camera or social media — and searches it against custody photographs held nationally. It is the highest-volume facial recognition activity in UK policing and the least visible, because nothing happens in public.\n\nMore than 25,000 searches are run each month against a pool of over 19 million custody images.\n\nTHE BIAS FINDING\n\nIn December 2025 the National Physical Laboratory published an accuracy and equitability evaluation of the Cognitec algorithm used for these searches. The Information Commissioner's Office described the outcome without ambiguity in August 2026: bias in the algorithm increases the likelihood of incorrect matches for people in some demographic groups.\n\nThe response is recorded rather than the criticism alone. The Home Office and the National Police Chiefs' Council put mitigations in place — staff training, oversight reporting and equality impact assessments — alongside plans to replace the algorithm. The ICO continues to engage with HMICFRS, the Home Office and the NPCC on that replacement, and has reserved the right to take further regulatory action.\n\nThis is one of the few instances in this catalogue where an independent evaluation found a demographic disparity, a regulator stated it plainly, and the operator committed to replacing the system. That sequence is worth acknowledging even while the underlying finding stands.\n\nTHE CUSTODY IMAGE PROBLEM\n\nThe searchable pool is the long-standing controversy. Custody images are taken on arrest, and arrest is not conviction. Images of people never charged, or charged and acquitted, have historically remained in the collection, and the practical burden of removal has fallen on the individual to request it. Nineteen million images is a substantial fraction of the adult population of England and Wales.\n\nEvery retrospective search therefore runs against a gallery containing large numbers of people who were never convicted of anything.\n\nTHE LEGAL POSITION CHANGED IN 2026\n\nAccounts describing an absence of any statutory framework are now out of date. The Crime and Policing Act 2026 opened police access to DVLA driver records including photographs. In March 2026 the House of Lords rejected an amendment that would have blocked facial recognition searching of the roughly 50 million image DVLA database, by 123 votes to 40.\n\nThe substantive framework governing which non-police galleries may be searched has been deferred to a further Police Reform Bill. A Home Office consultation on the facial recognition framework closed on 12 February 2026 with a government response promised within twelve weeks; that response remained unpublished more than three months past its due date as of August 2026, with the Home Secretary pressed on the delay in the Lords in July.\n\nTHE CASE FOR IT\n\nThe investigative value is straightforward and well evidenced by forces. Without it, identifying a suspect from a CCTV still can take around fourteen days; with it, minutes. That is not a marginal efficiency in a volume crime investigation.\n\nIt is also the least intrusive of the three facial recognition modes in one specific sense: it processes an image police already hold in connection with an actual incident, rather than scanning uninvolved members of the public. The NPL evaluation found retrospective use performed at a 100% true positive identification rate with no significant demographic difference in its overall assessment — a finding that sits alongside, and in tension with, the ICO's statement about the algorithm's bias, and both are recorded here.\n\nTHE CASE AGAINST\n\nSearching a gallery of unconvicted people is the central objection, and it is not answered by algorithmic improvement. Retention of custody images of people never charged has been criticised for years without being resolved.\n\nThe DVLA expansion is a categorical change rather than an incremental one. A custody image collection contains people who have at least been arrested. A driving licence database contains people who have done nothing but apply to drive, and searching it makes the great majority of adults biometrically identifiable to police.\n\nInvisibility compounds both. Live facial recognition is at least apparent to those near a van. A retrospective search happens without the subject knowing, and there is no mechanism by which a person learns their image was searched.\n\nWHAT IS NOT ESTABLISHED\n\nWhether the algorithm replacement has begun, and what will replace it, is not established.\n\nHow many searches result in a correct identification, a wrongful one, or no match is not published.\n\nHow many of the 19 million custody images belong to people never convicted is not published.\n\nThe Home Office consultation response remained outstanding as of August 2026.",
      "typical_vendor": "Police National Database facility",
      "related_subject": "facial-recognition",
      "url": "https://www.policeainews.com/technologies/retrospective-fr-pnd"
    },
    {
      "slug": "flock-safety-alpr",
      "name": "Flock Safety ALPR",
      "category": "Vehicle tracking",
      "summary": "A national network of automatic number plate reader cameras sold to US police departments and private communities, searchable across jurisdictions, and now the subject of over a hundred municipal cancellations.",
      "description": "Flock Safety sells fixed cameras that photograph passing vehicles, read the registration plate, and record a \"vehicle fingerprint\" — make, colour, body type, roof rack, bumper stickers and other distinguishing features. Records are retained and made searchable, and crucially they are searchable across the network rather than only by the agency that installed the camera.\n\nThat last point is what distinguishes Flock from a conventional local plate reader deployment. A camera bought by one small town becomes, by default, part of a national searchable index. Flock reports use by around 5,000 US law enforcement agencies alongside private customers including homeowner associations and businesses.\n\nWHAT IT IS FOR\n\nThe strongest use case is a vehicle-based investigation with a partial lead: a witness recalls a white pickup with a ladder rack, and the vehicle fingerprint search returns candidates without a plate ever being read. Departments credit the network with recovering stolen vehicles, locating missing people including in medical emergencies, and identifying vehicles used in serious crime.\n\nCameras are typically solar-powered and quick to install, which is why adoption spread fast among small agencies without the budget or infrastructure for traditional systems.\n\nTHE REVERSAL\n\nThe direction of travel changed sharply in 2026. The civic organisation Deflock has tallied more than a hundred municipalities cancelling Flock contracts since 2021, more than half of those in 2026, and over a quarter in July 2026 alone. This is an advocacy tally rather than an audited count and should be read as such, but the trend it describes is corroborated by individual council decisions, including Chandler, Arizona, which announced in August 2026 that it would stop using the cameras and not renew.\n\nThe reasons cluster around misuse rather than malfunction. Reporting by the Washington Post found Flock cameras featured in 46 of at least 50 recently documented misuse cases. This site's tracker records searches run for purposes unconnected to any local investigation, including in jurisdictions whose own laws restrict such use.\n\nTHE AUGUST 2026 MEASURES, PRECISELY\n\nOn 13 August 2026 Flock announced accountability changes, and the detail matters more than the headline. Default retention was reduced to seven days — but as a recommendation and a new default rather than a cap, meaning customers can still choose otherwise. The previously optional abnormal-search monitoring tool, Audit Assistance, was to become a requirement for customers by the end of 2026 rather than immediately. Several of the changes depend on customer enforcement to have any effect.\n\nThe announcement came days before a National Week of Action Against ALPRs held from 16 to 22 August 2026.\n\nTHE CASE FOR IT\n\nVehicles are involved in a large share of serious crime, and a plate reader is a comparatively narrow instrument: it records a vehicle in a public place, not a person's identity, communications or associations. Compared with facial recognition, the intrusion per record is genuinely lower.\n\nCross-jurisdictional search has real investigative value. Vehicles cross municipal boundaries and offenders do not confine themselves to one police district, so a network that stops at a city line is less useful than one that does not.\n\nThe recovery and missing-person cases are not rhetorical; they are the routine daily use, and they are why the network spread as quickly as it did.\n\nTHE CASE AGAINST\n\nThe intrusion argument changes with scale. One camera records a vehicle passing a junction. Thousands of cameras, searchable together and retained, produce a reconstructable history of where a vehicle has been — which is a record of where its owner has been, and inferentially who they visited, worshipped with, or sought medical treatment from.\n\nThe cross-jurisdictional design is simultaneously the main benefit and the main governance problem. A community consenting to cameras in its own town has not consented to its data being searched by agencies elsewhere, under different rules, for purposes its own council may prohibit. The documented misuse cases are largely instances of exactly this.\n\nThe private-to-state pathway matters too. Cameras bought by homeowner associations and businesses feed the same searchable index, so the network grows through decisions no elector participates in.\n\nFinally, the company has moved beyond plate reading: it now also sells a gunshot detection product, Raven, with at least one documented municipal purchase. A single vendor accumulating multiple surveillance capabilities across thousands of small agencies is a procurement concentration worth tracking.\n\nWHAT IS NOT ESTABLISHED\n\nThe current number of active agencies and cameras is a company figure, not independently verified.\n\nThe Deflock cancellation tally is an advocacy count with no published audit methodology.\n\nWhether the August 2026 measures have been implemented, and whether the seven-day default has been widely retained rather than overridden, is not established.\n\nNo independent evaluation of the network's contribution to case outcomes, as opposed to counts of vehicles located, has been identified.",
      "typical_vendor": "Flock Safety",
      "related_subject": "anpr",
      "url": "https://www.policeainews.com/technologies/flock-safety-alpr"
    },
    {
      "slug": "clearview-ai",
      "name": "Clearview AI facial recognition",
      "category": "Facial recognition",
      "summary": "Facial recognition matching against a database of tens of billions of images scraped from the public internet without consent, sold to police and intelligence agencies, and found unlawful by regulators across Europe.",
      "description": "Clearview AI is the most heavily sanctioned technology company in this catalogue, and it is still operating. Its business is simple to describe: scrape photographs of faces from the public internet, convert each into a biometric code, and sell police, military and intelligence agencies the ability to upload an image of an unknown person and get back a name.\n\nThe company was founded in the United States and is led by Hoan Ton-That. What distinguishes it from every other facial recognition supplier here is the source of its gallery. NEC and Idemia match against custody images a force already holds. Clearview matches against the internet — Facebook photographs, LinkedIn headshots, Instagram selfies, images from news articles and personal blogs, collected from people who never encountered the criminal justice system and never consented.\n\nHOW THE COMPANY DESCRIBES IT\n\nClearview's own material states that its platform includes the largest known database of more than 50 billion facial images, sourced from public-only web sources including news media, mugshot websites, public social media and other open sources. The \"public-only\" framing is the core of its defence: that photographs already visible on the open web are fair to collect.\n\nIts stated use case is investigative lead generation. The company positions results as indicative rather than definitive, requiring further verification — a framing this site's tracker entry already records.\n\nTHE SIZE OF THE DATABASE\n\nFigures vary by source and by date, and the variation is itself informative. The UK regulator worked from around 20 billion images in 2022. Dutch and other regulators cited 30 billion in 2024. The privacy organisation noyb put it above 60 billion in 2025. Clearview's own site claims more than 50 billion. The database grows continuously, so no single figure is stable, but every source agrees the order of magnitude is tens of billions — several images for every person alive.\n\nTHE REGULATORY RECORD\n\nNo other system in this catalogue has been found unlawful by this many authorities.\n\nThe Dutch Data Protection Authority fined Clearview €30.5 million in September 2024, the largest GDPR penalty against the company, with a further non-compliance order of up to €5.1 million. Its reasoning was that Clearview should never have built the database at all: the biometric codes are, like fingerprints, biometric data whose collection is prohibited absent a statutory exception the company cannot rely on. The regulator also stated that use of Clearview's services within the Netherlands is itself prohibited, and warned that company directors can be held personally liable where they know the GDPR is being violated, have authority to stop it, and consciously accept the violations.\n\nItaly, Greece and France each imposed fines of around €20 million. Greece's, in July 2022, was described at the time as record-breaking. Authorities in Germany and Austria ruled the company's processing unlawful without financial penalty — the Hamburg commissioner concluding in January 2021 that the extracted hash constituted biometric data under Article 4(14) GDPR. Regulators variously ordered deletion of European data and cessation of collection.\n\nThe United Kingdom's case is the most legally interesting. The ICO fined Clearview £7.5 million in May 2022. In October 2023 the First-tier Tribunal set that aside, finding the processing fell outside the material scope of UK and EU data protection law because it was carried out exclusively in support of foreign governments. The ICO appealed, with Privacy International intervening. In 2025 the Upper Tribunal reinstated the fine. Information Commissioner John Edwards said the ruling upheld the regulator's ability to protect UK residents from having their images unlawfully scraped into a global database without their knowledge, and confirmed the regulator can act regardless of where the company is based.\n\nTaken together, European authorities have imposed roughly €100 million in penalties.\n\nTHE ENFORCEMENT PROBLEM\n\nThat total is the point at which the story becomes about regulation rather than about technology. Clearview is a US company with no establishment in Europe. The Dutch regulator acknowledged directly that compelling an American company without a European presence to obey the law has proven difficult. Reporting indicates the company has largely ignored the orders.\n\nIn October 2025 the privacy organisation noyb filed a criminal complaint, on the reasoning that administrative fines had failed. Max Schrems argued that cross-border criminal procedures are run for stolen bicycles, and asked why not for the personal data of billions of people.\n\nTHE CASE FOR IT\n\nThe investigative argument is real and should not be dismissed. A photograph of an unknown suspect, or of an unidentified child in abuse material, is a genuine dead end for an investigation, and a tool that generates a candidate identity where none existed can move a case that would otherwise stall. Several documented uses concern victim identification rather than suspect identification.\n\nThe company's stated discipline — that a result is an investigative lead requiring corroboration — is the correct one, and matches how forces using it publicly describe their practice.\n\nTHE CASE AGAINST\n\nThe objection is not primarily about accuracy. It is that the gallery was assembled without consent from people with no connection to any investigation, which is what every regulator that has examined it has concluded, and which no accuracy improvement addresses.\n\nThere is also a documented harm record. This site's tracker records wrongful arrests attributed to Clearview searches, including a Georgia man arrested over a Louisiana offence he had no connection to, and a case settled for $200,000. A caution that matters for accuracy: not every US facial recognition wrongful arrest is a Clearview case, and this site is careful not to attribute them all to one vendor.\n\nThe regulatory record raises a distinct question for any force using it. Where a supplier's database has been found unlawful by multiple European authorities, a force procuring its services is relying on evidence derived from processing that regulators in its own or neighbouring jurisdictions have declared illegal. That is a live problem for admissibility and for public confidence, quite separate from whether the match was correct.\n\nThe enforcement gap also sets a precedent visible to every other vendor: that a company operating from outside the EU can absorb roughly €100 million in fines and continue.\n\nWHAT IS NOT ESTABLISHED\n\nThe current database size is unverifiable independently; every figure originates with the company or with regulators relying on the company's disclosures.\n\nWhether any European deletion order has been complied with is not established.\n\nThe number of active police customers is unclear. The roughly 3,100 US agency figure recorded on this site's tracker is a historical company estimate, not a verified current count.\n\nWhether the Upper Tribunal's reinstatement of the UK fine has resulted in payment is not established by the sources reviewed.",
      "typical_vendor": "Clearview AI (United States)",
      "related_subject": "facial-recognition",
      "url": "https://www.policeainews.com/technologies/clearview-ai"
    },
    {
      "slug": "france-vsa",
      "name": "Algorithmic video surveillance (VSA)",
      "category": "Video analytics",
      "summary": "Algorithmic video surveillance authorised in France for the 2024 Olympics as a time-limited experiment, extended by legislation in 2026 to run until December 2030 — the clearest case of a temporary measure becoming durable.",
      "description": "Algorithmic video surveillance, or vidéosurveillance algorithmique, applies software to existing camera feeds to detect predefined events automatically: crowd density above a threshold, movement against the flow, an abandoned bag, a person on the ground, a fire, a weapon.\n\nFrance's implementation is the most legally documented example in Europe, and its trajectory is the reason it matters.\n\nWHAT IT DOES, AND WHAT IT DOES NOT\n\nThe French framework deliberately excludes biometric identification. The systems detect events, not identities: an algorithm flags that someone has fallen, not who they are. This distinction is central to the legal case made for it, and it is a real one — event detection and facial recognition are different capabilities with different intrusion profiles.\n\nThat said, the same cameras and the same feeds could support identification if the law permitted it, and the constraint is legal rather than technical.\n\nTHE EXPERIMENT THAT DID NOT END\n\nAlgorithmic video surveillance was authorised as a time-limited experiment tied to the 2024 Olympic and Paralympic Games. Experimental authorisation was the basis on which it was accepted: a defined event, a defined end date, and an evaluation.\n\nIt has since been extended. The loi Ripost passed the Senate on 26 May 2026 and went to a joint committee on 21 July 2026, extending the framework to December 2030.\n\nThat sequence — exceptional measure for an exceptional event, then extension — is a pattern this site records wherever it appears, because the political cost of authorising a surveillance capability is paid once, at the point of the exception, and extension attracts far less scrutiny than the original authorisation did.\n\nTHE CASE FOR IT\n\nEvent detection genuinely helps in dense crowds. A person who falls in a packed concourse may not be visible to any operator watching dozens of screens, and an alert that draws attention to that camera is a safety improvement rather than a surveillance one.\n\nThe exclusion of biometric identification is a meaningful legislative choice, not a technicality. France drew a line that several jurisdictions in this catalogue have not, and the systems operate within it.\n\nExperimental authorisation with an evaluation requirement is also, in principle, the right way to introduce a contested capability — better than deployment under general powers, which is how much of the technology on this site arrived.\n\nTHE CASE AGAINST\n\nThe extension undermines the framework that justified the original authorisation. An experiment that is extended before its evaluation has settled the question is not functioning as an experiment.\n\nEvent detection is defined by whatever the operator configures it to detect. The categories are set administratively, and a system that detects abandoned bags today can detect gathering crowds tomorrow without new hardware or, potentially, new legislation.\n\nThe infrastructure argument applies here as elsewhere: cameras and processing installed for event detection are the same cameras and processing that biometric identification would require, and the only thing separating the two is a legal prohibition that a future parliament can amend — as this one amended the end date.\n\nAnd the evidence base for effectiveness is thin. No published evaluation establishing that algorithmic detection improved outcomes during the Games, as against additional staffing or conventional monitoring, was identified.\n\nWHAT IS NOT ESTABLISHED\n\nWhether the mandated evaluation of the Olympic period was completed and published, and what it concluded, is not established here.\n\nWhich specific detection categories are currently authorised, and whether the list has changed since 2024, is not documented in the sources reviewed.\n\nHow many cameras currently run algorithmic analysis in France is not published.\n\nWhether any assessment has been made of false alert rates, and the operational burden they impose, is not established.",
      "typical_vendor": "Multiple contractors (France)",
      "related_subject": "surveillance",
      "url": "https://www.policeainews.com/technologies/france-vsa"
    },
    {
      "slug": "skynet-sharp-eyes",
      "name": "Skynet / Sharp Eyes",
      "category": "Surveillance",
      "summary": "China's overlapping national camera networks, in which Skynet blankets urban areas with facial-recognition-linked cameras and Sharp Eyes extends the same monitoring into villages by absorbing private cameras into the state network.",
      "description": "China does not operate a single surveillance network but several overlapping programmes built up over two decades, and untangling them is the first difficulty in describing them accurately.\n\nThe Golden Shield Project, from 2003, established the underlying infrastructure and is best known for internet censorship. Safe Cities, also from 2003, focused on urban public security, disaster warning and traffic management. Skynet, launched in 2005, installed cameras connected to facial recognition algorithms across roads, districts, schools and business centres. Sharp Eyes, set out in a five-year plan from around 2015, extended the same monitoring into rural towns and villages.\n\nWHAT MAKES SHARP EYES DISTINCT\n\nSharp Eyes is not simply more cameras. Its defining feature is the incorporation of privately owned cameras — those outside shops, apartment buildings and homes — into the state network, and the enlistment of residents as participants.\n\nIn some implementations, households were given television boxes allowing them to view live security footage from their area and press a button to summon police if they saw something amiss. The programme's target, set in the five-year plan, was 100% coverage of public spaces by 2020. Publicly available reporting does not confirm whether that target was met, though it suggests the country came close.\n\nReported implementation detail gives a sense of the granularity: one city was divided into core, key and auxiliary monitoring grids, with cameras oriented around what were described as the four basic needs of daily life — transportation, clothing, food and housing — feeding facial recognition and other algorithmic analysis back into the wider system.\n\nTHE NUMBERS, AND WHY THEY SHOULD BE TREATED CAREFULLY\n\nCamera figures for China are widely cited and poorly sourced, and this site would rather explain the problem than repeat a number confidently.\n\nState media reported Skynet operating around 20 million cameras across 16 cities and provinces. Other accounts put the network at 170 million by 2017. IHS Markit research reported 176 million cameras nationwide by 2016 with a projection tripling to 626 million by 2020. Estimates of more than 200 million public and private cameras across all programmes are common.\n\nThe difficulty is that these figures measure different things. Chinese coverage frequently aggregates Skynet, Safe Cities and Sharp Eyes into single totals, and, as China Digital Times has noted, clear lines between cameras with and without integrated facial recognition are not readily available or verifiable. Projections made in 2016 for 2020 are also not evidence of what was actually installed.\n\nSpending figures are similarly indicative rather than precise: central government expenditure on Sharp Eyes was reported at 3.1 billion yuan by 2017, with Safe Cities investment reported at around $140 billion. One frequently cited datapoint is that in 2018 the city of Zhoukou spent more on surveillance networks than on education.\n\nState media claims that Skynet can scan the entire Chinese population in one second with 99.8% accuracy. Georgetown's Center for Security and Emerging Technology has noted this claim ignores glaring technical limitations, and this site records it as a state claim rather than a capability.\n\nTHE SUPPLIERS\n\nThe systems run on computer vision from domestic firms including Hikvision, Dahua, SenseTime and Yitu — companies that also export. Dahua appears elsewhere on this site's tracker supplying vehicle-mounted facial recognition to police in Brazil, which is the practical reason a catalogue of policing AI needs to describe the Chinese domestic market rather than treat it as a separate subject.\n\nTHE CASE MADE FOR IT\n\nThe stated rationale is public security, traffic management and emergency response, and the Safe Cities framing in particular covers functions — disaster warning, traffic flow — that are uncontroversial in themselves and pursued by many countries.\n\nThe participatory element of Sharp Eyes is presented domestically as community involvement in local safety rather than as surveillance, and rural coverage as an equity measure extending to villages what cities already had.\n\nTHE CASE AGAINST\n\nThe programmes are instruments of social management, which is the stated purpose rather than a critical characterisation. The objective described in Chinese policy language is grassroots stability maintenance and the establishment of a hyper-stable structure through new technology.\n\nThe private-camera integration removes any meaningful boundary between what a person consents to when they install a doorbell camera and what the state can observe. Enlisting residents as monitors of their neighbours changes the social texture of a place in ways no camera count captures.\n\nThe most serious objection is that these systems provide the general infrastructure into which targeted repression is plugged. Sharp Eyes connects to systems used for population control in Xinjiang, and a network built for traffic management does not stay confined to traffic management once it exists.\n\nWHAT IS NOT ESTABLISHED\n\nNo verifiable current camera count exists for any individual programme, and figures conflating programmes should not be treated as measurements.\n\nWhat proportion of cameras have integrated facial recognition, as opposed to being conventional CCTV, is not established.\n\nWhether the 100% public space coverage target was achieved is not confirmed by any source reviewed.\n\nIndependent verification is structurally constrained. Entries on this site concerning China rely on human rights researchers, investigative journalism and satellite or procurement analysis working around significant state restrictions on information, rather than on official disclosure. That limitation is real and is stated rather than glossed over.",
      "typical_vendor": "Multiple domestic Chinese manufacturers",
      "related_subject": "surveillance",
      "url": "https://www.policeainews.com/technologies/skynet-sharp-eyes"
    },
    {
      "slug": "ijop",
      "name": "Integrated Joint Operations Platform (IJOP)",
      "category": "Surveillance",
      "summary": "The Integrated Joint Operations Platform, a big-data policing system in Xinjiang that aggregates personal data, flags people and behaviours as suspicious, and prompts investigations leading to detention.",
      "description": "The Integrated Joint Operations Platform is the system Chinese authorities use for mass surveillance in Xinjiang. It aggregates data about individuals, applies criteria to flag those it deems potentially threatening, and prompts officials to investigate them. Some of those flagged are detained and sent to political education camps and other facilities.\n\nAlmost everything known about how it works in practice comes from a single piece of research, and its provenance is worth stating plainly because it determines how much weight the detail carries.\n\nHOW IT BECAME KNOWN\n\nHuman Rights Watch first reported on the platform in February 2018. In early 2018 the organisation obtained a copy of a mobile application used by police in Xinjiang to communicate with the system, and reverse engineered it. The resulting report, published May 2019 and led by senior China researcher Maya Wang, describes what the system actually collects and what it treats as suspicious.\n\nThis is not a leak, a whistleblower account or an official disclosure. It is a technical analysis of the software officials use, which is a stronger evidential basis than most reporting on Chinese surveillance, and it is also a snapshot of one application at one point in time.\n\nWHAT THE ANALYSIS FOUND\n\nOfficials use the app for three broad functions: collecting personal information, reporting activities or circumstances deemed suspicious, and prompting investigations of people the system flags as problematic.\n\nThe significance lies in what counts as suspicious. Reported triggers include ordinary behaviours that would be unremarkable elsewhere — the use of WhatsApp, or of a VPN. The system does not primarily identify criminal conduct; it identifies deviation from expected patterns, and treats deviation itself as grounds for investigation.\n\nThat is a different proposition from every other system in this catalogue. A risk score in a Western criminal justice system predicts reoffending among people already arrested. This flags people who have done nothing that any legal system treats as an offence.\n\nREPORTED DEVELOPMENTS, AT LOWER CONFIDENCE\n\nAnalysis published in 2024 to 2026, attributed to Citizen Lab and the Australian Strategic Policy Institute and relayed through secondary sources, describes an upgraded platform integrating real-time payment data, machine learning models trained on historical detention data, connection to national databases so that a Xinjiang flag can trigger restrictions elsewhere, and links to the Sharp Eyes network in rural areas.\n\nThis site has not been able to examine the underlying reports directly, and records these as reported rather than established. The machine-learning claim is notable if accurate, because a model trained on who was previously detained would encode and reproduce the demographic profile of prior detentions — the feedback problem documented elsewhere on this site, in a context where the consequence is internment rather than a patrol allocation.\n\nIS THERE A CASE FOR IT\n\nThis is one of the few entries in this catalogue where the balance section cannot be filled in on the usual terms. The stated official rationale is counterterrorism and stability maintenance. No independent evidence has been identified that the system prevents violence, and its documented function — flagging lawful behaviour for investigation leading to extrajudicial detention — is not a policing purpose that has a legitimate counterpart to weigh.\n\nRecording that plainly is more honest than manufacturing symmetry. Where this site sets out competing arguments elsewhere, it is because both sides have a case that can be evidenced.\n\nWHAT IS NOT ESTABLISHED\n\nThe current operational state of the system is not independently verifiable.\n\nThe proportion of people flagged who are subsequently detained is not established.\n\nThe reported 2024-2026 upgrades have not been verified against primary sources by this site.\n\nIndependent verification is structurally constrained, as it is for all entries concerning China on this site, and rests on human rights researchers and investigative analysis working around significant state restrictions rather than on official disclosure.",
      "typical_vendor": "State-developed; hardware reportedly including Hikvision (CETC)",
      "related_subject": "surveillance",
      "url": "https://www.policeainews.com/technologies/ijop"
    },
    {
      "slug": "moscow-facial-recognition",
      "name": "Moscow facial recognition CCTV network",
      "category": "Facial recognition",
      "summary": "Moscow's city-wide facial recognition camera network, among the largest live identification systems anywhere, and documented as used to identify people attending political protests.",
      "description": "Moscow operates one of the largest municipal facial recognition camera networks in the world, integrated with the city's wider surveillance infrastructure and connected to police systems.\n\nIts significance in this catalogue is not technical. Comparable capability exists in several countries. What distinguishes Moscow is what the system has been documented as being used for.\n\nWHAT IT IS USED FOR\n\nThe stated purposes are conventional: locating wanted persons, finding missing people, and public safety in transport and public spaces.\n\nThe documented uses extend further. The system has been reported as used to identify people attending political protests and to locate individuals subsequently detained in connection with attendance — that is, identification of people engaged in activity that is not, in itself, criminal conduct in most legal systems, in order to bring consequences to bear on them for it.\n\nThat distinguishes it from live facial recognition as deployed in Western Europe, where watchlists are constructed from people wanted for specific offences and where the legality of the practice has been tested in court. It places the system closer to the population-control applications documented elsewhere on this site.\n\nTHE CASE MADE FOR IT\n\nThe public safety rationale offered is the same as elsewhere: a large city with a large transport network benefits from being able to locate wanted and missing people quickly, and integration with existing camera infrastructure is cheaper than building new capability.\n\nWhere such systems locate missing children or vulnerable adults, that benefit is real and is not diminished by other uses.\n\nTHE CASE AGAINST\n\nIdentification at protests is the central objection, and it does not require any argument about accuracy. A system that reliably identifies everyone attending a demonstration produces a chilling effect on assembly whether or not it ever misidentifies anyone, because the deterrent operates through correct identification rather than error.\n\nThe absence of effective legal challenge compounds it. In the UK, live facial recognition has been through the Court of Appeal and the High Court; in Italy and Greece, data protection authorities have blocked or ruled against systems. No comparable independent constraint has operated here.\n\nAccess is also a documented concern: reporting has indicated that access to Moscow camera footage has been obtainable through informal channels, which if accurate means the system's outputs are not confined to the authorities operating it.\n\nWHAT IS NOT ESTABLISHED\n\nCurrent camera counts and match volumes are not independently verifiable.\n\nThe algorithms and suppliers currently in use are not confirmed by any source this site has examined directly.\n\nReports of informal access to footage have not been independently verified here.\n\nAs with all entries on this site concerning states that restrict information, independent verification rests on human rights researchers and investigative journalism working around significant constraints rather than on official disclosure. That limitation is real and is stated rather than glossed over.",
      "typical_vendor": "Domestic and NtechLab-linked systems reported",
      "related_subject": "surveillance",
      "url": "https://www.policeainews.com/technologies/moscow-facial-recognition"
    },
    {
      "slug": "ai-bodycam-facial-recognition",
      "name": "AI Facial Recognition Body-Worn Cameras",
      "category": "Facial recognition",
      "summary": "Facial recognition running on an officer's body-worn camera, turning a recording device into a live identification system — proposed and trialled rather than widely deployed, and resisted by at least one major supplier.",
      "description": "Body-worn cameras were introduced as accountability devices: a record of what an officer did, reviewable afterwards. Adding facial recognition changes what they are. A camera that identifies people as it records is not a record of an encounter; it is an identification system worn on the chest of every officer on patrol.\n\nThe distinction matters because the accountability rationale that justified body cameras does not extend to identification, and the two capabilities pull in opposite directions on public trust.\n\nWHERE IT STANDS\n\nDeployments recorded on this site's tracker are listed below, with Edmonton Police Service the documented example and Corsight AI the named supplier.\n\nAdoption has been slower than the technology's availability would suggest, and one reason is supplier reluctance. Axon, the largest body-worn camera supplier, has previously declined to add face matching to its cameras following advice from its own ethics board, which concluded the technology was not accurate enough for that use and raised concerns about the consequences of error in an operational setting.\n\nThat is unusual and worth recording: a commercial supplier declining a capability its customers wanted, on ethical advice it had commissioned.\n\nTHE CASE FOR IT\n\nAn officer who cannot identify the person in front of them has limited options, and the alternatives — detention for identification, or letting a wanted person walk away — are both unattractive. Identification at the point of encounter is faster than either.\n\nThere is also a safety argument: knowing that the person being approached is wanted for a violent offence changes how an officer approaches them.\n\nAnd unlike live facial recognition mounted on a van, a body camera is already present at the encounter. The marginal intrusion, on this argument, is the matching rather than the recording.\n\nTHE CASE AGAINST\n\nThe marginal intrusion argument understates the change. A body camera records; a matching body camera identifies everyone in frame, which in a street encounter includes bystanders who are not part of the interaction and have no involvement in it.\n\nContinuous operation is the deeper issue. A van-mounted system operates at announced deployments for limited periods. A body camera is worn for an entire shift, and if matching runs continuously the officer becomes a mobile identification checkpoint wherever they walk.\n\nAccuracy conditions are also worse than for handheld operator-initiated recognition, which this site records as performing well. A body camera captures moving subjects at variable angles in variable light — the conditions under which facial recognition performs least reliably, and under which demographic disparities in error rates are most pronounced.\n\nThe accountability inversion should be stated plainly. Body cameras were introduced substantially in response to public demand for oversight of police conduct. Repurposing them as identification devices converts a technology the public asked for into one applied to the public.\n\nWHAT IS NOT ESTABLISHED\n\nHow many forces worldwide have deployed or trialled face matching on body cameras is not centrally recorded.\n\nWhether matching runs continuously or is officer-initiated differs by system and is not documented in most deployments.\n\nNo independent accuracy evaluation of facial recognition under body-camera conditions has been identified.\n\nWhether images of bystanders captured during matching are retained is not documented.",
      "typical_vendor": "Axon (hardware) with third-party facial recognition models, including Corsight AI",
      "related_subject": "body-cameras",
      "url": "https://www.policeainews.com/technologies/ai-bodycam-facial-recognition"
    },
    {
      "slug": "autonomous-patrol-vehicles",
      "name": "AI Autonomous Police Patrol Vehicles",
      "category": "Drones and robotics",
      "summary": "Driverless vehicles operating as mobile surveillance platforms on patrol routes, distinct from the humanoid and wheeled robots covered elsewhere — currently documented almost entirely in China.",
      "description": "Autonomous patrol vehicles are driverless road vehicles that follow patrol routes carrying cameras and sensors, streaming to a control centre. They are neither the wheeled site-patrol robots nor the legged hazardous-entry machines covered on this site's police ground robots page; they operate on roads, at vehicle scale, over larger areas.\n\nDeployments recorded on this site's tracker are listed below and are concentrated in China, where they form part of the wider surveillance infrastructure documented on this site's Skynet and Sharp Eyes page.\n\nWHAT DISTINGUISHES THEM\n\nCoverage is the difference. A fixed camera sees one location. A patrol robot covers a site. An autonomous vehicle covers a route, and can be redirected — which makes it closer in function to a mobile camera network than to either of the other robot categories.\n\nBecause they are vehicles, they can also carry substantially more sensing equipment than a robot constrained by battery and weight, including the plate recognition and facial recognition capabilities documented elsewhere on this site.\n\nTHE CASE FOR IT\n\nContinuous patrol coverage without an officer present is the obvious efficiency, and vehicles can operate on routes and at hours where staffing a patrol would be difficult to justify.\n\nThere is a safety argument in specific settings — routes where a lone officer would be at risk, or environments hazardous to people.\n\nAnd unlike an armed or intervening system, these platforms observe and report; the decision to act remains with a person.\n\nTHE CASE AGAINST\n\nMobility is what makes the surveillance argument different. A fixed camera is a known quantity in a known place, and residents can see where it is. A vehicle that patrols and can be redirected offers no such certainty, and its coverage is defined by whoever sets the route.\n\nThe context of the documented deployments matters. Where these vehicles operate as components of a national surveillance programme with the stated purpose of social management, their function is not separable from that programme.\n\nVerification is also constrained. As with other entries concerning states that restrict information, capability claims originate substantially with manufacturers and state media, and independent assessment is limited.\n\nWHAT IS NOT ESTABLISHED\n\nDeployment numbers are not independently verifiable.\n\nWhat sensing capabilities are fitted in operational units, as opposed to advertised, is not established.\n\nWhether any comparable deployment exists outside the jurisdictions recorded on this site's tracker is not documented.\n\nNo independent evaluation of effect on crime or public safety has been identified.",
      "typical_vendor": "Not disclosed in available reporting",
      "related_subject": "surveillance",
      "url": "https://www.policeainews.com/technologies/autonomous-patrol-vehicles"
    },
    {
      "slug": "briefcam",
      "name": "BriefCam Video Analytics",
      "category": "Video analytics",
      "summary": "A video analytics platform that makes existing CCTV footage searchable, using object detection, attribute filtering and an optional face recognition module, sold to police forces as a way to review hours of video in minutes.",
      "description": "BriefCam is a video analytics platform that sits on top of an organisation's existing camera network and makes recorded footage searchable. It does not replace cameras or a video management system; it processes what those already record. The company was founded in Israel out of research at the Hebrew University of Jerusalem, was acquired by Canon in 2018, and now sits within Milestone Systems, itself a Canon subsidiary. That corporate chain matters when reading older sources, which variously describe it as Israeli, Canon-owned or Milestone-owned, all of which have been true at different points.\n\nHOW THE COMPANY DESCRIBES IT\n\nBriefCam's own marketing centres on a patented technique it calls VIDEO SYNOPSIS, which the company describes as condensing hours of surveillance into a short summary by overlaying multiple events onto a single frame, each tagged with its original timestamp, so a reviewer can filter them by object type. Milestone's product material promises to let a user \"review hours of video in minutes\" and states the platform is camera-agnostic.\n\nThe platform is sold in three modules. REVIEW handles retrospective investigation: video synopsis, multi-camera search, appearance similarity, face recognition and filtering. RESPOND handles real-time alerting against configurable rules. RESEARCH produces aggregate statistics from video for operational analysis rather than investigation. Vendor documentation states the REVIEW module is included in every pricing plan because it is the analytics engine the others depend on.\n\nWHAT IT ACTUALLY DOES\n\nThe core capability is object detection and classification across recorded video, generating searchable metadata. An investigator can search across multiple cameras simultaneously for people or vehicles matching attributes: apparel, colour, size, speed, direction, dwell time, path. The company states advanced multi-camera search identifies men, women, children and vehicles of interest using face recognition, appearance similarity and those filters.\n\nFace recognition is a module within the platform rather than the whole of it. BriefCam's v6.0 release notes claimed a 17% improvement in matching faces \"in the wild\" and added support for matching faces wearing masks. Later releases added licence plate recognition, edge processing on Axis deep learning cameras, and cloud deployment options.\n\nDeployment models include standalone single-site installations, multi-site architectures with a central hub, and hybrid edge processing on compatible cameras. The vendor cites reduced total cost of ownership and operation in low-bandwidth environments as the rationale for the edge option.\n\nWHERE IT IS DEPLOYED IN POLICING\n\nDeployments recorded on this site's tracker are listed below. In summary: French police forces, Dutch police, the Brussels region in Belgium, and in the UK, West Midlands Police, with Cumbria Police also reported to hold a contract. The Czech Police names BriefCam among the tools it classifies as machine learning in its own freedom-of-information disclosure.\n\nThe Brussels deployment illustrates a common pattern: there, BriefCam functions as an analytics extension inside a Genetec Security Center video management deployment operated by a regional IT body serving six police zones, rather than as a standalone force procurement. The platform is typically specified alongside, and constrained by, an existing VMS.\n\nTHE CASE FOR IT\n\nThe efficiency argument is the strongest and the best evidenced. Reviewing footage manually is slow, and forces consistently describe video review as a bottleneck in volume crime investigation. Search that reduces hours of footage to a filtered shortlist addresses a real operational constraint rather than a hypothetical one.\n\nIt is also additive rather than replacive. Because it processes existing camera feeds and integrates with common VMS platforms, it does not require a camera estate to be replaced, which materially lowers the cost of adoption compared with a new sensor network.\n\nThe security posture is reasonably documented: the vendor states support for HTTPS and TLS, role-based access control, user authentication, audit logging, and deployment in network-segregated hardened environments. Audit logging in particular is what makes misuse detectable after the fact.\n\nTHE CASE AGAINST, AND WHAT TO PROBE\n\nThe face recognition module is the part that carries the regulatory weight, and it is easy to buy the platform without deciding how that module will be governed. A force adopting BriefCam for video search also acquires a facial recognition capability, and in most jurisdictions those attract different legal bases, different oversight expectations and different public reaction. The UK's Biometrics and Surveillance Camera Commissioner has said forces acquire powerful technologies without fully understanding the capabilities and implications, and has found no evidence of forces following the national decision-making model for buying new technology. That criticism applies directly to a modular platform where a capability can be enabled without a fresh procurement decision.\n\nAccuracy claims are vendor-supplied. The 17% matching improvement and the mask-matching capability come from BriefCam's own release material, not from independent evaluation, and this site has not identified a published independent accuracy or equitability assessment of BriefCam's face recognition comparable to the National Physical Laboratory work on other systems.\n\nAttribute search raises a distinct question from face matching. Searching for people by apparel, colour and appearance similarity does not identify anyone by name, but it does enable tracking an individual across a camera estate without ever establishing who they are, and without the legal threshold that a facial recognition search would trigger. Whether that is a lesser intrusion or simply a less regulated one is genuinely contested.\n\nDocumentation of outcomes is thin. This site has not identified published figures from any police deployment showing what proportion of searches produced an identification, a charge, or a conviction. Procurement decisions are consequently being made on efficiency claims rather than evidenced investigative outcomes.\n\nWHAT IS NOT ESTABLISHED\n\nSeveral things a full account of this platform would need are not on the public record, and are recorded here as gaps rather than left implied.\n\nWhich modules any given force has licensed is generally undisclosed. Because face recognition is a separately licensed component, knowing that a force uses BriefCam does not establish whether it uses facial recognition, and in most of the deployments recorded on this site that question is unanswered.\n\nNo independent accuracy or equitability evaluation of BriefCam's face recognition has been identified, comparable to the National Physical Laboratory's published assessment of the algorithm used for retrospective searches of the UK Police National Database. The accuracy figures in circulation originate with the vendor.\n\nRetention of generated metadata, as distinct from the source video, is not documented in any deployment reviewed. The platform's output is a searchable index describing who and what appeared where; whether that index is deleted when the underlying footage is, or persists beyond it, is a materially different question from CCTV retention and one no reviewed source answers.\n\nOutcome data is absent. No published figures were found from any police deployment showing what proportion of searches produced an identification, a charge or a conviction. Claims for the technology rest on time saved rather than on evidenced investigative results, and the distinction is worth holding onto: a tool can genuinely accelerate review without changing how many crimes are solved.",
      "typical_vendor": "BriefCam, a Milestone Systems company (Milestone is a Canon subsidiary)",
      "related_subject": "surveillance",
      "url": "https://www.policeainews.com/technologies/briefcam"
    },
    {
      "slug": "ai-report-writing",
      "name": "AI-Assisted Police Report Writing",
      "category": "Generative AI",
      "summary": "Generative AI that turns body-worn camera audio into the first draft of a police report, most widely as Axon's Draft One, and the subject of the first US state laws requiring disclosure that a report was machine-written.",
      "description": "Draft One transcribes the audio from an officer's body-worn camera and generates a first draft of the incident report. Axon launched it in April 2024 and describes it as its fastest-growing product. It is built on OpenAI's GPT-4, with what one account describes as the creativity setting turned down to keep output tied to what was actually recorded.\n\nIt matters more than most tools in this catalogue because of where police reports sit. A report is the account an officer gives of what happened and why they acted; it informs whether a prosecutor charges, whether a judge grants bail, and what an officer relies on when testifying, sometimes years later. Changing how that document is produced changes an input to the whole criminal process.\n\nAxon is not alone — Truleo sells a comparable product called Field Notes — but Axon's position matters: it is the largest supplier of body-worn cameras in the United States, supports more than 5,000 police departments, and Draft One integrates directly with the camera and evidence systems those departments already run.\n\nWHAT IT DOES, AND WHAT FORCES SAY ABOUT IT\n\nThe time saving is the whole pitch, and the figures are striking. Scottsdale Police, running a pilot, reported saving at least thirty minutes of officer time per report. In Oklahoma City, an hour-long search with a K-9 unit was transcribed and summarised into a report in eight seconds, against thirty to forty-five minutes to write by hand.\n\nFort Collins in Colorado offered the clearest adoption account: a technology sergeant tested it after a demo, said he was struck by the quality and speed, initially made it available to around seventy officers, and later extended it to all of them. Other named deployments include Lafayette in Indiana, Tampa in Florida, and Campbell in California.\n\nForces consistently describe human review as the control. A Scottsdale sergeant emphasised that getting details right matters for courts, records and prosecution, and that the tool is not bringing in content that was not there in the first place.\n\nWHERE IT HAS GONE WRONG\n\nThe most quoted failure is also the clearest illustration of the risk. In Heber City, Utah, an AI-generated police report stated that an officer turned into a frog partway through responding to an incident. Nothing in the audio supported that; the model produced it.\n\nA frog is obvious and gets caught. The concern is the class of error that is plausible enough not to be.\n\nThe more systematic problem is what investigative reporting found about how departments configure the tool. Axon builds in transparency features — disclaimers marking a report as AI-assisted, and settings requiring officer input before submission. Of seven departments responding to records requests, only South Jordan in Utah maintained mandatory officer input, and even there officers could bypass it. Departments in Lafayette and Fort Collins were reported to have intentionally turned the transparency defaults off.\n\nWorse, some police chiefs did not know which of their reports had been AI-drafted at all, because officers were copying the generated narrative into reports they then indicated they had written themselves — bypassing every disclaimer and audit trail the system provides.\n\nMost departments do not restrict use by offence type. South Jordan alone generated more than 900 AI-assisted reports across a range of incidents, including serious felonies.\n\nOne detail from Campbell, California, suggests the tool changes the documents themselves rather than merely producing them faster. The department's records supervisor emailed staff noting a significant difference in narrative format, and asked to be told if the District Attorney's office had comments about it.\n\nTHE REGULATORY RESPONSE\n\nThis is one of the few technologies in this catalogue where legislatures moved quickly.\n\nUtah's SB 180 requires that any police report created wholly or partly by generative AI carry a disclaimer saying so. On 10 October 2025 California became the second state to act, when Governor Newsom signed SB 524. That law goes further: it requires AI-assisted reports to be marked as such, requires agencies to keep an audit trail identifying who used the AI and which audio and video was used, and requires the first AI-generated draft to be retained for as long as the official report.\n\nSeparately, the King County prosecuting attorney's office in Washington State barred police from using AI to write reports at all — a prosecutor, rather than a legislature, declining to accept the output.\n\nTHE CASE FOR IT\n\nReport writing genuinely consumes a large share of officer time, and time spent typing is time not spent on other duties. If the saving is anywhere near thirty minutes per report, across a shift and a force, that is substantial.\n\nTranscription from a recording also has a real accuracy argument in its favour: a report written from body-camera audio hours later, from memory, is a worse record than one drafted from what the microphone actually captured. Used properly, the tool anchors the narrative to the recording.\n\nAxon's own disclaimer and officer-input features are the right controls to have built. The failure documented above is that departments switched them off, not that they were absent.\n\nTHE CASE AGAINST\n\nGenerative models produce fluent text regardless of whether it is accurate, and fluency is precisely what makes an error hard to spot in a document a reviewer expects to be routine. The frog is memorable because it is absurd; a subtly wrong sequence of events, or an inferred intention nobody stated, would read perfectly.\n\nThe bundling concern is real. Axon supplies the cameras, the evidence storage and now the report generator, and civil liberties groups have argued that bundling makes it easy for departments to acquire more technology than they need or than the public has agreed to.\n\nThere is also a training-data question that nobody has answered publicly: a model that learns from past police reports learns the conventions, framing and any bias embedded in them.\n\nFinally, the audit problem. Where transparency defaults are disabled and officers paste output into reports they sign as their own, it becomes impossible after the fact to establish which reports were machine-drafted — which forecloses any later study of whether the technology is accurate. Axon's own SEC filings acknowledge that AI failures could create operational and legal challenges and that datasets may carry bias.\n\nWHAT IS NOT ESTABLISHED\n\nAxon has declined to say how many departments use Draft One, so the true scale is unknown.\n\nNo independent evaluation of the accuracy of AI-drafted reports against officer-written ones has been identified.\n\nWhether the claimed thirty-minute saving holds outside vendor and department accounts has not been independently tested; the EFF has questioned it directly.\n\nHow prosecutors and courts treat an AI-assisted report where disclosure was not made is untested in any reported case reviewed here.",
      "typical_vendor": "Axon (Draft One); Truleo (Field Notes) and others",
      "related_subject": "generative-ai",
      "url": "https://www.policeainews.com/technologies/ai-report-writing"
    },
    {
      "slug": "shotspotter",
      "name": "AI Gunshot Detection (ShotSpotter)",
      "category": "Sensors and detection",
      "summary": "Acoustic sensors that detect suspected gunfire and alert police within seconds, and the subject of the most consequential body of independent evidence against any technology in this catalogue.",
      "description": "ShotSpotter places acoustic sensors across a coverage area, listens for sounds consistent with gunfire, and alerts police to a location within seconds. Sounds flagged by the algorithm are reviewed by a human analyst at a company review centre before an alert reaches a police department. The company rebranded from ShotSpotter to SoundThinking in 2023; the product name remains.\n\nThe premise is intuitive and worth stating fairly: gunfire is substantially underreported to emergency services, and a sensor that hears every shot can get officers moving where no one has called. An Illinois Criminal Justice Information Authority review found that adopting acoustic detection produced a sharp increase in dispatches to shots-fired incidents, which does suggest a real gap between gunfire that occurs and gunfire that is reported.\n\nHOW THE COMPANY DESCRIBES IT\n\nSoundThinking states an accuracy rate of 97% for identifying gunshots, based on the combination of its algorithm and analyst verification. It has credited the technology with substantial operational results in Chicago specifically, citing 125 lives saved over five years and the recovery of 2,985 firearms, figures it attributes to the police department.\n\nResponding to independent criticism, the company said the MacArthur Justice Center's findings \"relied heavily on a flawed analysis of incomplete data\", and that the city inspector general's conclusions had been \"misrepresented by critics to spread a false narrative\" about accuracy. Its chief executive, on Chicago's cancellation, emphasised the technology's continuing positive impact for residents.\n\nBoth positions are recorded here because the accuracy dispute turns on a distinction the two sides do not share. The company's 97% figure is about whether a detected sound was gunfire. The critical research is about what happened after officers arrived. Those can both be true at once.\n\nTHE INDEPENDENT EVIDENCE\n\nThis is the most substantial adverse evidence base for any technology on this site, and it comes from two sources examining the same city.\n\nThe MacArthur Justice Center at Northwestern University School of Law analysed Chicago Police Department data covering roughly 21 months, from July 2019 to April 2021. It found that 89% of alerts turned up no gun-related crime, and 86% led to no report of any crime at all — approximately 40,000 deployments over that period that produced nothing.\n\nChicago's Office of Inspector General reached a compatible conclusion in August 2021, reporting that ShotSpotter alerts rarely produce evidence of a gun-related crime, rarely give rise to investigatory stops, and even less frequently lead to recovery of gun-crime evidence during such a stop. It also found the technology changes how officers interact with the communities where sensors are concentrated, and campaigners cite it as increasing stop-and-frisk activity.\n\nWHAT HAPPENED IN CHICAGO\n\nChicago's contract history illustrates how these systems persist. A three-year, $33 million contract ended in 2021 and was quietly extended for two years at $9 million a year. Mayor Lightfoot's administration extended it again for $10 million while her successor was campaigning on ending it. The total value reached roughly $49 million.\n\nMayor Brandon Johnson announced cancellation in February 2024. The system continued through 22 September 2024 with a two-month transition, covering the Democratic National Convention. Houston, Atlanta and Cambridge, Massachusetts have also ended contracts with SoundThinking.\n\nCampaigners point to a December 2018 incident in which two Chicago officers, Conrad Gary and Eduardo Marmolejo, were killed responding to a ShotSpotter alert near railway tracks — an argument about the risks created by rapid dispatch to an unverified location.\n\nTHE CASE FOR IT\n\nThe underreporting problem is real, and the Illinois review supports it. If a substantial share of gunfire generates no call, a technology that closes that gap has a genuine rationale, and the fastest possible response to a shooting can matter for a victim's survival.\n\nThe human verification layer is also a meaningful design choice. An analyst reviews flagged sounds before an alert is dispatched, which is a check most automated detection systems do not have.\n\nTHE CASE AGAINST\n\nThe strongest objection is not that the sensors mishear. It is that even accepting the company's accuracy claim, the operational result documented in Chicago was tens of thousands of deployments producing no evidence of any crime. A technology can correctly identify gunfire and still fail as a policing intervention.\n\nThe distributional consequence follows from where sensors are placed. Coverage is concentrated in particular neighbourhoods, so the additional police presence, the investigatory stops and the changed pattern of interaction all fall on those neighbourhoods, whether or not crime is found.\n\nThere is also a structural limitation no configuration fixes: a gunshot detection system cannot alert anyone until a shot has been fired. In an open street a fast post-incident alert retains value; as a preventive measure it does not exist.\n\nFinally, the company now also owns the remains of Geolitica and HunchLab, giving one supplier both acoustic detection and place-based prediction — a consolidation worth watching given the evidence record of both product lines.\n\nWHAT IS NOT ESTABLISHED\n\nThe 125 lives saved and 2,985 firearms figures are attributed to the police department by the vendor; no independent verification was identified.\n\nThe MacArthur analysis covers one city over 21 months. Whether the hit rate is comparable elsewhere has not been tested at similar rigour.\n\nWhether cancellation changed gun crime outcomes in Chicago, Houston, Atlanta or Cambridge is not established by any source reviewed — a control the debate would benefit from.\n\nHow many agencies currently operate the system is not published.",
      "typical_vendor": "SoundThinking (formerly ShotSpotter, Inc.)",
      "related_subject": "surveillance",
      "url": "https://www.policeainews.com/technologies/shotspotter"
    },
    {
      "slug": "place-based-predictive-policing",
      "name": "Place-Based Predictive Policing",
      "category": "Predictive policing",
      "summary": "Forecasting where crime will occur so patrols can be directed there — the approach behind PredPol and its successors, and the one with the clearest independent evidence that it does not work as claimed.",
      "description": "Place-based predictive policing forecasts where and when crime is most likely to occur, usually as small boxes on a map, so patrols can be directed to those locations. It is distinguished from individual risk scoring, which forecasts what a named person will do; here the unit of prediction is a place.\n\nThe approach spread rapidly from around 2010 and has contracted sharply since. The leading product, PredPol — later Geolitica — ceased operations at the end of 2023 and is covered in detail on its own page.\n\nHOW IT WORKS\n\nModels ingest historical crime incident reports and identify spatial and temporal clustering. The best-known implementation drew on earthquake aftershock mathematics, on the theory that crime clusters in space and time in comparable ways: an offence in a location raises the near-term probability of another nearby.\n\nThe output is a patrol instruction, not an investigative lead. Officers are directed to be present in the predicted area during the predicted window.\n\nTHE EVIDENCE\n\nIndependent analysis has been unusually damaging. The Markup's examination of 23,631 predictions generated for one US department found a success rate below half a percent, and — more consequentially — that arrest rates in predicted areas were the same whether or not a prediction had been made. The full findings are set out on the PredPol page.\n\nAcoustic gunshot detection, covered separately, is a related but distinct proposition: it reports where a shot has already been fired rather than forecasting where one will be.\n\nTHE CASE FOR IT\n\nPatrol resources are finite and must be allocated somehow. The realistic alternative is officer intuition and habit, which carries its own well-documented biases and cannot be audited. A model at least applies consistent criteria and can be examined — as it was.\n\nPlace-based prediction is also a narrower intrusion in principle than scoring individuals. It directs attention to a location rather than attaching a risk label to a person, and nobody acquires a record because a model flagged their street.\n\nTHE CASE AGAINST\n\nThe feedback problem is structural rather than a defect of any particular product. A model trained on recorded crime learns where police have previously been, because recorded crime is a record of enforcement as much as of offending. Directing patrols there produces more recorded incidents there, which feeds back into the model. The scholar Ruha Benjamin characterised this as a crime production algorithm, and better engineering does not resolve it.\n\nThe distributional consequence follows: predictions concentrate in neighbourhoods already heavily policed, with the additional stops and interactions falling on those communities regardless of what is found.\n\nAnd the accuracy evidence, where it exists, is poor. A defence of prediction over intuition requires the prediction to be better than intuition, and the strongest independent analysis available found no measurable difference in outcomes.\n\nWHAT IS NOT ESTABLISHED\n\nWhether successor products from other vendors reproduce the same weaknesses is untested; no independent evaluation of the replacements has been identified.\n\nHow many forces worldwide currently operate place-based prediction is not centrally recorded.\n\nNo study identified compares crime outcomes in areas using place-based prediction against comparable areas that do not.",
      "typical_vendor": "PredPol / Geolitica (US, wound down); in-house systems including the Dutch Crime Anticipation System",
      "related_subject": "predictive-policing",
      "url": "https://www.policeainews.com/technologies/place-based-predictive-policing"
    },
    {
      "slug": "individual-risk-scoring",
      "name": "Individual Risk Scoring and Targeting",
      "category": "Predictive policing",
      "summary": "Systems that score named individuals by predicted risk of offending or victimisation, directing police attention to people rather than places — the most consequential and least evaluated category in this catalogue.",
      "description": "Individual risk scoring predicts what a named person will do. It is distinct from place-based prediction, which forecasts where crime will occur, and the difference matters: a place cannot be stigmatised, arrested or refused bail.\n\nThe category covers police force models scoring individuals for likelihood of offending or of becoming a victim, multi-agency systems combining police, council and social data, and central government research programmes.\n\nWHAT IS ACTUALLY DEPLOYED\n\nThis site's tracker records several variants. Avon and Somerset Police operate predictive analytics models scoring individuals. Bristol's Think Family Database combined data across agencies to identify families for intervention and was discontinued as not fit for purpose. New South Wales operated the Suspect Target Management Plan, which scored individuals for proactive policing attention and was discontinued after sustained criticism. Durham's HART, covered on its own page, classified arrestees for diversion.\n\nThe UK Ministry of Justice, with the Home Office, Greater Manchester Police and the Metropolitan Police, developed a model originally titled the Homicide Prediction Project and since renamed. A data sharing agreement covered between 100,000 and 500,000 people including names, dates of birth, ethnicity and gender. Documents obtained by Statewatch record it as completed in December 2024 but not deployed.\n\nWhere sources conflict, both positions are recorded: Statewatch states sensitive data including domestic abuse and self-harm information was used, covering people never convicted; officials deny this, stating only data from people with at least one conviction was used.\n\nTHE PATTERN WORTH NOTICING\n\nOf the individual-scoring systems documented on this site, a striking proportion have been discontinued: the Bristol database, NSW STMP, HART. Others were developed and never deployed.\n\nThat is a different trajectory from facial recognition or plate reading, which have expanded. It suggests individual risk scoring is harder to sustain in operation than in prospect — because the outputs are contestable, because the consequences attach to named people, and because a score is difficult to defend when the person asks why they are being visited.\n\nTHE CASE FOR IT\n\nPolice already make judgements about which individuals warrant attention, and those judgements are currently unstructured, undocumented and shaped by whoever happens to be in the room. A model at least applies consistent criteria and leaves a record.\n\nSome applications are protective rather than punitive. Scoring for risk of victimisation — in domestic abuse, child exploitation or serious youth violence — directs support rather than enforcement, and the tools on this site with the clearest social case are of that type.\n\nMulti-agency data combination also addresses a real failure. Serious case reviews repeatedly find that information held separately by police, health and social services would, if combined, have identified a child at risk.\n\nTHE CASE AGAINST\n\nThe feedback problem is more acute here than for place-based prediction. A model trained on prior police contact identifies people who have been policed before, and directing attention to them generates further contact, which confirms the score. The person cannot escape the classification by behaving lawfully, because the input is police attention rather than offending.\n\nConsequences attach without process. A person on a targeting list is not charged with anything, has no hearing, is generally not told, and has no route to challenge the classification.\n\nThe Ministry of Justice's own finding that OASys algorithms had greater predictive validity for white offenders than for offenders of Asian, Black and Mixed ethnicity applies to this category directly.\n\nAnd the protective framing does not neutralise the objection. A family identified for intervention by an algorithm experiences state attention it did not seek, on the basis of a calculation it cannot see.\n\nWHAT IS NOT ESTABLISHED\n\nHow many UK forces currently operate individual risk scoring, and under what models, is not centrally recorded.\n\nWhether individuals are ever informed they have been scored, in any deployment reviewed, is not established.\n\nNo independent evaluation comparing outcomes for scored individuals against comparable unscored individuals has been identified.\n\nThe reasons for discontinuation of several of these systems are documented only partially, which limits what can be learned from them.",
      "typical_vendor": "Various, including in-house police and council systems",
      "related_subject": "predictive-policing",
      "url": "https://www.policeainews.com/technologies/individual-risk-scoring"
    },
    {
      "slug": "offender-risk-assessment",
      "name": "OASys and Actuarial Reoffending Prediction",
      "category": "Courts and corrections",
      "summary": "The UK's national tool for assessing reoffending risk and criminogenic need, informing sentencing, parole and supervision decisions — and, importantly, actuarial rather than machine learning.",
      "description": "The Offender Assessment System is the structured assessment used across prisons and probation in England and Wales to evaluate an individual's likelihood of reoffending, their criminogenic needs and their risk of serious harm. It informs sentence planning, supervision intensity, and decisions about parole. The current automated version has been in operation since 2013, managed by the Ministry of Justice, with assessments collated centrally by the OASys Data, Evaluation and Analysis Team.\n\nA distinction this site tries to draw consistently applies here, and cuts against the assumption that everything on a police AI catalogue is AI. OASys is actuarial. Its algorithmic components — principally the Offender Group Reconviction Scale, which is also used independently for short pre-sentence reports — apply defined statistical weightings to recorded factors. They do not currently use machine learning. Published analysis notes there is some indication that machine learning methods are under consideration, and that the willingness to use them in policing risk tools suggests a possible direction of travel, but that is a prospect rather than a description of the current system.\n\nThat matters because criticisms of opaque, unexplainable models do not straightforwardly apply to a system whose weightings are documented, while criticisms about validity, disparity and over-reliance apply just as forcefully.\n\nHOW IT WORKS AND WHAT HAS BEEN BUILT ONTO IT\n\nAssessment combines actuarial prediction with structured professional judgement by the assessor. Alongside the general reoffending prediction sits a Risk of Serious Harm section allowing the assessor to identify factors specific to that construct, which is a different question from frequency of reoffending.\n\nSpecialised components have been added over time. A Sexual Predictor component, developed to predict contact sexual reoffending, was validated on a study of almost fifteen thousand individuals who had committed sexual offences. An offence-free time component allows estimates to be adjusted for those who have already spent periods in the community without reoffending, reflecting that risk is highest immediately after sentence.\n\nThe scale is substantial: in 2007/08 around 700,000 assessments were completed on almost 350,000 individuals.\n\nTHE EVIDENCE BASE\n\nOASys is, by the standards of this catalogue, unusually well studied. The Ministry of Justice published a compendium of research and analysis covering 2009 to 2013, examining predictive validity, assessor experience and the performance of individual components. Few systems on this site have anything comparable.\n\nThe most consequential finding for present purposes concerns disparity. Ministry of Justice analysis found that the predictive validity of the algorithms used in OASys was greater for white offenders than for offenders of Asian, Black and Mixed ethnicity. That is a government finding about a government tool, not an advocacy claim, and it is the reason this system connects directly to the wider fairness debate documented on this site's COMPAS page.\n\nTHE CASE FOR IT\n\nStructured assessment applied consistently is a defensible improvement on unaided judgement, which varies between assessors and is not auditable. Because OASys is actuarial and its components are documented and published, it can be scrutinised in ways that a proprietary model cannot — and it has been.\n\nThe design also retains the assessor. OASys informs a professional judgement rather than replacing it, and the Risk of Serious Harm section is explicitly a space for the assessor's own analysis.\n\nThe research programme deserves credit on its own terms. Publishing a multi-year compendium including validation of individual components, and analysis that surfaces uncomfortable findings about ethnic disparity, is more transparency than most operators of risk tools have offered.\n\nTHE CASE AGAINST\n\nDifferential predictive validity by ethnicity is the central problem. A tool that predicts less accurately for some groups produces assessments that are, on average, less reliable for those groups — and it informs decisions about liberty, supervision and parole.\n\nThere is a longer-standing critique that assessment tools of this kind were adopted and promoted by correctional services without robust evidence that they improve prediction of reoffending or the management of individuals. Some practitioners have argued the process displaces the relationship-building that effective supervision depends on, though this has itself been contested as mistaking the tool for a substitute rather than an aid.\n\nScale creates its own risk. When a tool is applied to hundreds of thousands of assessments a year, a score becomes the default frame through which a person is understood by every subsequent decision-maker, whatever the assessor originally intended.\n\nAnd the direction of travel deserves attention. If machine learning components are introduced, the documented-weighting transparency that currently distinguishes OASys from proprietary systems would be diminished, and that change should be visible rather than incremental.\n\nWHAT IS NOT ESTABLISHED\n\nWhether machine learning has since been introduced into any OASys component is not established by the sources reviewed.\n\nWhether the ethnic disparity in predictive validity has been addressed in subsequent revisions, and whether any fix has been independently evaluated, is not documented.\n\nHow often assessors depart from the actuarial score, and in which direction, is not published.\n\nCurrent annual assessment volumes are not available in the sources reviewed; the figures above are from 2007/08.",
      "typical_vendor": "UK Ministry of Justice and HM Prison and Probation Service; technical support from Capita",
      "related_subject": "courts-sentencing",
      "url": "https://www.policeainews.com/technologies/offender-risk-assessment"
    },
    {
      "slug": "us-criminal-justice-risk-assessment",
      "name": "COMPAS and US Pretrial Risk Assessment",
      "category": "Courts and corrections",
      "summary": "Actuarial tools scoring a defendant's likelihood of reoffending, used in US bail, sentencing and parole decisions, and the subject of the most consequential algorithmic fairness dispute yet conducted.",
      "description": "COMPAS — Correctional Offender Management Profiling for Alternative Sanctions — scores a defendant's likelihood of reoffending from factors including criminal history and personal circumstances. It was built by Northpointe, a company founded in the 1990s, which merged with two sister companies to become Equivant in January 2017. The scores inform bail, sentencing and parole decisions across many US jurisdictions.\n\nIt matters far beyond its own deployment because the argument about it produced the clearest public demonstration of something that turns out to be mathematically unavoidable in every risk-scoring system on this site.\n\nTHE PROPUBLICA FINDING\n\nIn May 2016 ProPublica published \"Machine Bias\", analysing COMPAS scores for more than 10,000 defendants in Broward County, Florida.\n\nIt found the tool correctly predicted reoffending about 61% of the time. More importantly, it found the errors fell unevenly: Black defendants were almost twice as likely as white defendants to be labelled higher risk and then not reoffend, while white defendants were considerably more likely to be labelled lower risk and then go on to offend. The mistakes ran in opposite directions by race.\n\nTHE COMPANY'S REBUTTAL\n\nNorthpointe responded with a 37-page technical document, and its argument was not a denial. The company accepted the overall accuracy figure of around 60%, and demonstrated — using ProPublica's own data — that this accuracy was equivalent for Black and white defendants. A defendant assigned a given risk score had roughly the same actual probability of reoffending regardless of race.\n\nIts criticisms of the analysis were specific: that ProPublica had not accounted for different underlying base rates of reoffending between the groups, and that combining the High and Medium categories into a single \"higher risk\" band inflated the apparent false positive rate.\n\nWHY BOTH SIDES ARE RIGHT\n\nThis is the part that generalises, and it is why this page matters more than its deployment count suggests.\n\nProPublica measured whether error rates were equal across groups — whether a Black defendant and a white defendant who both did not reoffend were equally likely to have been wrongly labelled high risk. Northpointe measured whether a given score meant the same thing for both groups — whether a high-risk label carried the same actual probability of reoffending.\n\nBoth are reasonable definitions of fairness. Both were satisfied by the evidence each side presented. And subsequent academic work established that where two groups have different underlying base rates, a scoring system cannot satisfy both simultaneously. It is not a flaw in COMPAS that better engineering removes. It is a property of the mathematics.\n\nThat has a direct consequence for every risk tool in this catalogue: a vendor claiming its system is fair is claiming something incomplete unless it says which definition of fairness it means, and what it therefore gives up.\n\nWHAT THE COURTS DID\n\nIn State v. Loomis, decided by the Wisconsin Supreme Court in July 2016, Eric Loomis challenged the use of a COMPAS score in his sentencing as a violation of due process, arguing he could not examine a proprietary algorithm. A judge had cited the assessment in describing him as high risk to the community before imposing eight and a half years.\n\nThe court rejected the challenge. Its reasoning included that identical COMPAS reports were available to both the defendant and the state — that is, neither side could inspect the algorithm, so neither was disadvantaged relative to the other. The court permitted use of the score while cautioning against reliance on it as a determining factor.\n\nTHE CASE FOR IT\n\nThe honest comparison is not between an algorithm and perfect justice; it is between an algorithm and unaided human judgement, which carries its own well-documented and less measurable biases. A structured tool applies the same factors to every defendant, which is more than can be said for intuition, and it can be audited — as it was.\n\nNothing in the ProPublica analysis showed the tool was worse than the judges it advised. That comparison has largely not been made.\n\nTHE CASE AGAINST\n\nSixty-one percent accuracy is the figure that should give most pause, and it is not disputed by either side. A tool that is right about three times in five is informing decisions about liberty.\n\nProprietary opacity remains unresolved. Loomis reasoned that equal access to the same limited information satisfied due process; it did not address whether anyone can meaningfully contest a score whose workings nobody outside the vendor can inspect.\n\nThe feedback problem applies here as it does to place-based prediction: a model trained on rearrest learns patterns of enforcement as much as patterns of offending.\n\nAnd there is a use question distinct from the accuracy question. Northpointe's founder later indicated that race-correlated factors may have been in play, and that he never intended COMPAS to be the sole basis for a decision. A Broward County judge who oversaw most pretrial hearings said he had stopped relying on it years earlier, preferring his own judgement. How a score is weighed sits outside the vendor's control and is rarely documented.\n\nWHAT IS NOT ESTABLISHED\n\nHow many jurisdictions currently use COMPAS or comparable tools, and in what decisions, is not centrally published.\n\nNo study identified compares outcomes under COMPAS-informed decisions against unaided judicial decisions on equivalent cases, which is the comparison that would actually answer whether it helps.\n\nWhether the tool has been revised since 2016 in response to the findings, and whether any revision has been independently evaluated, is not established.\n\nHow often judges depart from a score, and in which direction, is not systematically recorded.",
      "typical_vendor": "Equivant (formerly Northpointe) for COMPAS; Arnold Ventures for the Public Safety Assessment",
      "related_subject": "courts-sentencing",
      "url": "https://www.policeainews.com/technologies/us-criminal-justice-risk-assessment"
    },
    {
      "slug": "prison-ai-systems",
      "name": "AI in Prisons and Custody",
      "category": "Courts and corrections",
      "summary": "AI applied inside prisons and custody, from retrospective analysis of seized devices to continuous monitoring of communications and behaviour — a setting where the usual consent and proportionality arguments work differently.",
      "description": "Prisons are an unusual environment for the technologies in this catalogue. Monitoring that would be plainly disproportionate in public is lawful and expected inside a custodial institution; people held there have reduced privacy rights by design, and are told so.\n\nThat makes the interesting question not whether monitoring is permissible, but where its limits sit, and who else it captures.\n\nTWO DISTINCT MODELS\n\nThe first is retrospective analysis of seized material. HM Prison and Probation Service uses AI language analysis on phones seized from prisoners — devices already unlawfully held, examined after seizure, in connection with a specific investigation. This is closest to conventional digital forensics: a bounded examination of a specific item.\n\nThe second is continuous monitoring of live infrastructure at population scale, covered in detail on this site's page on AI prison communications monitoring. Every call transcribed, every conversation analysed, extending in the current generation into voice identification and video behaviour analysis.\n\nThe distinction matters more than the shared setting. A seized device belongs to the person it was taken from. A monitored telephone system captures both parties, and the person outside is not in custody.\n\nWIDER APPLICATIONS\n\nBeyond communications, AI is applied in custodial settings to automated headcounts, behaviour detection and real-time alerting from CCTV, and to risk assessment for categorisation, parole and release planning — the latter covered by this site's offender risk assessment page.\n\nTHE CASE FOR IT\n\nPrisons are dangerous places for the people held in them and for staff. Violence, coercion, drug supply and self-harm are persistent problems, and staffing levels in many systems are inadequate to observe what is happening.\n\nMonitoring is also lawful, disclosed and long established. All parties to a prison telephone call are notified it is monitored, and courts in multiple jurisdictions have upheld wide latitude where the purpose is safety and crime prevention.\n\nSome applications are unambiguously protective: identifying a person at risk of self-harm, or detecting that someone is being coerced by another prisoner, are uses where automation genuinely extends what limited staff can observe.\n\nTHE CASE AGAINST\n\nThe people on the other end of the call are the central problem. They have been convicted of nothing, have no realistic alternative means of contact with their relative, and cannot consent in any meaningful sense. Continuous voice identification, as now marketed, builds biometric profiles of them specifically.\n\nScope creep in this setting is documented rather than hypothetical. An investigation of one widely used system found calls flagged for content unrelated to any security concern, and a vendor pitch citing the value of monitoring in defending an institution against litigation brought by the people it holds.\n\nContact with family is also one of the strongest predictors of successful resettlement. A monitoring regime experienced as total may reduce the contact that reduces reoffending, which sets a security measure against a rehabilitative one.\n\nAnd the environment weakens the usual safeguards. A person in custody who believes a system has flagged them wrongly has less practical ability to contest it than almost anyone else in this catalogue.\n\nWHAT IS NOT ESTABLISHED\n\nThe scale of AI use across UK custodial estates is not published.\n\nWhether outside parties to monitored calls are informed that voice profiles may be built from their speech is not established in any deployment reviewed.\n\nNo evaluation of whether communications monitoring reduces violence or contraband, as opposed to generating intelligence, has been identified.\n\nRetention periods for transcripts, voice profiles and behavioural analysis outputs are not documented.",
      "typical_vendor": "Not publicly specified",
      "related_subject": "prisons-probation",
      "url": "https://www.policeainews.com/technologies/prison-ai-systems"
    },
    {
      "slug": "police-data-integration-platforms",
      "name": "Police Data Integration and Link Analysis Platforms",
      "category": "Investigative analytics",
      "summary": "Software that merges separate police databases into one searchable environment and maps relationships between people, places and events — constrained in Germany by a Federal Constitutional Court judgment, and unconstrained in the United States.",
      "description": "Data integration platforms do not collect anything. They connect what a police force already holds — incident reports, custody records, intelligence, vehicle data, communications data — into a single searchable environment, and surface relationships between people, places, vehicles and events that would be invisible while the data sat in separate systems.\n\nPalantir's Gotham is the best-known product. The same vendor's immigration enforcement systems are covered separately on this site, and the contrast between the two is the most instructive thing about this category.\n\nWHAT IT DOES\n\nThe core capability is link analysis. Given a name, the platform returns not only that person's records but their associations: who they have been stopped with, which addresses connect them to other individuals, which vehicles recur. Analysts describe it as replacing weeks of manual cross-referencing.\n\nThe value proposition is genuine. Police forces routinely hold information in systems that cannot talk to each other, and failures to connect existing intelligence feature in a great many serious case reviews.\n\nTHE GERMAN CONSTRAINT\n\nGermany is where this technology has been most seriously tested against law, and the outcome is the reason this page exists.\n\nThe Federal Constitutional Court has constrained automated data analysis by police, holding that merging databases and applying automated analysis to them constitutes a distinct and serious interference with informational self-determination — separate from the collection of the underlying data, and requiring its own justification and threshold.\n\nThe reasoning matters beyond Germany. It rejects the argument that integration is neutral because no new data is gathered. Combining lawfully held datasets creates a capability that did not previously exist, and the court treated that capability as the thing requiring authorisation.\n\nTHE CONTRAST\n\nThe same vendor's platform operates in the United States with essentially the inverse legal position. No court has constrained it, the contracts have generally been awarded sole-source rather than competitively, and in the immigration context the output is an operational target list rather than an investigative link chart.\n\nTwo jurisdictions, one vendor, broadly one capability, and diametrically different governance. That comparison is more informative about the state of police AI regulation than any amount of description of the software.\n\nTHE CASE FOR IT\n\nFragmented data is a real and documented cause of police failure. Where information sufficient to prevent harm existed but sat unconnected across systems, an integration platform addresses precisely that.\n\nLink analysis also produces auditable outputs. A chart showing why the system connected two people is inspectable in a way an opaque risk score is not.\n\nAnd unlike most systems in this catalogue, the platform adds no new collection. Every input was already lawfully held.\n\nTHE CASE AGAINST\n\nThe German court's reasoning is the strongest argument against, and it is a judicial finding rather than an advocacy position: integration is not neutral, and the capability it creates requires its own justification.\n\nAssociation is not evidence. A link chart shows that two people were stopped together; it does not show why, and a person can acquire a prominent position in a network diagram through proximity rather than conduct.\n\nScope is the practical problem. Once every dataset is searchable together, the constraints on what may be searched become policy settings rather than technical limits — the same structural issue this site records for border biometric databases.\n\nProcurement concentration compounds it. Where one vendor supplies the integration layer across many agencies, the analytical frame those agencies use is shaped by a single commercial supplier.\n\nWHAT IS NOT ESTABLISHED\n\nWhich German state forces currently operate the platform, and under what post-judgment authorisation regime, is not established here.\n\nWhat audit trail exists for a link-analysis query, and whether authorisation is required before running one, is not documented in the deployments reviewed.\n\nNo independent evaluation of whether integration platforms improve case outcomes has been identified.",
      "typical_vendor": "Palantir Technologies (Gotham); others including Microsoft and IBM",
      "related_subject": "surveillance",
      "url": "https://www.policeainews.com/technologies/police-data-integration-platforms"
    },
    {
      "slug": "ai-emergency-call-triage",
      "name": "AI Emergency and Non-Emergency Call Triage",
      "category": "Control rooms",
      "summary": "AI layered over emergency call handling to transcribe, translate, detect duplicate reports and prioritise calls — and, in at least one US city, to handle non-emergency calls end to end without a human.",
      "description": "AI in emergency call handling covers a spectrum, and the distinction between its ends is the most important thing on this page.\n\nAt one end, AI assists a human call-taker: transcribing live, translating in real time, extracting addresses and incident details, and identifying when several callers are reporting the same incident. The human retains every decision.\n\nAt the other end, AI handles a call from arrival to completion with no human involved.\n\nTHE MARKET CONSOLIDATED\n\nAxon acquired Carbyne in a deal announced on 4 November 2025 for $625 million, closing in the first quarter of 2026, having acquired Prepared in late 2025. The two now anchor a combined product line branded Axon 911. Axon states the pair work with over 1,000 agencies processing more than 80 million emergency calls a year.\n\nThat places emergency call handling, real-time crime centres, body cameras, plate recognition and AI report writing with one supplier — a procurement concentration this site notes wherever it appears.\n\nWHAT THE ASSISTIVE VERSION DOES\n\nDuplicate detection is the clearest use case. Orleans Parish Communication District uses AI to identify calls relating to an incident already known to the centre — several callers reporting the same road collision — so human call-takers can prioritise calls needing intervention. Carbyne and the district report more than 3,500 events processed over a 90-day period and redundant calls reduced by more than 30% among those triaged. These are vendor and customer figures rather than independently validated measures.\n\nPrepared's layer transcribes calls live, translates between languages using a synthetic voice so a dispatcher can converse with a speaker of another language without conferencing an interpreter, and extracts addresses and incident details automatically.\n\nTHE AUTONOMOUS VERSION\n\nTrade reporting in August 2026, citing Axon's 911-division chief executive, describes Fort Worth, Texas using the platform to handle non-emergency calls fully autonomously — thousands a month, from call arrival through to completing the call for service, with no human call-taker. The executive stated non-emergency wait times drop to zero and that agencies can offload more than a third of call volume without additional headcount.\n\nThis claim is vendor-sourced — a company executive in trade press, not a Fort Worth city or emergency-centre source — and is recorded with that caveat rather than as established fact. If independently confirmed it would be the first fully automated call-handling deployment identified in this research.\n\nTHE UK CONTRAST\n\nIt is worth setting against the stated UK position. This site's tracker records the UK government as not supporting police use of AI to answer 999 calls, with AI triage confined to 101 non-emergency routing. Two jurisdictions have reached opposite conclusions about the same capability.\n\nTHE CASE FOR IT\n\nEmergency call centres are chronically understaffed and call-taker turnover is high. Anything that reduces time spent on duplicate calls or on arranging interpretation is time returned to calls that need a person.\n\nReal-time translation is a genuine equity improvement. A caller who does not speak the local language currently waits for a third-party interpreter to be conferenced in, during an emergency.\n\nDuplicate detection is also low-risk: the failure mode is a duplicate reaching a human, which is the status quo.\n\nTHE CASE AGAINST\n\nThe autonomous end is a different proposition entirely, and the distinction is easily lost when both are described as AI call triage. A system that decides a call needs no human has made a consequential judgement, and calls categorised as non-emergency are not always non-emergency — the classification is the risk.\n\nVendor-sourced performance claims dominate this category. The duplicate-reduction figures, the wait-time claims and the volume-offload estimates all originate with the supplier or its customer, and no independent evaluation of any of them was identified.\n\nConsolidation matters here more than in most categories, because a single supplier now handles a large share of US emergency call infrastructure alongside the response equipment.\n\nWHAT IS NOT ESTABLISHED\n\nThe Fort Worth autonomous handling claim has not been confirmed by any source independent of the vendor.\n\nWhat happens when an autonomous system misclassifies an emergency as non-emergency, and whether any such case has occurred, is not documented.\n\nNo independent evaluation of duplicate-detection accuracy has been identified.\n\nWhether callers are informed they are speaking to an automated system is not established.",
      "typical_vendor": "Axon (Carbyne and Prepared, combined as Axon 911); others",
      "related_subject": "emergency-response",
      "url": "https://www.policeainews.com/technologies/ai-emergency-call-triage"
    },
    {
      "slug": "ai-call-content-analysis",
      "name": "AI Analysis of Control Room Call Content",
      "category": "Control rooms",
      "summary": "AI applied to the content of calls already handled by a control room — analysing what was said after the fact to identify risk, missed indicators and patterns across incidents, rather than routing calls in real time.",
      "description": "This is distinct from the call triage covered elsewhere on this site. Triage acts on a call as it arrives — routing it, detecting duplicates, prioritising response. Content analysis examines what was actually said, generally after the call has ended, to find things a call-handler working at speed could not.\n\nThe clearest documented application is retrospective review of calls involving domestic abuse and vulnerability, looking for indicators of escalating risk that were present in the words used but not acted on at the time.\n\nDeployments recorded on this site's tracker are listed below.\n\nTHE CASE FOR IT\n\nThe rationale is grounded in a well-documented failure. Serious case reviews and domestic homicide reviews repeatedly find that warning signs were present in earlier contacts with police — often in what a caller said — and were not recognised because each call was handled in isolation by a different person under time pressure.\n\nAnalysis across calls addresses precisely that. A pattern visible across six calls over eighteen months is invisible to six separate call-handlers, and no amount of individual diligence would surface it.\n\nBecause it runs on calls already recorded and already lawfully held, it also adds no new collection, and its purpose is protective rather than investigative: identifying people at risk rather than building a case against them.\n\nTHE CASE AGAINST\n\nAnalysing what a caller said means analysing the words of people seeking help, most of whom are victims or witnesses rather than suspects. A system that scores callers for risk is scoring people at their most vulnerable, and the same information could support less benign uses without any technical change.\n\nThere is also a false-reassurance risk. A force that has deployed risk analysis may treat the absence of a flag as evidence of safety, when it may reflect only that the caller did not use the language the system recognises. Under-reporting and understatement are characteristic of exactly the cases these systems target.\n\nTranscription accuracy is a practical constraint that is easy to overlook. Calls to control rooms are made in distress, with background noise, in many accents and languages, and automatic speech recognition performs worst under exactly those conditions — with documented disparities in error rates across accents and dialects.\n\nAnd the retrospective framing can drift. A capability built to review past calls for missed risk is technically the same capability as one applied live, and the governance difference between them is policy rather than architecture.\n\nWHAT IS NOT ESTABLISHED\n\nHow many forces analyse call content, and whether retrospectively or in real time, is not centrally recorded.\n\nNo independent evaluation of whether these systems identify risk that human review missed has been identified.\n\nTranscription accuracy across accents and languages in operational control-room conditions is not published for any deployment reviewed.\n\nWhether callers are informed their words may be analysed is not established.",
      "typical_vendor": "Built in-house by West Yorkshire Police with the NPCC Digital Public Contact programme",
      "related_subject": "emergency-response",
      "url": "https://www.policeainews.com/technologies/ai-call-content-analysis"
    },
    {
      "slug": "drone-first-responder",
      "name": "Drone as First Responder",
      "category": "Drones and robotics",
      "summary": "Drones launched automatically on a 999 or 911 call, arriving before ground units to give responders an aerial view — the fastest-spreading police technology in the United States, with over a thousand agencies cleared in ten months.",
      "description": "Drone as First Responder inverts the usual sequence of police drone use. Rather than an officer taking a drone to a scene, a drone launches the moment a call comes in, flies autonomously to the location, and streams an aerial view to dispatchers and responding units — often arriving before any vehicle.\n\nChula Vista Police Department in California ran the first programme in 2018 under the Federal Aviation Administration's UAS Integration Pilot Program. It secured a beyond-visual-line-of-sight waiver in May 2019 allowing flight up to three miles from the launch site, and in March 2021 became the first US department authorised to launch from anywhere in its city. Four launch points now give it over 90% coverage.\n\nWHY IT SUDDENLY SPREAD\n\nThe growth curve is the most important fact about this technology, and it is entirely a story about regulation rather than about the drones.\n\nBetween 2018 and 2024 the FAA approved just over 50 DFR waivers. Each took eleven months or more and required both a certified remote pilot and a human visual observer, usually stationed on a rooftop.\n\nIn May 2025 the FAA introduced a simplified Certificate of Waiver process, following safety data indicating minimal collision risk with general aviation and the addition of requirements such as parachutes, anti-collision lighting and detection technology. Critically, it permitted technology to substitute for the human visual observer.\n\nMore than a thousand waivers were approved in the following six months, running at roughly 125 a month. A FOIA-obtained list published in August 2026 made the full set public. Axon told shareholders in early 2026 that a DFR mission launches approximately every thirty seconds in the United States.\n\nA standardised national framework, FAA Part 108, is expected to replace case-by-case waivers.\n\nTHE OPERATIONAL RESULTS\n\nThe performance figures are among the strongest for any technology in this catalogue.\n\nChula Vista reports response times cut by half. Across programmes, drones have returned ground units to service in around 25% of responses before those units arrived — meaning the aerial view resolved the call. Reported averages include a 97-second response on calls the drone reaches first, and roughly one call in five cleared without any vehicle rolling.\n\nChief Roxana Kennedy has described DFR as one of the department's best de-escalation tools, on the basis that verifying whether a reported weapon actually exists changes how officers approach a scene.\n\nChula Vista also runs a public dashboard showing daily flights, flight paths, call types and response times — genuinely more transparency than most deployments in this catalogue offer.\n\nWHAT THE DATA ACTUALLY SHOWS ABOUT USE\n\nThat same transparency has produced the sharpest criticism, which is a point worth noting in itself.\n\nAnalysis of Chula Vista's own published data found that more than 20% of flights were welfare checks or mental health crises, while only around 6% responded to assault calls. A Government Technology analysis found many flights answering low-level calls: noise complaints, mental health checks, and reports about unhoused people.\n\nA WIRED investigation found roughly one in ten flights lacked a stated purpose, and that for nearly 500 recent flights the recorded reason was \"unknown problem\". Experts quoted described this as a classic case of self-perpetuating mission creep, with the drones' existence both justifying and necessitating their use.\n\nTHE CASE FOR IT\n\nSending a camera before a person to a scene nobody has yet assessed is a genuinely good idea. Officers and firefighters routinely arrive with incomplete information, and traditional air support is expensive and rarely immediately available.\n\nThe de-escalation argument has substance. Knowing before arrival that a reported weapon is a phone, or that a reported fight has dispersed, changes the posture of the officers who turn up.\n\nAnd clearing one call in five without dispatching a vehicle is a real resource saving in a service where response times are the dominant public measure.\n\nTHE CASE AGAINST\n\nThe mission-creep evidence is strong precisely because it comes from the operator's own dashboard. A capability justified by armed incidents is predominantly used for welfare checks, mental health calls and noise complaints — and sending an aerial camera to a person in mental health crisis is a different proposition from sending one to an active shooter.\n\nThe scale is now the issue. As the EFF has put it, a thousand agencies are acquiring aircraft that can see backyards and windows from angles no patrol car reaches, and most of their cities have not written a single rule about the resulting footage. The regulatory change enabled acquisition; it did not create any governance of what is recorded.\n\nWarrant requirements exist in some states — Florida and Virginia require search warrants absent exigent circumstances — but they are the exception.\n\nThe unlabelled flights matter too. Where one in ten flights has no stated purpose, the transparency dashboard that makes this programme look accountable is also demonstrating the limits of its own accountability.\n\nWHAT IS NOT ESTABLISHED\n\nRetention of DFR footage, and whether it is treated as evidential or operational, is not documented for most of the thousand-plus agencies now cleared.\n\nWhether the response-time and call-clearance figures from early adopters hold for agencies deploying at scale is untested.\n\nHow many of the thousand-plus waiver holders have actually launched programmes is not established; the list records regulatory clearance, not operation.\n\nNo independent evaluation of whether DFR changes case outcomes, as opposed to response times, has been identified.",
      "typical_vendor": "Skydio, among others",
      "related_subject": "drones",
      "url": "https://www.policeainews.com/technologies/drone-first-responder"
    },
    {
      "slug": "police-ground-robots",
      "name": "Police Ground Robots",
      "category": "Drones and robotics",
      "summary": "Wheeled and legged robots used by police to patrol fixed sites or enter hazardous scenes, covering two very different propositions that are frequently conflated.",
      "description": "Police ground robots fall into two categories that share almost nothing beyond being robots, and conflating them has muddled most public discussion of both.\n\nThe first is the autonomous patrol robot: a wheeled machine that circulates a fixed site continuously, recording and alerting. It has no ability to intervene and is generally deployed to car parks, parks, campuses and transport hubs. Knightscope is the dominant supplier and has its own page.\n\nThe second is the remotely operated hazardous-entry robot: a machine an operator directs into a place a person should not go — a suspected explosive device, an armed standoff, a structurally unsafe building. Boston Dynamics' Spot is the best-known current example and also has its own page, though police have used tracked robots for this purpose for nearly fifty years.\n\nWHY THE DISTINCTION MATTERS\n\nThe arguments differ entirely.\n\nA patrol robot is a surveillance question. It records continuously, often in public space, frequently with licence plate recognition, and the issues are retention, access and whether persistent recording of a park is proportionate.\n\nA hazardous-entry robot is not primarily a surveillance question at all. It is deployed to specific incidents, for minutes at a time, and the argument is about militarisation, public perception, and whether the stated use case matches the observed one.\n\nPublic reaction has also diverged sharply. The clearest failure of a patrol robot was a machine unable to help a bystander reporting an assault. The clearest failure of a hazardous-entry robot was a contract cancelled after a subpoena revealed it had never been disclosed as required.\n\nTHE COMMON GROUND\n\nBoth categories share three features worth noting.\n\nNeither can act. A patrol robot alerts; an entry robot observes. Everything either produces depends on a human response, and the gap between detection and response is where both have failed publicly.\n\nBoth are acquired as equipment rather than as surveillance systems, which frequently means less consultation than a camera network attracts, despite carrying cameras.\n\nAnd both are sold substantially on perception — deterrence in one case, officer safety in the other — rather than on measured outcomes, and neither has a substantial independent evidence base.\n\nWHAT IS NOT ESTABLISHED\n\nNo independent evaluation of deterrent effect or operational effectiveness has been identified for either category.\n\nRetention and access rules for footage captured by robots patrolling public space are undocumented in every deployment reviewed.\n\nThe current number of police forces operating either type is not published by any source.",
      "typical_vendor": "Boston Dynamics (Spot), Knightscope (K5), AiMOGA Robotics / Chery (humanoid units)",
      "related_subject": "robotics",
      "url": "https://www.policeainews.com/technologies/police-ground-robots"
    },
    {
      "slug": "mobile-device-extraction",
      "name": "Mobile Device Extraction",
      "category": "Digital forensics",
      "summary": "Tools that download the full contents of a seized phone — including deleted data and, increasingly, linked cloud accounts — used by most UK forces on devices belonging to suspects, witnesses and victims alike.",
      "description": "Mobile device extraction connects a phone to a forensic device that copies its contents: messages, emails, contacts, photographs, video, call logs, calendars, GPS history and application data, including material the owner believed deleted. Encrypted messaging databases from WhatsApp, Signal, Telegram and Facebook Messenger are routinely obtainable.\n\nThe dominant supplier is Cellebrite, an Israeli company whose Universal Forensic Extraction Device is the industry standard. Competitors include MSAB and Oxygen Forensic.\n\nSCALE IN THE UK\n\nA Bristol Cable investigation found around 28 of 44 UK police forces using Cellebrite technology. The Metropolitan Police has distributed the UFED widely among officers. Financial records show a Home Office contract worth £81,017.50 and forces such as Cleveland allocating six-figure sums for extraction systems. Durham examined 90% of mobile phones locally as early as 2015.\n\nPrivacy International's \"Digital Stop and Search\" report, based on freedom of information requests to 47 UK forces, found the technology had been deployed for over six years largely without public knowledge, applied to phones belonging to people convicted of no crime, and to witnesses and victims as well as suspects.\n\nWHY THIS IS AN AI ENTRY\n\nMost of extraction is not AI: it is data copying and file-system parsing. The machine learning element sits in analysis — classification, categorisation and pattern-finding across the extracted material — which is why the Czech Police names Cellebrite UFED among the tools it classifies as applying machine learning while excluding its statistical forensic software.\n\nTHE SELECTIVE EXTRACTION QUESTION\n\nCellebrite states its platform supports selective extraction, meaning investigators need only collect data strictly relevant to the case.\n\nPrivacy International's examination of what that means in practice is the most useful detail on this page. Selecting an application to extract does not mean extracting one or two messages. It means all of that application's data — every WhatsApp conversation, with every contact, going back potentially years, including shared links, images, contact information and associated databases and keys.\n\nInvestigation of Police Scotland's cyber kiosks raised a further question: where police state they can limit what is viewed, it is unclear whether everything is nonetheless extracted first and search parameters applied afterwards. Those are materially different propositions for the person whose phone it is.\n\nCLOUD EXTRACTION\n\nThe capability has extended beyond the device. Cloud extraction reaches data hosted in cloud applications, services and web pages linked to the phone rather than stored on it. Lancashire Constabulary began testing in 2014 and by 2018 was using Cellebrite's cloud tool weekly in two or three investigations; Hampshire Constabulary confirmed use in response to a freedom of information request.\n\nThat is a significant expansion. Seizing a phone becomes a route to material held on servers elsewhere, potentially in other jurisdictions, that was never on the seized device.\n\nTHE CASE FOR IT\n\nDigital evidence is central to a large and growing share of investigations, and a phone frequently contains the material that establishes or refutes an allegation. Extraction is often what exonerates as much as what convicts.\n\nIt is also frequently consensual and victim-initiated: a complainant offering their phone to corroborate an account. In serious cases — child sexual exploitation, homicide, grooming — the capability is genuinely important, as Lancashire's stated use cases indicate.\n\nDeleted-data recovery addresses a real evidential problem, since deletion is a common response to investigation.\n\nTHE CASE AGAINST\n\nThe scope problem is fundamental and is not solved by better technology. A phone examined in connection with one allegation yields years of communications belonging to the owner and to everyone who ever messaged them, none of whom are party to the investigation.\n\nThe framework problem is worse. Privacy International found forces not obtaining warrants, not seeking permission from the people whose phones were searched, and operating without a clear legal framework — which it argued was potentially unlawful. The contrast with the United States is instructive, where Riley v California established a warrant requirement for phone searches.\n\nThe victim and witness dimension is the most damaging in practice. Complainants in serious cases have been asked to surrender phones for full extraction as a condition of an investigation proceeding, which is a substantial disincentive to reporting.\n\nCompetence is a documented concern. The House of Lords Science and Technology Committee's report on forensic science highlighted police not understanding the technology being used, the risk of incorrect inferences from extracted data, and a general lack of understanding of these tools' capabilities.\n\nAnd there is a supply-chain point rarely raised: Cellebrite's own privacy policy indicates it can collect usage information from products in police possession, which means a commercial supplier has visibility into how forces use devices holding evidential material.\n\nWHAT IS NOT ESTABLISHED\n\nWhether selective extraction limits what is copied or only what is displayed is not resolved in the material reviewed.\n\nHow many extractions UK forces perform annually, and what proportion involve witnesses or victims rather than suspects, is not published.\n\nRetention periods for extracted data, as distinct from the seized device, are not consistently documented.\n\nWhether the legal framework concerns raised have since been addressed by statute or guidance is not established here.",
      "typical_vendor": "Cellebrite (UFED); competitors include MSAB and Oxygen Forensic",
      "related_subject": "digital-forensics",
      "url": "https://www.policeainews.com/technologies/mobile-device-extraction"
    },
    {
      "slug": "border-biometric-systems",
      "name": "Border Biometric Identity Systems",
      "category": "Borders and immigration",
      "summary": "Central biometric registration at national borders, principally the EU Entry/Exit System, which records the face and fingerprints of every non-EU traveller entering the Schengen area and shares them across 29 states.",
      "description": "Border biometric identity systems register a traveller's face and fingerprints on first crossing, store them centrally, and verify against that record on every subsequent journey. The largest current example is the European Union's Entry/Exit System, which replaced passport stamping across the Schengen area and offers an unusually clear case study in the difference between a system's design and its deployment.\n\nWHAT THE EU SYSTEM DOES\n\nThe Entry/Exit System began a phased launch on 12 October 2025, with full implementation across all external border points scheduled for April 2026. It applies to non-EU nationals entering or leaving the Schengen area, and covers 29 participating countries.\n\nAt the border it records alphanumeric data — name, travel document details, nationality, time and place of crossing — alongside biometric identifiers comprising a facial image and four fingerprints. The stated purposes are to replace error-prone manual stamping, to track overstays and enforce the 90-days-in-180 rule, and to prevent identity fraud. It also underpins ETIAS, the separate pre-travel authorisation scheme for visa-exempt visitors.\n\nData is held centrally by eu-LISA, the EU agency for large-scale IT systems in justice and home affairs, and shared across member states. Retention is three years after a traveller's last entry, exit or refusal of entry — extended to five years where an overstay is recorded.\n\nTHE DESIGN, WHICH IS SOUND\n\nThe intended experience is worth stating because it is a reasonable proposition. A visitor gives fingerprints and a facial image once. That first crossing was designed to take three to seven minutes. Every subsequent return within the three-year window would then clear an automated gate on a facial scan in well under a minute.\n\nFor frequent travellers, that is a genuine improvement on queuing for a manual stamp. The security argument is also coherent: a stamped passport is a poor instrument for detecting overstays, and a central record makes the 90/180 rule enforceable rather than nominally applicable.\n\nWHAT ACTUALLY HAPPENED\n\nThe rollout has been the clearest recent demonstration that a well-designed border system can fail entirely on implementation.\n\nAt Prague's Václav Havel Airport, among the first major airports to operate it, travellers reported queues of up to ninety minutes. The reported cause is instructive: although EES was designed around self-service kiosks, Czech authorities directed border officers to collect biometric data manually at passport control, removing the throughput the design depended on.\n\nAirports Council International reported processing times increasing by 70% in some locations. On 18 December 2025 its director general, Olivier Jankovec, said airports supported the security goals but could not accept the new system translating into \"mayhem for travelers and chaos at our airports\", and called for an immediate review by the Commission, Frontex, eu-LISA and member states.\n\nThe most striking single incident came in April 2026. One day after the system became fully operational, border officers at Lisbon, Porto and Faro switched EES off, allowed the queues to drain, and switched it back on after lunch. Across the Schengen area passengers reported waits of up to three hours; at one Italian airport a hundred travellers missed their flight. At Paris Charles de Gaulle, non-EU arrivals were routed into manual lanes never built for those numbers.\n\nThe legal framework anticipated some of this: member states may partially suspend EES checks for up to 90 days after rollout completes, extendable by 60 days, specifically to prevent queues at peak periods.\n\nThe Commission's position has been considerably more positive. A spokesman for Magnus Brunner, the commissioner responsible, described the rollout as \"smooth and well-managed\".\n\nTHE CASE FOR IT\n\nManual stamping is genuinely inadequate for its purpose, and overstay enforcement without a central record is close to unworkable. If the system reaches its design performance, most travellers will spend less time at borders than before, not more.\n\nBiometric verification against a record the traveller themselves created is also a narrower proposition than most systems in this catalogue. It answers \"are you the person who registered?\" rather than searching a population against a watchlist, and it does not attempt to infer anything about behaviour or risk.\n\nRetention is defined, time-limited and published, which is more than can be said for several systems here.\n\nTHE CASE AGAINST\n\nThe clearest objection is that the system creates a comprehensive, centrally held biometric record of every non-EU national who crosses a Schengen border, shared across 29 states. That database exists regardless of whether any individual traveller has done anything to warrant suspicion, and its future uses depend on decisions not yet taken.\n\nFunction creep is the standing concern with any large biometric store. The EES was built for border management; it sits alongside SIS II, VIS and EURODAC, and this site's tracker already records handheld police devices in another member state querying that family of databases during street stops. Once the record exists, the question of who may search it is a policy setting rather than a technical constraint.\n\nThe implementation record also matters on its own terms. Border officers switching a system off to clear queues, and national authorities bypassing the kiosks the design depended on, are not teething problems — they indicate a system deployed ahead of the physical and staffing capacity needed to run it.\n\nInconsistent application has its own consequences: travellers through Spain and Switzerland reported biometric checks applied irregularly or repeated on return trips, which undermines both the traveller experience and the integrity of the record.\n\nWHAT IS NOT ESTABLISHED\n\nWhether processing times have since converged on the three-to-seven-minute design figure is not established by the sources reviewed.\n\nHow many travellers are currently registered, and what proportion of crossings are handled at automated gates rather than manually, is not published in the material examined.\n\nWhether law enforcement access to EES data beyond border management has been granted or sought in any member state is not established here.\n\nNo independent evaluation of whether the system has improved overstay detection — its principal stated purpose — has been identified.",
      "typical_vendor": "Operated by eu-LISA for the European Union",
      "related_subject": "borders-immigration",
      "url": "https://www.policeainews.com/technologies/border-biometric-systems"
    },
    {
      "slug": "ai-weapons-detection",
      "name": "AI Weapons Detection Screening",
      "category": "Sensors and detection",
      "summary": "Walk-through scanners marketed as AI-powered weapons detection for schools, stadiums and transport hubs, and the subject of an FTC enforcement action over deceptive claims about what they can detect.",
      "description": "These systems screen people walking through a portal, using sensors and pattern recognition intended to identify the signature of a weapon while ignoring everyday items. They are marketed as a faster, less intrusive alternative to metal detectors: no emptying pockets, no queue, high throughput.\n\nThe dominant supplier is Evolv Technologies, whose Express scanners are used in over 800 schools across 40 states, alongside stadiums, hospitals and transport systems. School systems make up around half its business. The company has been backed by prominent investors including Bill Gates.\n\nWHAT THE COMPANY CLAIMED\n\nEvolv marketed the scanners as able to distinguish weapons from everyday objects using AI, claiming they could identify the unique signatures of tens of thousands of weapons and separate guns, bombs and large tactical knives from keys and laptops. It advertised screening of up to 1,000 students in fifteen minutes, ten times faster than traditional metal detectors, and positioned the technology as more accurate, efficient and cost-effective than metal detection.\n\nTHE FTC ACTION\n\nIn November 2024 the Federal Trade Commission filed a complaint and proposed settlement alleging Evolv deceptively advertised that its scanners would detect all weapons, ignore harmless personal items, and outperform metal detectors through the use of artificial intelligence.\n\nThe specific findings are the substance of this page.\n\nThe scanners failed in several instances to detect weapons in schools while flagging harmless items students routinely carry — laptops, binders, water bottles. In October 2022 an Express scanner failed to detect a seven-inch knife brought into a school, which was then used to stab a student. After that incident the school raised the sensitivity setting, which produced a 50% false alarm rate.\n\nA 2021 test at a Major League Soccer stadium in Columbus, Ohio found the scanners missed two small handguns and detected knives only 58% of the time. In one deployment across twenty subway stations over a month, the scanners produced 118 false positives, recovered 12 knives, and detected no firearms at all. In some school settings the scanners flagged more than 85,000 false positives on laptops between August 2023 and April 2024.\n\nThe FTC also alleged that Evolv altered a report from the National Center for Spectator Sports Safety and Security to remove negative findings about its scanners.\n\nThe commission further noted that Evolv's own remedies — a higher sensitivity setting, plus conveyor belts and manual diversion of harmless items — made the system function more like the lower-cost metal detectors it was sold as superior to, while requiring additional staffing.\n\nTHE SETTLEMENT\n\nEvolv was barred from making unsupported claims about its products' ability to detect weapons using artificial intelligence, and required to notify certain K-12 customers with contracts signed between April 2022 and June 2023 that they could cancel. Customers who had run a 30-day trial or bought fifteen or more scanners were excluded from that option.\n\nThe company's own position is recorded: it stated it disagrees with the FTC's allegations, has not admitted wrongdoing, and resolved the matter to focus on its mission. It emphasised that the FTC did not challenge the core efficacy of its products including its use of AI, and pointed to incidents in which its scanners did detect guns students attempted to bring into schools.\n\nTHE CASE FOR IT\n\nThe problem is real and the alternative is unattractive. Metal detectors are slow, create queues that are themselves a vulnerability, and are experienced as prison-like in a school. A system that screens a thousand people quickly without emptying pockets addresses a genuine operational constraint.\n\nThe scanners do detect weapons. Documented interceptions exist, and neither the FTC nor this site suggests the technology detects nothing.\n\nTHE CASE AGAINST\n\nThe gap between marketing and performance is the issue, and it was established by a federal regulator rather than by campaigners. A 58% knife detection rate and zero firearms detected across a month of subway screening are not incidental shortfalls; they are the central capability failing.\n\nThe false positive burden inverts the value proposition. Eighty-five thousand laptop alarms is not a screening system, it is a queue with extra steps — and the remedies proposed reintroduce exactly the staffing and equipment the product was sold as replacing.\n\nThe most serious allegation is the alteration of a third-party test report. Independent evaluation is the only mechanism by which a purchaser can check a vendor claim, and interfering with it undermines the entire basis on which these decisions are made.\n\nThe consequences fall on people who cannot assess the claims. School districts spent $3.7 million, $2.5 million and $17 million respectively on systems they had no capacity to evaluate, and a student was stabbed with a knife the scanner did not detect.\n\nWHAT IS NOT ESTABLISHED\n\nCurrent detection rates for the deployed configuration, after the sensitivity changes, are not independently published.\n\nHow many customers exercised the cancellation right is not established.\n\nThe outcome of the SEC investigation reported in 2024 is not established here.\n\nNo independent evaluation comparing these scanners against metal detectors in matched conditions has been identified.",
      "typical_vendor": "Evolv Technologies (Evolv Express); competitors include ZeroEyes",
      "related_subject": "surveillance",
      "url": "https://www.policeainews.com/technologies/ai-weapons-detection"
    },
    {
      "slug": "bwc-audio-analytics",
      "name": "Body-Worn Camera Audio Analytics",
      "category": "Body-worn cameras",
      "summary": "Automated analysis of body-worn camera audio to score officer conduct and flag encounters for supervisory review — one of the few technologies in this catalogue pointed at police rather than at the public.",
      "description": "Body-worn camera audio analytics transcribes and analyses what officers and members of the public say during recorded encounters, classifying interactions and flagging those that warrant supervisory attention.\n\nIt is unusual in this catalogue in that its subject is the officer. Almost every other system here is directed at the public; this one is directed at police conduct, and that inversion changes most of the arguments.\n\nWHAT IT DOES\n\nBody-worn camera footage is generated in enormous volume and almost never watched. A force may record tens of thousands of hours a month, of which a supervisor reviews a fraction, usually only after a complaint.\n\nAudio analytics addresses that by processing all of it: transcribing encounters, identifying professional and unprofessional language, detecting indicators of escalation, and surfacing interactions for review that no one would otherwise have looked at.\n\nDeployments recorded on this site's tracker are listed below.\n\nTHE CASE FOR IT\n\nThe oversight gap is real. Body cameras were introduced for accountability, but accountability requires someone to look, and at current volumes nobody does. A system that reviews everything and flags the concerning encounters is a genuine answer to a genuine failure.\n\nIt also identifies good practice, not only bad. Analytics that surface effective de-escalation give training a real evidence base rather than anecdote, and several vendors position the technology primarily as a training tool.\n\nCrucially, it operates on encounters that are already recorded with the knowledge of both parties. No new collection occurs.\n\nAnd it addresses the selection problem in police oversight: currently, review is triggered by complaint, which means conduct toward people least likely to complain is least likely to be examined. Automated review is indifferent to whether the person complained.\n\nTHE CASE AGAINST\n\nThe obvious risk is that a technology introduced for accountability becomes one for surveillance of the workforce, and then, incrementally, for evidence-gathering about the public. The footage is the same footage; only the analysis differs, and expanding the analysis requires no new equipment or consultation.\n\nAutomated classification of conduct is also harder than it appears. Whether language was appropriate depends on context a transcript does not capture, and a system that flags raised voices in a genuinely dangerous situation generates noise while potentially missing quiet coercion.\n\nOfficer trust matters practically rather than sentimentally. If officers believe every word is being scored, the documented risk is that they say less, deactivate cameras more, or engage less with the public — which would defeat the purpose of body cameras entirely.\n\nAnd the finding this site records elsewhere applies here: where a technology's outputs are used in disciplinary processes, the reliability of the analysis becomes material, and no independent evaluation of these systems' classification accuracy has been identified.\n\nWHAT IS NOT ESTABLISHED\n\nHow many forces use audio analytics on body-camera footage is not centrally recorded.\n\nWhether outputs are used for training only, or also in disciplinary proceedings, differs by force and is rarely published.\n\nNo independent evaluation of classification accuracy, or of whether flagged encounters correlate with substantiated complaints, has been identified.\n\nWhether officers are informed which behaviours are being scored is not documented in the deployments reviewed.",
      "typical_vendor": "Truleo",
      "related_subject": "body-cameras",
      "url": "https://www.policeainews.com/technologies/bwc-audio-analytics"
    },
    {
      "slug": "private-lfr-networks",
      "name": "Private Live Facial Recognition Networks",
      "category": "Facial recognition",
      "summary": "Facial recognition operated by businesses and community organisations rather than police, creating watchlists and exclusions outside any statutory framework, and increasingly feeding information to forces.",
      "description": "Private live facial recognition covers systems run by retailers, venues, community organisations and private security firms rather than by police. The technology is largely the same as police deployments. The governance is not.\n\nWHY THIS IS A SEPARATE CATEGORY\n\nWhen a police force deploys facial recognition it is subject to data protection law, the public sector equality duty, judicial review, and — in the UK — a body of case law from Bridges to Thompson. A shop deploying the same technology is subject to data protection law and very little else.\n\nThat asymmetry produces the defining feature of private deployment: consequences imposed without process. A person flagged by a police system may be stopped and questioned, with the constraints that attach to a police stop. A person flagged by a retail system may be removed from a shop and added to a watchlist shared across a network of businesses, with no hearing, no notification and no clear appeal.\n\nTHE MODELS\n\nRetail networks are the most developed. This site's Facewatch page documents the largest UK example, where watchlists compiled from incidents reported by shop staff are shared across participating businesses.\n\nCommunity and non-profit hubs are a second model. Project NOLA in New Orleans operates as a private non-profit running a camera network that feeds police, inverting the usual direction: the surveillance is privately operated and the state is the customer rather than the operator.\n\nA third model runs the other way. Police-operated platforms such as Axon's Fusus allow private camera owners to enrol their own cameras into a police operating picture, so private infrastructure becomes part of a public system by voluntary contribution.\n\nThe three models differ in who owns the camera, who runs the analysis and who acts on the result — but all three move surveillance capability outside the framework that governs police deployment.\n\nTHE CASE FOR IT\n\nBusinesses have a legitimate interest in preventing crime on their own premises and protecting staff from assault, and have always been permitted to exclude individuals. Facial recognition automates a judgement that shop staff already make and are entitled to make.\n\nPrivate camera networks also fill genuine gaps. Police cannot camera every street, and voluntary contribution of existing footage is cheaper than public installation.\n\nWhere private systems feed police with material relating to specific documented incidents, the investigative value is real, and this site's tracker records cases where it has contributed to identifying suspects.\n\nTHE CASE AGAINST\n\nThe core objection is regulatory arbitrage. A capability that would attract sustained scrutiny in police hands attracts far less in private hands, while producing comparable consequences for the individual — and in the retail case, exclusion from essential shops.\n\nWatchlist governance is the practical failure point. Inclusion typically rests on a report by an employee rather than any finding of fact, and network sharing multiplies the consequence of a single erroneous report.\n\nThe private-to-state pathway raises a distinct question. When a private network feeds police, material collected under private-sector rules enters a criminal justice process that has its own evidential standards, and the transition between the two is rarely documented.\n\nFinally, accumulation. Each individual deployment is defensible on its own terms; the aggregate is a parallel surveillance infrastructure that no one decided to build.\n\nWHAT IS NOT ESTABLISHED\n\nThe total scale of private facial recognition deployment in the UK or elsewhere is not published by anyone.\n\nHow often private systems pass information to police, and under what basis, is not systematically recorded.\n\nWhether individuals on private watchlists have any effective route to discover or challenge their inclusion is unclear in every deployment reviewed.",
      "typical_vendor": "Project NOLA (non-profit operator)",
      "related_subject": "facial-recognition",
      "url": "https://www.policeainews.com/technologies/private-lfr-networks"
    },
    {
      "slug": "national-police-ai-programmes",
      "name": "National Police AI Programmes",
      "category": "Governance",
      "summary": "Government programmes coordinating AI adoption across all forces in a country, of which the UK's PoliceAI is the most developed, funded under a named statutory mechanism rather than dedicated AI legislation.",
      "description": "Most entries in this catalogue are individual systems bought by individual forces. National programmes are different: government-level efforts to coordinate AI adoption across all forces in a jurisdiction, set standards, and fund common capability.\n\nThey matter because they determine what individual forces will be buying in five years, and because they are the level at which governance could be set — or not set.\n\nTHE UK PROGRAMME\n\nPoliceAI is administered by a team hosted in the College of Policing and will transition into the National Police Service once that body is established.\n\nIts legal basis is more specific than is usually reported. PoliceAI is funded through a grant agreement issued by the Home Office under Section 57 of the Police Act 1996. That is not dedicated AI legislation, but it is a named statutory funding mechanism rather than an absence of legal basis.\n\nA component often omitted is the Policing AI Threat Hub, which leads the national policing response to AI-enabled crime, including deepfake intimate images. This is the inverse of most of this catalogue: police responding to AI used against the public rather than deploying it.\n\nFunding detail from the January 2026 White Paper includes £26 million for the development and delivery of a national facial recognition system and a further £11.6 million on live facial recognition capabilities.\n\nRELATED LEGISLATIVE CHANGE\n\nFrom May 2026 it became illegal in the UK to create, adapt or supply a tool that generates non-consensual intimate images, with developers as well as users criminally liable. The government describes its legislation banning generators of AI child sexual abuse material as the first of its kind in the world.\n\nThe government has also announced plans for all police forces to use AI to review and summarise evidence by 2027.\n\nTHE CASE FOR IT\n\nCoordination is genuinely preferable to forty-three forces independently procuring incompatible systems at forty-three different price points, with forty-three different standards of assurance. Central programmes can impose testing requirements that individual forces lack the capacity to run — the National Physical Laboratory evaluation of facial recognition algorithms is a direct example.\n\nA national programme is also a single point of accountability. It is far easier to scrutinise one funded programme than to track procurement across every force.\n\nThe Threat Hub addresses a real and growing harm, and it is the clearest case in this catalogue of police AI capability being built for public protection against AI rather than for surveillance.\n\nTHE CASE AGAINST\n\nCoordination accelerates adoption, and a national programme that funds facial recognition capability is making a policy decision about the direction of policing that has not been separately legislated. £26 million for a national facial recognition system is a substantial commitment made through a funding mechanism rather than through a statute debated on its merits.\n\nThe Biometrics and Surveillance Camera Commissioner has observed that forces acquire powerful technologies without fully understanding their capabilities and implications, and that despite a national decision-making model for buying new technology existing, his office has found no evidence of forces following it. A national programme that funds capability without ensuring that model is followed reproduces the problem at scale.\n\nThere is also a sequencing question that runs through this site. The facial recognition framework consultation response remained unpublished well past its due date as of August 2026, while national funding for facial recognition capability proceeded. Capability is being built ahead of the framework intended to govern it.\n\nWHAT IS NOT ESTABLISHED\n\nWhether the transition of PoliceAI into the National Police Service has occurred is not established.\n\nWhat proportion of the announced funding has been spent, and on what, is not published.\n\nWhether the 2027 evidence-summarisation commitment is on track is not established.\n\nHow the programme's outputs will be evaluated, and by whom, is not documented in the sources reviewed.",
      "typical_vendor": "UK Home Office and College of Policing (PoliceAI); City of London Police with Capita and PwC (FCCRAS)",
      "related_subject": "police-tech-procurement",
      "url": "https://www.policeainews.com/technologies/national-police-ai-programmes"
    },
    {
      "slug": "flock-os-investigate",
      "name": "Flock OS Investigate",
      "category": "Investigative analytics",
      "summary": "An AI investigation layer over Flock's number plate camera network that lets police search for people and vehicles by movement pattern, without a plate, a name or a named crime, reconstructed from files exposed on Flock's own login pages.",
      "description": "This is a materially different proposition from the automatic number plate recognition network it sits on, and is recorded separately for that reason. Formerly called Nightshift and now marketed as OS Investigate, it combines vehicle movement data from Flock cameras with police databases and commercial identity information, and allows investigators to search in natural language.\n\nThe change is in what question can be asked. Conventional plate recognition answers where a specific vehicle has been. This allows searches that begin without a known registration number and without a known individual: identifying vehicles exhibiting a pattern, then moving from those vehicles to potential individuals, associates and suspects.\n\nWHAT THE INVESTIGATION FOUND\n\nMost of what is publicly known about the tool comes from a WIRED investigation by Dhruv Mehrotra and Dell Cameron, which reconstructed its capabilities from more than 450 files exposed through Flock's own login pages rather than from any disclosure by the company.\n\nThe reporting describes a set of built-in prompts that let officers search for people by behaviour. A substantial number of them focus on movement patterns, and a large share require no registration number, no name and no physical description to run. The system is described as cross-referencing roughly 120,000 cameras against case files, emergency dispatch records, ballistics data and commercial identity databases.\n\nFlock has stated the tool is still in testing.\n\nWHY THIS SITS AWKWARDLY WITH THE COMPANY'S PUBLIC POSITION\n\nFor years Flock told the public that its cameras cannot recognise, identify or track individuals — a claim that underpinned much of the case made to city councils considering the network. A tool that starts from a movement pattern and works towards an individual is difficult to reconcile with that assurance, and the tension is the substance of the criticism rather than an inference from it.\n\nTHE CASE FOR IT\n\nThe underlying investigative problem is genuine. A vehicle seen leaving a scene, with no legible plate, is a dead end under conventional plate recognition. Pattern-based search offers a route where none existed, and Flock's network is large enough that the route may actually lead somewhere.\n\nIntegration is also a real efficiency. Investigators currently correlate camera data, dispatch records and case files by hand across separate systems, and doing it in one place is faster and less error-prone.\n\nTHE CASE AGAINST\n\nThe central objection is that searching by pattern rather than by identifier inverts the usual sequence of an investigation. Conventionally, suspicion attaches to a person and evidence is then gathered about them. A query that returns everyone whose movements match a described behaviour reverses that: the population is searched first, and suspicion is assigned to whoever the query returns. The ACLU's Chad Marlow has characterised this class of search as a fishing expedition.\n\nThe disclosure route matters too. This capability became public because files were exposed on the company's own login pages, not because Flock, or any police department using it, told anyone. A tool of this reach entering service without public notice is a governance failure independent of whether the tool is useful.\n\nThe scale of cross-referencing compounds it. Combining camera data with commercial identity databases means the search reaches material that no police force collected and no court authorised.\n\nContext is relevant: the Washington Post reported that Flock cameras featured in 46 of at least 50 recent documented misuse cases, and the company announced a set of accountability measures on 13 August 2026, shortly before a national week of action against plate readers.\n\nWHAT IS NOT ESTABLISHED\n\nWhich agencies have access, and whether any have used it operationally rather than in testing, is not established.\n\nWhether the prompt set reported by WIRED reflects the current product is unknown, since the reconstruction came from exposed files rather than documentation.\n\nWhat audit trail exists for a pattern-based query, and whether any authorisation is required before running one, is undocumented.\n\nFlock's own account of the tool's capabilities has not been published in any detail comparable to the reporting.",
      "typical_vendor": "Flock Safety",
      "related_subject": "anpr",
      "url": "https://www.policeainews.com/technologies/flock-os-investigate"
    },
    {
      "slug": "mobile-fortify-federal-biometric-id",
      "name": "Federal Mobile Biometric Identification (Mobile Fortify)",
      "category": "Facial recognition",
      "summary": "A smartphone app letting US federal officers identify people during street encounters by photographing their face or capturing contactless fingerprints, matched against border databases holding around 200 million images.",
      "description": "Mobile Fortify is used by Immigration and Customs Enforcement and Customs and Border Protection officers to identify people during encounters on the street. An officer photographs the person; the image is sent to CBP's Traveler Verification Service — the same infrastructure used for airport processing — and compared against government-held images. If no facial match returns, the app allows contactless fingerprint capture matched against IDENT, the largest biometric repository in the US government.\n\nIt reportedly also supports a \"Super Query\" searching databases covering individuals, vehicles, aircraft, vessels and addresses simultaneously. Its existence was first reported by 404 Media, via leaked internal emails, rather than disclosed.\n\nWHAT MAKES IT DIFFERENT\n\nThree features distinguish it from the handheld facial recognition covered elsewhere on this site.\n\nThe gallery is border and travel biometric infrastructure rather than a policing watchlist. It therefore includes people who have never been arrested, and in many cases people whose only interaction with the state was entering the country lawfully.\n\nIt is multimodal. Where a facial match fails, fingerprints follow, which makes refusal considerably harder than with a face-only system.\n\nAnd it operates federally. More than twenty US cities and states have banned or restricted local police use of facial recognition; a federal officer using this app in one of those jurisdictions is not covered by those bans. The ACLU has made this its central argument, and it is the most consequential structural point.\n\nRETENTION AND CONSENT\n\nA Privacy Threshold Analysis states that ICE does not provide individuals the opportunity to decline or consent to biometric collection, that photographs and fingerprints are retained for fifteen years including images that do not match existing DHS data, and that the app may be used regardless of a person's citizenship or immigration status.\n\nThat last point deserves emphasis: it is not restricted to non-citizens.\n\nSCALE AND OVERSIGHT\n\nA lawsuit brought by Illinois and Chicago in January 2026 states DHS had used the app in the field more than 100,000 times since launch. A Privacy Threshold Analysis indicates a related module launched in September 2025 and could be extended to more than a thousand state and local agencies operating under 287(g) agreements.\n\nIt was deployed without a completed Privacy Impact Assessment — the specific procedural failure driving the oversight response. A DHS Office of Inspector General audit is examining immigration enforcement data practices, prompted by senators in January 2026. On 5 February 2026 legislators introduced the \"ICE Out of My Face Act\", which would ban ICE and CBP use of facial recognition and other biometrics and require deletion. A coalition including EFF, EPIC and the Project on Government Oversight has demanded the app be shut down.\n\nTHE CASE FOR IT\n\nEstablishing identity during an encounter is a legitimate function, and officers frequently encounter people carrying no documents. Resolving that in seconds is faster than the alternative, which is often detention for the purpose of identification.\n\nContactless fingerprint capture is also more reliable than facial matching in poor conditions, and a system that verifies rather than guesses reduces one category of error.\n\nTHE CASE AGAINST\n\nThe absence of a completed Privacy Impact Assessment before deployment is not a technicality. It is the mechanism by which the US federal government is supposed to evaluate exactly these questions before a system is used on people, and it was skipped.\n\nFifteen-year retention of biometrics from people who did not match anything is difficult to justify on any operational rationale, and no consent mechanism exists.\n\nThe federal bypass of local bans is the deepest issue. Where a city has democratically decided its police will not use facial recognition, a federal officer conducting the same check on the same street defeats that decision entirely, and no local process can address it.\n\nAnd a tool designed for immigration enforcement that may be used regardless of citizenship status means US citizens are subject to biometric identification with no ability to decline.\n\nWHAT IS NOT ESTABLISHED\n\nWhether a Privacy Impact Assessment has since been completed is not established.\n\nHow many of the 100,000-plus uses resulted in a correct identification, a wrongful one, or no match is not published.\n\nWhether the app has been extended to 287(g) partner agencies in practice is not confirmed.\n\nThe outcome of the DHS OIG audit was not available at the time of writing.",
      "typical_vendor": "Built on CBP's Traveler Verification Service; app first reported by 404 Media",
      "related_subject": "border-immigration",
      "url": "https://www.policeainews.com/technologies/mobile-fortify-federal-biometric-id"
    },
    {
      "slug": "ai-prison-communications-monitoring",
      "name": "AI Prison Communications Monitoring",
      "category": "Custody, prisons and probation",
      "summary": "Continuous AI transcription and analysis of telephone calls made by incarcerated people, at population scale, now extended into voice-identity tracking and video analysis.",
      "description": "These systems ingest a facility's existing telephone infrastructure, transcribe every call using automatic speech recognition, and apply keyword and semantic analysis to flag calls for investigators. Transcripts become searchable by keyword and metadata. Keywords are typically selected jointly by facility officials and the vendor.\n\nThe dominant example is Verus, from LEO Technologies, which uses Amazon Transcribe for speech-to-text and layers keyword, semantic and generative analysis on top.\n\nTHE SCALE\n\nSuffolk County, New York was an early pilot from 2019, scanning as many as 600,000 minutes a month. The Florida Department of Corrections paid $2.5 million from August 2023 under a contract allowing up to 50 million minutes. Georgia's Department of Corrections began in March 2020, with the company reporting more than 7.5 million calls monitored by April 2021 and actionable evidence of criminal activity in 1,612 of them. In 2022 the company stated Verus monitored more than 100,000 calls originating in California, Florida, Georgia, Missouri, Texas, New York and Louisiana — over 65 million calls a year.\n\nWHAT THE INVESTIGATION FOUND\n\nA Thomson Reuters Foundation investigation obtained contracts and emails from eight states and found the system used well beyond its stated gang-and-violence rationale.\n\nDocumented examples include a call flagged in which a man told his father about a COVID-19 outbreak he alleged officials were covering up, which was then shared with prison officials and company employees; and a call flagged because the person was collecting unemployment benefits. In a sales pitch to Cook County, Illinois, a company representative cited an Alabama sheriff's belief that the calls would help him fend off pending liability from civil action by inmates and activists.\n\nThe system was also reportedly configured to flag Spanish words including \"abogado\" and \"abogada\" — lawyer — which raises an obvious legal privilege question.\n\nTHE 2026 EXTENSION\n\nLEO Technologies has extended Verus from call content into voice biometrics and video. Verus Voice AI continuously builds voice identities across monitored calls and alerts investigators when a detected speaker does not appear to match the person assigned to the account used to place the call, targeting PIN sharing. It can also act as an investigative voice seeker for a voice whose account is unknown. The company separately markets AI video analysis for correctional facilities with automated counts, behaviour detection and real-time alerts.\n\nBiometric Update characterised this as a shift from investigator-initiated search to a continuously operating intelligence system connecting communications with physical movement and institutional history.\n\nTHE CASE FOR IT\n\nPrisons are places where serious crime is organised, and calls are a documented vector. Monitoring is lawful, disclosed to all parties, and long established — US courts have given authorities wide scope to surveil prisoner conversations where the purpose is safety and crime prevention.\n\nAutomating what was previously spot-checking means a facility can act on what is actually said rather than on a sample, and the Georgia figures suggest it does surface genuine criminal activity.\n\nLegitimate safeguarding uses exist too, including identifying people at risk of self-harm or being coerced.\n\nTHE CASE AGAINST\n\nThe defining problem is who else is on the call. Every conversation captured includes a person outside the institution — usually a family member — who has been convicted of nothing, has no realistic alternative means of contact, and whose only choice is to accept monitoring or not speak to their relative.\n\nThe documented scope creep is not speculative. Flagging a call about a COVID outbreak, or about unemployment benefits, is not gang intelligence, and the Alabama sales pitch describes using the system to defend the institution against litigation by the people it holds.\n\nThe reported flagging of the word \"lawyer\" is the most serious single allegation, because privileged communication is the one category monitoring is supposed to exclude.\n\nThe 2026 extension changes the character of the system. Continuous voice identification builds a biometric profile of everyone who calls a prisoner, including outside parties, and no consent mechanism exists for them.\n\nWHAT IS NOT ESTABLISHED\n\nNo named facility has been confirmed as deploying the voice biometrics extension; it is recorded as a launched product rather than a verified deployment.\n\nWhether the legal-privilege keyword configuration was corrected, and whether any privileged call was reviewed, is not established.\n\nWhat proportion of flagged calls result in disciplinary or criminal action is not published.\n\nRetention periods for transcripts, and for voice profiles of outside callers, are not documented in any deployment reviewed.",
      "typical_vendor": "LEO Technologies (Verus)",
      "related_subject": "custody-prisons",
      "url": "https://www.policeainews.com/technologies/ai-prison-communications-monitoring"
    },
    {
      "slug": "csam-detection-classification",
      "name": "CSAM Detection and Classification",
      "category": "Child protection",
      "summary": "Deep-learning classifiers that detect previously unseen child sexual abuse material, including deepfake CSAM, and grade its severity — going beyond hash-matching of already-known images.",
      "description": "National child abuse image databases have historically relied on hash matching: comparing a file's digital fingerprint against fingerprints of already-identified abuse material. That catches what is known and nothing else.\n\nClassification is the newer capability. Deep learning models trained on an existing corpus detect characteristics of abuse imagery in files never previously catalogued, and grade severity against standard government categories — work that would otherwise require an officer to view the material directly.\n\nTHE UK IMPLEMENTATION\n\nThe Child Abuse Image Database went live in December 2014 and is used by all UK territorial forces, Police Scotland, the PSNI and the National Crime Agency. It holds images, videos, case reference data, data about faces and objects appearing in material, grading data, hash values and victim characteristics — but not victim names or contact details.\n\nContent arrives from suspect devices seized in raids, from overseas forces and agencies, and from the Internet Watch Foundation, which has had access since 2018 and can upload directly.\n\nThe AI layer is the Vigil AI CAID Classifier, developed with the Home Office CAID team and delivered by Roke. Roke describes it as, to its author's knowledge, the first operationally viable CSAM classifier able to both detect and grade, and states it handles previously unknown material including deepfake CSAM. Those claims are attributed to the supplier rather than stated as established.\n\nThe Home Office had invested £18.2 million by 2019, with AI and fast-forensic innovations costing a further £1.76 million, developed with UK suppliers including Qumodo, Vigil AI and Cyan Forensics. The database held 13 million images at that point, growing by around half a million every two months.\n\nTHE CASE FOR IT\n\nThe wellbeing argument is the strongest and is rarely made explicit elsewhere in this catalogue: automated grading reduces how much abuse material a human officer must personally view in order to categorise a case. That is a genuine benefit to the people doing this work, whose exposure has documented psychological consequences.\n\nThe capability gap is also real. Hash matching cannot identify a newly produced image, which means it cannot help identify a child currently being abused. A classifier can flag material no one has seen before, and that is the difference between historic and live safeguarding.\n\nThe deepfake capability, if it performs as described, addresses a category of harm that is growing and that hash matching cannot touch at all.\n\nTHE CASE AGAINST\n\nThe claims are supplier claims. Roke's white paper is a commercial document, and no independent evaluation of the classifier's accuracy — its false positive or false negative rate, or whether performance varies by the apparent ethnicity or age of the child depicted — has been identified.\n\nThe consequences of error are asymmetric and severe in both directions. A false negative leaves material uncatalogued; a false positive attaches the most serious possible allegation to an individual. Grading errors affect sentencing.\n\nThere is also no dedicated statute or named external regulator for CAID, which for a database of this sensitivity is a notable gap. A Home Office market notice in May 2024 signalled a possible move to public cloud hosting, reversing a 2020 legal opinion that data protection concerns meant the data should not be cloud hosted; the hosting contract expired in early 2026 and the outcome is not established.\n\nWHAT IS NOT ESTABLISHED\n\nNo independent accuracy evaluation of the classifier has been identified.\n\nWhether the cloud hosting move proceeded is not established.\n\nCurrent database size and growth rate are not published beyond the 2019 figures.\n\nWhether the deepfake detection capability has been tested against current generative models, which have advanced substantially since the tool was built, is not documented.",
      "typical_vendor": "Various; UK deployment built by Roke",
      "related_subject": "child-protection",
      "url": "https://www.policeainews.com/technologies/csam-detection-classification"
    },
    {
      "slug": "multimodal-handheld-biometric-id",
      "name": "Multimodal Handheld Biometric Identification",
      "category": "Biometrics",
      "summary": "Portable devices combining facial recognition, fingerprint capture and document scanning in one unit for street-level identity checks against national and international databases — ruled unlawful in Greece after full procurement and delivery.",
      "description": "Unlike single-modality handheld facial recognition, which checks a photographed face against one watchlist, multimodal devices combine several biometric and identity checks in a single unit: a face, a fingerprint, a scanned document and sometimes a vehicle plate, all checked in the same street stop against multiple databases at once.\n\nTHE GREEK CASE\n\nThe fullest documented example is Greece's Smart Policing programme, which equipped Hellenic Police officers with 1,000 devices.\n\nOfficers capture biometric data during an on-the-spot check and compare it instantly against national, EU and international databases including the Schengen Information System, the Visa Information System, EURODAC, and systems linked to Europol and third-country authorities. Human Rights Watch and Homo Digitalis reported comparison against roughly twenty databases including Interpol. Police stated that biometric data processed during stops would not be retained and would be used solely for instantaneous comparison before being discarded.\n\nThe contract was signed with Intracom Telecom in spring 2019, valued at €4 million, with 75% funded by the European Commission's Internal Security Fund. The state paid in full and the devices were delivered fully functional.\n\nTHE RULING\n\nAfter Homo Digitalis brought the programme to light with AlgorithmWatch in December 2019, filed an access-to-documents request, received an inadequate response and lodged a formal complaint in March 2020, the Hellenic Data Protection Authority ruled the procurement, operation and use of the Smart Policing system unlawful in Decision 45/2025, dated 31 December 2025.\n\nAn earlier June 2020 ruling had already established that the CCTV law the ministry relied on did not cover facial recognition or comparable identification methods.\n\nThis is a regulator finding an EU-funded, fully procured and delivered biometric system unlawful after the money was spent.\n\nWHY THIS MATTERS BEYOND GREECE\n\nSet against two other entries on this site, it completes a picture. In Italy, the data protection authority blocked a real-time facial recognition system before deployment. In the United Kingdom, handheld operator-initiated facial recognition proceeds with no dedicated statutory framework and a consultation response long overdue. In Greece, a system was fully built and paid for before being found unlawful.\n\nThree jurisdictions, one broad capability, three entirely different regulatory outcomes — determined by when the regulator engaged rather than by anything about the technology.\n\nTHE CASE FOR IT\n\nEstablishing identity in the field is a legitimate function, and a device resolving it in seconds is faster than the alternative of detention for identification.\n\nMultimodal checking is also more reliable than face alone: where a facial match is inconclusive, a fingerprint resolves it, reducing one category of error.\n\nThe stated non-retention position — instantaneous comparison then discard — is a narrower proposition than systems that build databases from field encounters.\n\nTHE CASE AGAINST\n\nThe gallery is the fundamental issue. Checking a person against SIS II, VIS and EURODAC is not a policing watchlist check; it is a check against EU immigration and border infrastructure, and the programme was reported as targeted at stops of undocumented migrants living in the country. That places it in immigration enforcement conducted through ordinary street policing.\n\nThe procurement sequence is the governance failure. €4 million of largely EU money was spent, and a thousand devices delivered, before the lawfulness question was resolved — and it was resolved only because a civil society organisation pursued it for five years.\n\nNon-retention is a stated policy, not a technical guarantee, and no independent verification of it was identified.\n\nWHAT IS NOT ESTABLISHED\n\nCurrent operational status following the ruling is unclear; this site records the programme as discontinued pending confirmation.\n\nWhether the devices were withdrawn, repurposed or destroyed is not established.\n\nWhether the European Commission sought recovery of Internal Security Fund money spent on a system later ruled unlawful is not documented.\n\nHow many stops were conducted using the devices before the ruling is not published.",
      "typical_vendor": "Varies by deployment",
      "related_subject": "border-immigration",
      "url": "https://www.policeainews.com/technologies/multimodal-handheld-biometric-id"
    },
    {
      "slug": "real-time-crime-center-integration",
      "name": "Real-Time Crime Center Camera Integration",
      "category": "Surveillance",
      "summary": "A cloud platform pulling camera, sensor and dispatch feeds into one live operational picture for a police real-time crime centre, including private cameras voluntarily enrolled by businesses and households.",
      "description": "Fusus is an integration layer rather than a sensor. It ingests video, automatic number plate recognition, drone and computer-aided dispatch feeds that an agency or its neighbours already operate, and presents them as a single live operating picture for staff in a real-time crime centre. Axon acquired the company in February 2024.\n\nIts distinguishing feature is the fususCORE device, which allows private camera owners — small businesses, retailers, shopping centres, healthcare providers, schools, households — to connect their own cameras into the police platform on terms they choose. Several jurisdictions run donor programmes providing the connecting hardware free to local businesses.\n\nHOW THE COMPANY DESCRIBES IT\n\nAxon presents the platform as solving a fragmentation problem rather than adding surveillance: cameras, sensors and dispatch systems that already exist but cannot be seen together. In a February 2026 presentation at Stanford Law's CodeX, an Axon product manager stated that privacy compliance is governed by memoranda of understanding and local retention law, described the donor programme placing cores with local businesses, and stated that there is no facial recognition in the platform.\n\nThat last claim is specific and checkable, which is unusual and worth recording as such. If it changes, the change is itself a story.\n\nDOCUMENTED DEPLOYMENTS\n\nAtlanta Police Department is the reference case Axon publicises. Providence, Rhode Island signed a $750,000 contract in June 2024 running to at least July 2028, funded partly by a federal earmark. Knoxville, Tennessee runs a \"Connect Knoxville\" programme asking residents and businesses to buy fususCORE devices to link private cameras into the police platform. Lawrence, Kansas is also a customer.\n\nChula Vista, California proposed in a January 2026 council report to replace Motorola Solutions with Fusus, stating that the platform \"is not a new crime fighting tool\" but an integration layer for systems already subject to community input. Chula Vista is separately recorded on this site's tracker for Drone as First Responder.\n\nA related product, Axon Lightpost, is a streetlight-mounted plate reader that pairs with Fusus to let officers track suspect vehicles and build hotlists. This makes Axon a direct competitor to Flock Safety in plate recognition, having previously been a Flock investor with deeply integrated systems before the two separated.\n\nTHE CASE FOR IT\n\nThe fragmentation problem is real. A city may have municipal cameras, transport cameras, school cameras and dispatch data on four systems that cannot be viewed together, and in a live incident that costs time that matters.\n\nThe platform adds no new sensors of its own, which is a genuine distinction from a camera procurement. Chula Vista's framing — that the underlying systems were already subject to community input — is defensible as far as it goes.\n\nAxon's statement that the platform contains no facial recognition is a meaningful limitation, and a checkable one.\n\nTHE CASE AGAINST\n\nThe counter-argument is that integration is not neutral. A dozen camera systems that cannot be searched together are, in practice, a weaker surveillance capability than the same dozen unified, searchable and live. Chula Vista's claim that this is \"not a new crime fighting tool\" is true of each component and questionable of the whole.\n\nThe private enrolment model raises the private-to-state pathway question this site's Facewatch and Project NOLA entries already track, but at municipal infrastructure scale. A camera a business installed to watch its own doorway becomes part of a police operating picture, and the customers of that business are unlikely to know.\n\nAdvocates note that real-time crime centres accumulate large volumes of surveillance data with little public oversight, and that adopting one tends to encourage adoption of further surveillance tools — the platform creates a place to plug the next thing into.\n\nIt is also worth noting what this inverts. Project NOLA is a private non-profit hub feeding police; Fusus is a police-operated platform into which private cameras are enrolled. The direction of travel is opposite and the endpoint is similar.\n\nWHAT IS NOT ESTABLISHED\n\nHow many private cameras are enrolled in any deployment, and on what terms, is not published.\n\nRetention of the unified feed, as distinct from retention by each contributing camera owner, is not documented in the deployments reviewed.\n\nWhether the no-facial-recognition position is contractual or simply a current product state is not established.\n\nNo evaluation of whether real-time crime centres change case outcomes, as opposed to response times, has been identified.",
      "typical_vendor": "Axon (acquired Fusus, February 2024)",
      "related_subject": "surveillance",
      "url": "https://www.policeainews.com/technologies/real-time-crime-center-integration"
    },
    {
      "slug": "immigration-case-management-targeting",
      "name": "Immigration Case Management and Targeting",
      "category": "Automated decision-making",
      "summary": "Case-management and targeting software for US immigration enforcement that aggregates federal and commercial data into prioritised deportation target lists, each with a dossier and a confidence score for the target's address.",
      "description": "This category covers software that moves beyond storing case files into active operational targeting: aggregating data from multiple sources, generating a prioritised list of enforcement targets, and producing a dossier on each with an automatically generated confidence score estimating how likely a given address is to be current.\n\nThe documented deployment is a set of three related Palantir systems built for US Immigration and Customs Enforcement.\n\nTHE THREE SYSTEMS\n\nICM, the Investigative Case Management system, logs and manages criminal and civil investigation files. The contract was awarded in 2022 at $139.3 million, expiring April 2026 and slated for sole-source renewal on the stated basis that ICE cannot find a comparable system. DHS first contracted Palantir for case management in 2014. The combined deal has reportedly grown beyond $145 million.\n\nImmigrationOS, the Immigration Lifecycle Operating System, is a roughly $30 million no-bid contract modification awarded in April 2025 to augment ICM. Contract language describes streamlining selection and apprehension, providing near real-time visibility into instances of self-deportation, and managing the end-to-end immigration lifecycle from identification to removal.\n\nELITE generates a digital map of potential targets, each with a dossier containing name, date of birth, Alien Registration Number, photograph, and a confidence score out of 100 estimating the accuracy of the target's known address.\n\nTHE EVIDENTIAL HIGH POINT\n\nIn M-J-M-A v. Wamsley, in the US District Court for the District of Oregon, ICE officers confirmed under oath on 2 December 2025 that ELITE was used to identify arrest targets during Oregon raids driven by daily arrest quotas.\n\nSworn testimony is stronger evidence than any of the reporting around this system, and it establishes two things at once: that the targeting tool is operational, and that it was used in an enforcement operation organised around numerical targets.\n\nWHERE SOURCES DISAGREE\n\n404 Media reported ELITE drawing on data sourced from the Department of Health and Human Services, including Medicaid records. Palantir disputes that characterisation, calling it misleading and denying that ICE has unfettered access to HHS records through its software.\n\nWhat is not in dispute is that a data-sharing agreement between the Centers for Medicare and Medicaid Services and ICE exists and potentially covers tens of millions of Medicaid enrollees. A preliminary injunction in August 2025 temporarily blocked that sharing in twenty plaintiff states; a December 2025 ruling allowed some basic Medicaid data to be shared while the case proceeds.\n\nBoth positions are recorded because neither has been resolved.\n\nTHE CASE FOR IT\n\nImmigration enforcement agencies hold case data across many systems, and integration addresses a genuine administrative problem. Case management software is unremarkable in itself.\n\nAn address confidence score also has a defensible safety rationale: officers arriving at an address expecting one person and finding another is dangerous for everyone present, and an estimate of data reliability is better than no estimate.\n\nTHE CASE AGAINST\n\nThe combination of quota-driven operations and an automated target list is the central concern, and it is established by sworn testimony rather than inference. A system that produces a ranked list of people to arrest, deployed in an operation with daily numerical targets, creates pressure to work the list rather than to assess individual cases.\n\nThe data sources are the second issue. Aggregating federal benefit records with enforcement data means information people provided to obtain healthcare becomes an input to locating them for removal — a use nobody consented to and which, if widely understood, would plausibly deter people from seeking medical care.\n\nProcurement is the third. Sole-source and no-bid awards for systems of this consequence remove the scrutiny competitive tendering provides, and the stated justification — that no comparable system exists — is partly a consequence of the incumbent's position.\n\nAnd a confidence score out of 100 attached to an address is precisely the kind of number that acquires unwarranted authority in operational use, because it looks like a measurement.\n\nWHAT IS NOT ESTABLISHED\n\nWhether ELITE currently draws on HHS-sourced data remains contested and unresolved.\n\nHow many people have been identified through ELITE, and how many of those arrests proceeded, is not published.\n\nWhat the address confidence score is calculated from, and how accurate it is, is not documented.\n\nWhether the ICM sole-source renewal proceeded is not established here.",
      "typical_vendor": "Palantir Technologies",
      "related_subject": "border-immigration",
      "url": "https://www.policeainews.com/technologies/immigration-case-management-targeting"
    },
    {
      "slug": "retrospective-facial-recognition",
      "name": "Retrospective Facial Recognition",
      "category": "Facial recognition",
      "summary": "Post-event facial matching: an image from CCTV, a phone, or a crime scene is compared one-to-many against a reference photo database, returning ranked candidates for a trained human operator to review.",
      "description": "Retrospective facial recognition (RFR) differs from live facial recognition in a single defining way: it runs after an event, not during one. An image of an unknown person -- typically pulled from CCTV, a mobile phone, social media, or a crime scene -- is converted into a numerical representation of facial features and compared against a reference database of already-known people, most often custody photographs or a similar offender image gallery. The system does not return a single answer. It reorders the entire reference database from most to least likely match and returns a ranked list of candidates, commonly the top 100 to 200, for a trained operator to review. A human decides whether any candidate is actually a match; the software's role stops at generating and ranking the shortlist.\n\nThis is what UK police forces describe as the practical advantage: South Wales Police has stated that without RFR an identification can take around fourteen days, and with it, minutes. The technology depends entirely on the quality and content of whatever reference database a given deployment searches, which is also where it varies most between countries: Italy's SARI Enterprise searches the national AFIS photographic database and is described by Italian police as producing an investigative lead rather than a legally determinative identification; Toronto Police Service compares against lawfully obtained booking photographs using NEC's NeoFace Reveal; Victoria Police in Australia uses a system called iFace against an offender image database; York and Peel Regional Police in Canada share an Idemia-based system searching combined booking photographs. The underlying technique -- one-to-many search against a specific gallery, human adjudication of a ranked shortlist -- is consistent across all of these; what changes is which gallery is searched, under what legal authority, and with what oversight.\n\nThis is a distinct technology from a country's own specific implementation of retrospective search against a particular database -- for example, UK retrospective search against the Police National Database has its own dedicated page on this site, since the PND's scale, the vendor history around it, and the specific legal and oversight questions it has generated are substantial enough to warrant separate treatment. This page covers the general capability as deployed elsewhere.",
      "typical_vendor": "Varies by deployment: NEC, Idemia, and others",
      "related_subject": "facial-recognition",
      "url": "https://www.policeainews.com/technologies/retrospective-facial-recognition"
    },
    {
      "slug": "viogen-algorithmic-risk-assessment",
      "name": "VioGén Algorithmic Risk Assessment",
      "category": "Predictive policing",
      "summary": "Spain's national system for scoring domestic and gender-violence risk, sorting cases into five bands that determine the police protection a victim receives, in use since 2007 and audited only by reverse engineering.",
      "description": "VioGen assesses the risk faced by victims of gender and domestic violence in Spain, and the band it assigns determines what police protection they receive. It has run since 2007 under Organic Law 1/2004 on Comprehensive Protection Measures against Gender Violence, operated by the Secretariat of State for Security within the Interior Ministry. It was updated to VioGen 5.0 in 2019, with an AI pilot to refine the model from 2022.\n\nIt runs nationwide except in the Basque Country and Catalonia, which operate separately.\n\nHOW IT WORKS\n\nOfficers complete two structured questionnaires: the Police Risk Assessment at first report, and the Police Risk Evolution Assessment to reassess over time. Together they cover 35 risk indicators. The output is a weighted score placing the case in one of five bands — unappreciated, low, medium, high or extreme — and that band determines the protection measures assigned.\n\nA precision worth stating: the Eticas Foundation, which audited it, describes the underlying method as classical statistical weighting rather than machine learning. This site draws that distinction consistently, and it applies here.\n\nWHAT THE AUDIT FOUND\n\nNo dedicated external regulator oversees VioGen. Eticas offered a confidential pro-bono audit from 2018 and received no response. It eventually conducted an external audit in 2021 by reverse engineering the system with the Ana Bella Foundation — that is, the only independent examination of a system determining the safety of tens of thousands of women was conducted without the operator's cooperation.\n\nIts central finding: officers keep the system's automatic outcome in 95% of cases, despite the score being designed as a recommendation an officer may raise.\n\nThat figure matters more than any accuracy statistic. A system described as decision support, in which the human overrides the machine one time in twenty, is functionally an automated decision with a human present.\n\nTHE OUTCOMES\n\nNew York Times reporting found that at least 247 women were killed by partners after being assessed by VioGen since 2007, of whom at least 55 had been graded negligible or low risk.\n\nAs of 2025 the system carried roughly 92,000 active cases, 83% classified low or negligible risk.\n\nTHE CASE FOR IT\n\nRisk assessment in domestic abuse is a genuinely difficult task that police must perform whether or not a tool exists, and unstructured judgement in this area has a poor record. A structured instrument applying the same 35 indicators to every case is more consistent than intuition and produces a record of the reasoning.\n\nThe purpose is also protective rather than punitive. Unlike most risk scoring in this catalogue, the output determines what support a victim receives, not what happens to a suspect.\n\nSpain has maintained the system for nearly two decades and refined it, which is more sustained investment in this problem than most countries have made.\n\nTHE CASE AGAINST\n\nThe 95% concordance rate is the core problem. If the design assumption is that an officer exercises judgement over the score, and officers almost never do, the safeguard does not exist in practice.\n\nThe 55 deaths among women scored negligible or low risk are the consequence in its starkest form. No risk tool can be perfect, but the distribution of error matters when the error means no protection was assigned.\n\nThe refusal of external audit is a governance failure independent of the system's accuracy. A tool this consequential, operating for eighteen years, has been independently examined once, by reverse engineering, by an organisation the operator declined to work with.\n\nAnd 83% of active cases classified low or negligible risk is a distribution worth interrogating rather than accepting, given what the fatality figures suggest about that band.\n\nWHAT IS NOT ESTABLISHED\n\nThe current model's weightings and the effect of the 2022 AI pilot are not published.\n\nWhether the 95% concordance figure has changed since 2021 is not established.\n\nNo published data was found on how many women assessed at high or extreme risk were subsequently harmed, which would be the counterpart to the low-risk fatality figure.\n\nWhether the Basque and Catalan systems perform differently is not documented in the sources reviewed.",
      "typical_vendor": "Built for Spain's Secretariat of State for Security (Interior Ministry)",
      "related_subject": "predictive-policing",
      "url": "https://www.policeainews.com/technologies/viogen-algorithmic-risk-assessment"
    },
    {
      "slug": "ai-ml-forensic-tools-bundle",
      "name": "AI and Machine Learning Forensic Tools",
      "category": "Digital forensics",
      "summary": "The category a police force uses to name which of its forensic tools it considers AI or machine learning, as distinct from the fixed statistical systems it does not — a distinction few forces draw publicly.",
      "description": "This page exists because of a distinction one force drew in public that almost no other has: the Czech Police, in a freedom-of-information response published on its own website, named which of its forensic tools it classifies as applying machine learning, and separately named those it does not.\n\nThe tools it identified as machine learning were ABBYY FineReader for optical character recognition, BriefCam for video analytics, Magnet AXIOM for digital forensic examination, Cellebrite UFED for mobile device extraction, and speech-to-text products from Phonexia and Newton. Alongside them it named TrueAllele, STRmix, LRmix, StatistiX, CODIS and FamLinkX as fixed statistical or probabilistic systems that it does not consider AI.\n\nThat second list matters as much as the first. Probabilistic genotyping software produces likelihood ratios from DNA mixtures using defined statistical models; it is complex, consequential and contested, but it is not machine learning, and a force saying so plainly is unusually precise. Much reporting on forensic AI does not make the distinction at all.\n\nThe force's own scope statement is equally worth recording: it states these tools are used for preliminary information processing and do not independently collect, evaluate or generate evidence. That is a limitation described by the force rather than inferred from the technology.\n\nWHY THIS IS AN OVERVIEW RATHER THAN A DESCRIPTION\n\nSeveral of the products named here are substantial systems in their own right, deployed well beyond the Czech Republic, and each raises different questions. Grouping them under one heading obscured that. They now have their own pages, linked below, and this page covers only the category and the disclosure that defined it.\n\nTwo of the named tools do not have pages. ABBYY FineReader is general-purpose optical character recognition sold across many industries rather than a policing product, and its inclusion reflects that OCR uses machine learning rather than that it is a law enforcement system. Newton could not be identified with confidence from any source reviewed beyond its appearance in this disclosure, and no page has been created asserting the existence of a product this site cannot describe accurately.\n\nWHAT THIS TELLS YOU ABOUT DISCLOSURE GENERALLY\n\nThe Czech response is valuable less for what it reveals about Czech policing than as a demonstration of what a force can disclose when asked. It names products, distinguishes categories of software, states a scope limitation, and records that a fundamental-rights impact assessment under the EU AI Act is planned but cannot yet be scoped because the necessary methodology has not been published.\n\nMost forces in this catalogue have disclosed considerably less. Where entries elsewhere on this site record a technology as \"not specified in the source\", that is usually because no comparable disclosure exists.\n\nWHAT IS NOT ESTABLISHED\n\nWhether other forces use the same combination of tools is unknown; this list reflects one force's disclosure, not a survey.\n\nThe scale of use within Czech policing — how often these tools are used, in what proportion of cases — is not addressed by the disclosure.\n\nWhether the fundamental-rights impact assessment has since been completed is not established.",
      "typical_vendor": "BriefCam, Magnet AXIOM, Cellebrite UFED, Phonexia, Newton, ABBYY FineReader",
      "related_subject": "digital-forensics",
      "url": "https://www.policeainews.com/technologies/ai-ml-forensic-tools-bundle"
    },
    {
      "slug": "real-time-biometric-id-judicial-authorisation",
      "name": "AI-Supported Real-Time Biometric Identification (judicial authorisation model)",
      "category": "Facial recognition",
      "summary": "Germany's model for live facial recognition: targeted search for one named person in a defined zone, requiring judicial authorisation, deployed in Frankfurt from July 2025 — the first live facial recognition by any German state.",
      "description": "Hessen operates the first live facial recognition deployment by a German federal state. It began on 10 July 2025 in Frankfurt's Bahnhofsviertel, with reporting citing around 50 cameras in the district.\n\nThe model differs from live facial recognition as deployed in the UK in a way that matters more than the shared technology.\n\nTARGETED SEARCH, NOT A STANDING WATCHLIST\n\nSection 14(9) of the Hessian Security and Order Act permits targeted search for a specific person only where necessary to avert an actual and present or actual and foreseeable danger, to avert a terrorist offence, or to search for victims of kidnapping, human trafficking or sexual exploitation, or for missing persons recorded in police search systems. Section 14(8) additionally allows identification of an individual already flagged by pattern-recognition alerting where there is danger to life or limb.\n\nThe distinction from a UK deployment is that the system is oriented to finding one named individual, in a defined place, for a defined reason — rather than continuously screening a crowd against a standing list of many people.\n\nTHE AUTHORISATION REQUIREMENT\n\nJudicial authorisation is required in principle under Section 14(11) before deployment. Where the police authority orders use itself, judicial confirmation must be obtained within 24 hours. Each use must be logged under Section 14(10), and use must be limited in time and place to what is strictly necessary.\n\nThis is the most substantial procedural constraint on live facial recognition documented anywhere on this site. UK deployment rests on force policy assessed by a court; here a judge authorises each use in advance.\n\nHOW THE POWER WAS CREATED\n\nThe statutory basis arrived by a route worth recording. The facial recognition and pattern-recognition powers were added to the Bill by a short-notice amendment on 5 December 2024 — after the public expert hearing on the legislation had already taken place on 12 November 2024. The amended Section 14 came into force on 2 February 2025.\n\nTHE CASE FOR IT\n\nRequiring a judge to authorise each deployment addresses the central objection to live facial recognition, which is that a force decides for itself when to scan the public. Prior judicial authorisation, mandatory logging, and a strict necessity limit are meaningful controls rather than stated principles.\n\nRestricting the purpose to specific serious circumstances — terrorism, kidnapping, trafficking, missing persons — narrows the technology to uses where the case is strongest.\n\nTHE CASE AGAINST\n\nThe people scanned are the same people either way. A targeted search for one individual still requires processing the biometrics of everyone who passes the camera, and the constraint operates on what police may do with matches rather than on whose face is captured.\n\nThe legislative route is a genuine criticism. Adding a surveillance power by amendment after the expert hearing has concluded avoids the scrutiny that hearing exists to provide.\n\nCritics including netzpolitik.org note that Hessian video surveillance is not confined to statistically crime-affected locations but may cover police-defined spaces perceived as unsafe, which imports a discretionary element into where the capability sits, and raise the accuracy-disparity objection documented elsewhere on this site.\n\nWHAT IS NOT ESTABLISHED\n\nThe vendor and algorithm in use are not identified in the sources reviewed.\n\nHow many authorisations have been sought and granted, and how many refused, is not published.\n\nWhether any identification has resulted in an arrest is not established.\n\nWhether other German states intend to follow is not established here.",
      "typical_vendor": "Not specified in the source",
      "related_subject": "facial-recognition",
      "url": "https://www.policeainews.com/technologies/real-time-biometric-id-judicial-authorisation"
    },
    {
      "slug": "sa-investigative-surveillance-bundle",
      "name": "AI Investigative and Surveillance Tools (South Australia bundle)",
      "category": "Investigative analytics",
      "summary": "A set of AI-type capabilities the South Australian Auditor-General found SAPOL using or planning, spanning facial recognition, mobile phone detection cameras, social media monitoring and lead-generation analytics.",
      "description": "South Australia's Auditor-General reported that South Australia Police was using AI-type tools for facial recognition matched against CCTV and custody photographs, mobile phone detection cameras identifying drivers using a phone at the wheel, and social media monitoring. Further plans were reported for AI-assisted interview transcription and machine learning tools to identify investigative leads and previously hidden connections between cases.\n\nWHY THESE ARE BUNDLED\n\nThe Auditor-General described these collectively as part of a wider review of the force's technology use, rather than confirming each capability's operational status, scale or vendor separately.\n\nThat is a real limit on what this page can say: it is an accountability body's summary of what a force is using and planning, not a procurement record or a force disclosure naming individual systems. Where any one of these is later confirmed with a named vendor and deployment scale, it would warrant its own entry rather than remaining here.\n\nThe mobile phone detection cameras are the most concrete of the group, being a road-safety application with a defined function, and the most distinct from the others in purpose.\n\nWHAT IS NOT ESTABLISHED\n\nWhich of the reported capabilities are operational and which remain planned is not separated in the source.\n\nNo vendors are named.\n\nDeployment scale, legal basis and oversight arrangements are not documented for any of the capabilities listed.\n\nWhether the reported plans for interview transcription and lead-generation analytics have proceeded is not established.",
      "typical_vendor": "Not specified in the source",
      "related_subject": "surveillance",
      "url": "https://www.policeainews.com/technologies/sa-investigative-surveillance-bundle"
    },
    {
      "slug": "dhs-surveillance-detection-analytics-bundle",
      "name": "AI-Enabled Surveillance and Detection Analytics (US DHS bundle)",
      "category": "Investigative analytics",
      "summary": "A Government Accountability Office finding that US Department of Homeland Security law enforcement components use more than 20 categories of detection, observation and monitoring technology, some incorporating AI.",
      "description": "This entry exists because of an accountability review rather than a disclosure, and that provenance determines what can and cannot be said.\n\nA Government Accountability Office review examining fiscal year 2023 usage, published December 2024, found Department of Homeland Security law enforcement components using more than 20 categories of detection, observation and monitoring technology. These include facial recognition and automatic licence plate readers, alongside analytic software — some of which uses AI — and access to third-party plate reader networks.\n\nWHAT THIS DOES AND DOES NOT ESTABLISH\n\nThe review establishes that this breadth of capability exists across DHS components. It does not name systems, vendors, deployment scales or operational rules, because that was not its purpose.\n\nThat is a meaningfully weaker evidential basis than the entries on this site drawn from procurement records or agency disclosures, and it is recorded as such rather than presented as equivalent.\n\nWhere individual DHS-component systems are named with enough specificity to source independently — CBP's Traveler Verification Service, ICE's Mobile Fortify, Palantir's immigration systems — they are catalogued as their own entries rather than folded into this one. This page covers only what the GAO review itself established.\n\nWHAT IS NOT ESTABLISHED\n\nWhich specific systems make up the 20-plus categories is not published.\n\nCurrent usage, as distinct from fiscal year 2023, is not established.\n\nWhether the GAO's recommendations have been implemented is not established here.",
      "typical_vendor": "Multiple, not fully disclosed",
      "related_subject": "surveillance",
      "url": "https://www.policeainews.com/technologies/dhs-surveillance-detection-analytics-bundle"
    },
    {
      "slug": "remote-biometric-id-proposed",
      "name": "Remote Biometric Identification (proposed, early-stage)",
      "category": "Facial recognition",
      "summary": "An early-stage policy discussion reported by Malta's data protection authority about deploying remote biometric identification at an airport — not a procurement, trial or confirmed system.",
      "description": "Malta's Data Protection Authority states that law enforcement authorities began preliminary discussions in 2025 about deploying remote biometric identification technology at an airport.\n\nThat is the entirety of what is established. A dedicated search found no procurement notice, tender, named vendor, trial, ministerial statement or parliamentary record beyond that single reference.\n\nWHY IT IS RECORDED AT ALL\n\nThe discussion is a matter of public record from a national regulator, and the direction of policy travel is itself informative on a site tracking where police AI is heading rather than only where it has arrived.\n\nBut this entry should not be read as evidence that Malta has deployed, procured, or is close to deploying any biometric identification system. Recording the absence of further evidence is the accurate position, and this site would rather state that than fill the page with material about the technology in general that says nothing about Malta.\n\nWhere this progresses to procurement or trial, it would be reclassified and expanded accordingly.\n\nWHAT IS NOT ESTABLISHED\n\nEverything beyond the fact of preliminary discussions: vendor, technology type, airport, timetable, legal basis, and whether the discussions are continuing.",
      "typical_vendor": "Not specified; no system has been procured",
      "related_subject": "facial-recognition",
      "url": "https://www.policeainews.com/technologies/remote-biometric-id-proposed"
    },
    {
      "slug": "poland-ai-investigative-programme-proposed",
      "name": "AI Investigative Analysis Programme (Poland, proposed)",
      "category": "Investigative analytics",
      "summary": "A 2026 Polish Interior Ministry announcement of intent to apply AI across criminal investigation, evidence analysis, emergency-call handling and border facial recognition — a programme announcement, not a deployed system.",
      "description": "On 22 April 2026 Poland's Interior Minister announced the creation of a dedicated AI team covering the Police, Border Guard and the 112 emergency notification system.\n\nThe ministry's own statement describes intended uses: analysis of evidential material, with AI able to search thousands of hours of recordings or thousands of documents; detection of patterns linking apparently unconnected facts; and automated anomaly detection for cybercrime. The Border Guard is to be supported by automation covering biometrics and image recognition, including automatic detection of border violations. The minister separately stated AI would be used for facial image recognition in border protection.\n\nWHAT THIS IS\n\nA statement of policy direction. No specific named system, vendor or confirmed operational deployment was found in the material reviewed.\n\nThe ministry gives the team's principal task as developing safe and transparent rules for AI use, consistent with European data protection and civil rights standards — that is, the announced work is governance as much as capability.\n\nPolish reporting notes the announcement contained no implementation timetable, no indication of which systems would go live first, no costs, and no stated mechanism for scrutinising how the algorithms operate. The minister's stated starting point was that existing IT tools for data processing were built more than twenty years ago. The team held its first meeting in May 2026.\n\nWHAT IS NOT ESTABLISHED\n\nNo system has been confirmed as procured or deployed.\n\nNo vendor has been identified.\n\nNo timetable, budget or audit mechanism has been published.\n\nWhether the announced governance rules have been drafted is not established.",
      "typical_vendor": "Not specified",
      "related_subject": "automation",
      "url": "https://www.policeainews.com/technologies/poland-ai-investigative-programme-proposed"
    },
    {
      "slug": "veripol-ai-report-analysis",
      "name": "VeriPol AI Report Analysis",
      "category": "Automated decision-making",
      "summary": "A discontinued Spanish tool that applied natural language analysis to written robbery reports to flag statistically probable false reports, withdrawn in October 2024 three months after the EU AI Act came into force.",
      "description": "VeriPol analysed the written text of a robbery report — what a victim told police — and flagged characteristics statistically associated with false reporting. It is a materially different kind of police AI from the video and biometric systems that dominate this catalogue: it analysed language in a written account rather than an image or a biometric sample, and its subject was the complainant rather than a suspect.\n\nIt was launched nationally in October 2018 following a 2017 pilot in Malaga and Murcia, promoted at the time as more than 90% accurate. Spain's Directorate General of Police confirmed to the transparency organisation Civio that the tool stopped being used on 21 October 2024.\n\nTHE EVIDENCE BASE WAS THIN FROM THE START\n\nThe model was trained on 1,122 robbery reports from 2015. That is a small corpus for a tool deployed nationally, and analysis by the University of Valencia identified serious design deficiencies.\n\nUsage collapsed well before withdrawal: from roughly 84,000 reports analysed by October 2020 to 3,762 in 2022.\n\nWHY IT WAS WITHDRAWN\n\nThe timing is the most informative fact. The tool was withdrawn three months after the EU AI Act came into force, and this is widely reported as a factor, since the Act classifies tools of this kind as high risk and imposes obligations accordingly.\n\nThe Interior Ministry has declined to release which police stations deployed it or how often, and a transparency complaint on that point remains open.\n\nTHE CASE FOR IT\n\nFalse reporting is a real phenomenon that consumes investigative resource, and identifying it earlier would in principle free capacity for genuine victims.\n\nAnalysing text is also a less intrusive proposition than biometric identification: no image is captured, no database is searched, and no innocent third party is processed.\n\nTHE CASE AGAINST\n\nThe objection specific to this tool is what it did to the relationship between police and complainants. A system that scores a victim's account for probable falsity, applied at the point of reporting, risks discouraging reports and directing suspicion at people seeking help.\n\nThat risk is not evenly distributed. Under-reporting is already a documented problem in categories of crime where victims fear disbelief, and a tool trained to identify statistically improbable accounts will tend to flag accounts that differ from the norm — which is not the same as accounts that are false.\n\nThe training corpus of 1,122 reports cannot support a 90% accuracy claim applied nationally to a diverse population, and the University of Valencia's findings reinforce that.\n\nFinally, the transparency record. The tool ran for six years and the Ministry will not say where it was used or how often, which means no one can assess how many complainants were affected.\n\nWHAT IS NOT ESTABLISHED\n\nWhich police stations used the tool, and how many reports were assessed in total, remains undisclosed and subject to an open transparency complaint.\n\nWhat happened to complainants whose reports were flagged is not documented.\n\nWhether any report flagged by VeriPol was later shown to be genuine is not published.\n\nThe basis for the original 90% accuracy claim, beyond the 1,122-report training set, is not established.",
      "typical_vendor": "Developed for Spain's Dirección General de la Policía",
      "related_subject": "automation",
      "url": "https://www.policeainews.com/technologies/veripol-ai-report-analysis"
    },
    {
      "slug": "knightscope-autonomous-security-robots",
      "name": "Knightscope Autonomous Security Robots",
      "category": "Drones and robotics",
      "summary": "Wheeled autonomous robots that patrol a fixed site continuously with cameras, licence plate recognition, thermal imaging and acoustic sensors, recording and alerting but unable to intervene.",
      "description": "Knightscope builds wheeled autonomous security robots that patrol a fixed site continuously, recording and alerting rather than intervening. They are not police equipment in the ordinary sense: the company's customers are mostly private, and the police deployments that exist are typically a force or city placing a unit in a park, car park or campus. The robots are unarmed and have no power to detain.\n\nThe company was founded in 2013 by William Santana Li, a former automotive executive, and Stacy Dean Stephens, a former police sergeant, explicitly in response to mass-casualty events. Its flagship K5 is roughly five feet tall, weighs around 400lb, and carries 360-degree cameras, licence plate recognition, thermal imaging and acoustic sensors.\n\nHOW THE COMPANY DESCRIBES IT\n\nKnightscope's own framing is built around evidence of work performed. Its marketing states that physical security \"has run on the honor system for thirty years\" — you paid for patrols and hoped they happened — and positions the robots as producing an auditable record instead: every patrol logged, every sensor check timestamped, every action exportable. The company summarises its own pitch as trading on receipts rather than logos.\n\nOperationally the robots are sold as a subscription rather than a purchase, bundling hardware, site assessment, installation, maintenance and access to the browser-based Knightscope Security Operations Center (KSOC), through which operators issue commands, view camera and acoustic feeds, and monitor position. Company material describes the K5 as running continuously without teleoperation. A remote human layer sits above the machines, with analysts verifying incidents and coordinating response.\n\nThe stated purpose is deterrence and documentation, not force. Executives have described the aim as granting \"superpowers\" to human officers by extending situational awareness across an area without requiring a person to be present, and taking over monotonous work.\n\nWHERE IT IS DEPLOYED\n\nDeployments recorded on this site's tracker are listed below. The most-cited police example is Huntington Park, California, which was Knightscope's first police department client and whose chief subsequently promoted the experience to other agencies, directly influencing at least one further deployment. Beyond policing, units operate at retail sites, hospital campuses, casinos, utility sites and residential communities.\n\nKnightscope's own site now claims deployment across 42 US states, though that figure covers all sectors rather than policing specifically. In November 2025 the company announced the K7, a larger off-road model aimed at extensive outdoor perimeters, with an early access waitlist rather than general availability.\n\nProof News, in an August 2026 investigation, found evidence of at least 21 security robot deployments across the sector since 2015 — a useful corrective to the impression of ubiquity that marketing can create, and a figure hard to reconcile with a 42-state presence unless most of those states host very few units.\n\nTHE CASE FOR IT\n\nThe cost argument is real. Continuous physical presence is expensive, and a machine that patrols around the clock without shift patterns addresses a genuine constraint for any organisation trying to cover a large outdoor area cheaply.\n\nThe documentation argument is arguably stronger than the deterrence one. A patrol that logs itself produces an audit trail that a human patrol generally does not, and the eye-level timestamped footage has been credited with helping police identify suspects after the fact.\n\nHuntington Park published before-and-after figures for its park deployment. Comparing June to December 2018, with no robot, against the same period in 2019 with the K5 operational, the department recorded a 46% reduction in crime reports, a 10% reduction in calls for service, a 68% reduction in citations and a 27% increase in arrests. These numbers are genuinely from a police department rather than invented by the vendor, though they were published through Knightscope's press channel.\n\nTHE CASE AGAINST\n\nThe Huntington Park figures should be read with care. They are a simple before-and-after comparison of one location over two seven-month windows, with no control site, no adjustment for wider crime trends, and no separation of the robot's effect from any other change in policing that park. A 68% fall in citations alongside a 27% rise in arrests is not an obviously coherent picture, and the department's own conclusion was that the robot was \"having a positive impact\" rather than a measured effect size.\n\nIndependent findings are considerably less flattering. Proof News reported a trail of cancelled contracts and companies rethinking the model, in some cases pivoting back to human guards. It cited one city whose robot patrolled a downtown park, failed to identify any criminal incidents, and produced no tickets or arrests; the city stated it \"did not fully meet our operational needs.\"\n\nThe clearest limitation is that these machines cannot act. A robot that detects and reports but cannot intervene depends entirely on a human response arriving, and the gap between the two is where the technology's most publicised failure occurred: in Huntington Park, a bystander attempting to summon help for an assault reported being unable to get a useful response from the machine. Whatever the specifics, it illustrates a design reality that marketing language about autonomous security forces tends to obscure.\n\nThere is also an alert-quality problem. A system that generates frequent alerts without reliable fidelity increases workload rather than reducing it, which is the opposite of the force-multiplier claim, and is one reason cited for customers reverting to fixed cameras and human guards.\n\nFinally, the surveillance question. Licence plate recognition and continuous 360-degree recording in a public park is a materially different proposition from the same sensors on private property, and the public consultation preceding these deployments is often minimal because the unit is procured as security equipment rather than as a surveillance system.\n\nWHAT IS NOT ESTABLISHED\n\nNo independent evaluation of Knightscope's deterrent effect has been identified. The Huntington Park figures are the most substantial published data and they are uncontrolled.\n\nRetention and access rules for footage and plate reads captured by units patrolling public space are not documented in any deployment reviewed, nor is whether that data is shared with police as a matter of routine or on request.\n\nThe current scale of police-specific deployment is unclear, and the two available figures pull in opposite directions. The company advertises presence in 42 states; independent reporting counts 21 deployments sector-wide since 2015. Neither is necessarily wrong — a state presence can mean a single unit — but the number of active law enforcement clients specifically is not published by anyone.\n\nWhether the recorded crime reductions persist, or decay as novelty wears off, is unaddressed by any source reviewed.",
      "typical_vendor": "Knightscope, Inc. (Nasdaq: KSCP)",
      "related_subject": "robotics",
      "url": "https://www.policeainews.com/technologies/knightscope-autonomous-security-robots"
    },
    {
      "slug": "boston-dynamics-spot",
      "name": "Boston Dynamics Spot",
      "category": "Drones and robotics",
      "summary": "A remotely operated four-legged robot used by police to enter hazardous scenes such as suspected explosive devices, armed standoffs and unsafe structures, best known for the NYPD deployment that was cancelled after public backlash.",
      "description": "Spot is a four-legged robot made by Boston Dynamics, leased or bought by police forces to enter places officers would rather not: hostage scenes, suspected explosive devices, structurally unsafe buildings, and the aftermath of fires. It is remotely operated rather than autonomous in any meaningful sense — an operator directs where it goes, while onboard software handles the mechanics of walking, balancing and climbing stairs. It weighs around 32kg, and in its policing configuration carries cameras and lighting. An arm attachment allows it to open doors and move objects.\n\nHOW THE COMPANY DESCRIBES IT\n\nBoston Dynamics frames Spot around work that is hazardous or tedious for people. A company spokesperson told Scientific American that Spot has been \"particularly resourceful in tackling dull, dirty and dangerous tasks\", citing inspecting a bomb, searching the remnants of an explosion or fire, and de-escalating potentially dangerous situations. Responding to the New York controversy, the company said it supported communities reviewing the allocation of public funds and described Spot as a cost-effective tool comparable to robotic devices public safety agencies have long used for hazardous inspection.\n\nThe company's public position has consistently been that Spot is not a weapons platform, and its terms prohibit weaponisation.\n\nHOW POLICE HAVE USED IT\n\nThe most documented deployment is the New York Police Department's, which leased a Spot in 2020 under a contract worth roughly $94,000 and nicknamed it \"Digidog\". It was used in a small number of incidents: a home invasion in the Bronx in February 2021, an incident at a Manhattan public housing building, and a hostage situation in which officers strapped food to the robot and sent it in to the location.\n\nThe NYPD's operational case was made by John Miller, then deputy commissioner for intelligence and counterterrorism. He noted the department had used robots for hazardous response for nearly fifty years, and argued Spot was more agile than its predecessors because it could climb stairs, had better video recording, and was substantially cheaper than existing tracked robots costing between $200,000 and $270,000 depending on model.\n\nWHAT HAPPENED IN NEW YORK, AND WHY IT MATTERS\n\nThe deployment collapsed under public and political pressure rather than technical failure. Footage of the robot at a public housing building and at the Bronx incident circulated widely, and critics framed it as emblematic of aggressive policing in poor communities and of the wider militarisation of US police. Mayor Bill de Blasio's office said it was glad the Digidog was \"put down\", calling it creepy and alienating. Comparisons to the Black Mirror episode Metalhead were common enough to become part of the story.\n\nThe procedural failure is the more instructive part. The ACLU pointed out that the NYPD had not included the Digidog on its list of surveillance devices, which the 2020 POST Act required it to disclose. Councilman Ben Kallos and Council Speaker Corey Johnson subpoenaed the department for records relating to the device; in response, officials disclosed that the lease had been terminated on 22 April 2021, four months before it was due to expire.\n\nThe technology was not withdrawn because it did not work. It was withdrawn because it was acquired without the disclosure the law required, and because the department had not established public consent for it.\n\nIt also did not stay withdrawn. In April 2023 Mayor Eric Adams announced the robot's return as part of a wider fleet, declaring that \"Digidog is out of the pound\". That reversal is a reminder that public backlash suspends deployments more often than it ends them.\n\nTHE CASE FOR IT\n\nThe core argument is straightforward and hard to dismiss: sending a machine into a situation that might kill a person is better than sending a person. Bomb disposal, structural collapse, fire scenes and armed standoffs are the clearest examples, and police have used remote robots for exactly this purpose for decades. Spot's contribution is mobility — stairs, rubble, uneven ground — that tracked robots handle poorly, at a lower price point than the machines it replaces.\n\nThe hostage-food example is a genuinely useful illustration: a task that would otherwise require a person to approach an armed subject was performed by a machine.\n\nTHE CASE AGAINST\n\nThe strongest objection is not about the robot's function but about what it signals. A four-legged machine moving through a public housing block reads very differently to the residents of that block than a tracked bomb-disposal unit at a cordon, even where the underlying capability is similar. That perception is not irrational: it reflects a reasonable concern about which neighbourhoods experience new policing technology first.\n\nThe disclosure failure is a concrete, documented governance problem rather than a matter of perception. A department subject to a statutory obligation to publish its surveillance technologies did not list this one.\n\nThere is also a mission-creep question that the New York record raises directly. The justification offered was hazardous entry, but the deployments that caused the backlash were a home invasion and a public housing incident, neither obviously a bomb-disposal scenario. The gap between the stated use case and the observed use case is the substance of the criticism.\n\nCost framing cuts both ways. $94,000 is cheap against a $270,000 tracked robot and expensive against nothing at all — and the political argument in New York, made among others by Congresswoman Alexandria Ocasio-Cortez, was that the money should have gone to housing, education or health rather than to policing technology.\n\nWHAT IS NOT ESTABLISHED\n\nNo evaluation of operational effectiveness has been identified for any police Spot deployment. The NYPD used it in a handful of incidents and no published analysis assesses whether outcomes differed from what conventional equipment would have achieved.\n\nThe current number of police forces operating Spot is not established. Boston Dynamics told Scientific American that other US departments were testing their own units, but no comprehensive figure has been published.\n\nWhat data Spot records during a deployment, how long it is retained, and whether it is treated as evidential material or as operational telemetry, is not documented in any source reviewed. For a device that carries cameras into private dwellings, that is a substantial gap.",
      "typical_vendor": "Boston Dynamics",
      "related_subject": "robotics",
      "url": "https://www.policeainews.com/technologies/boston-dynamics-spot"
    },
    {
      "slug": "predpol-geolitica",
      "name": "PredPol / Geolitica",
      "category": "Predictive policing",
      "summary": "The best-known place-based predictive policing product, which produced daily maps of where crime was likely to occur, found by independent analysis to predict at under half a percent accuracy before ceasing operations in 2023.",
      "description": "PredPol was the best-known predictive policing product in the world, and its collapse is the most thoroughly documented failure in this catalogue. It ingested historical crime incident reports and produced daily predictions of where and when specific crime types were most likely to occur, marking small boxes on a map for officers to patrol. Created in 2010, it was a leading vendor by 2012, rebranded to Geolitica in 2021, and ceased operations at the end of 2023.\n\nIts remains were absorbed by SoundThinking, the company formerly known as ShotSpotter, which hired Geolitica's engineering team, acquired patents, and migrated its customers onto its own patrol platform. SoundThinking had already bought HunchLab, a competing predictive policing product, from Azavea in 2018.\n\nOne detail of that acquisition is unusually revealing. SoundThinking Senior Vice President Sam Klepper stated that the company \"was not interested in and did not purchase the source code for the modeling of the Geolitica prediction system\", and that the algorithms it did take \"do not include any crime prediction modeling source code at all.\" The buyer of a crime prediction company explicitly declined to buy the crime prediction.\n\nHOW IT WAS SOLD\n\nThe pitch was resource allocation on an evidential basis rather than intuition: feed the system what has already happened, and it tells patrols where to be. The underlying model drew on earthquake aftershock mathematics, on the theory that crime clusters in space and time in comparable ways. In its later years the company shifted its positioning away from prediction and towards being a broader platform for managing police department data, which is itself an indication of how the core claim was faring.\n\nWHAT INDEPENDENT ANALYSIS FOUND\n\nTwo investigations by The Markup are the substance of what is known, and both are unusually well evidenced because they rest on the company's own outputs rather than on interviews.\n\nThe first, published with Gizmodo in December 2021, analysed predictions Geolitica had supplied to police departments in 38 jurisdictions over nearly three years. The data was obtained from an unprotected cloud storage server linked from a page on the Los Angeles Police Department's own website. It found the software disproportionately directed officers to neighbourhoods with higher proportions of low-income, Black and Latino residents relative to the cities as a whole, while predictions for white middle and upper-class areas were largely absent.\n\nThe second, in October 2023, tested accuracy directly. The Markup examined 23,631 predictions generated for the Plainfield Police Department in New Jersey between 25 February and 18 December 2018. Fewer than 100 lined up with a crime of the predicted category subsequently reported to police. The success rate was below half of one percent. The Markup's own summary was that the software predicted crime at a rate slightly better than throwing darts at a map.\n\nThe most consequential finding is the least quoted. Comparing predictions against arrest data from eleven departments willing to share it, The Markup found arrest rates in predicted areas stayed the same whether or not a prediction had been made. The software did not measurably change what policing achieved.\n\nWHAT THE CUSTOMERS SAID\n\nPlainfield paid $20,500 for an annual subscription and a further $15,500 to extend. Captain David Guarino's account is worth quoting at length because it is rare for a force to be this direct: asked why the department bought it, he said they wanted to be more effective at reducing crime and thought knowing where to be would help, adding \"I don't know that it did that.\" He continued: \"I don't believe we really used it that often, if at all. That's why we ended up getting rid of it.\" He told The Markup the money could have been better spent elsewhere.\n\nPlainfield was the only one of 38 departments willing to provide information.\n\nTHE STRUCTURAL CRITICISM\n\nBeyond accuracy, the persistent objection is that a model trained on recorded crime learns where police have previously been, not where crime has previously happened — and those are different things. Directing patrols to those areas generates more recorded incidents there, which feeds back into the model. The scholar Ruha Benjamin characterised PredPol as a \"crime production algorithm\" on exactly this reasoning: officers patrol predicted zones expecting to find crime, and the expectation is self-fulfilling.\n\nThis is not a criticism that better engineering resolves. It is a property of training a predictive system on enforcement data and then using its output to direct enforcement.\n\nIS THERE A CASE FOR IT\n\nIn fairness, the underlying problem is real: patrol resources are finite and have to be allocated somehow, and the alternative to a model is officer intuition, which carries its own well-documented biases. Place-based prediction was also, in principle, a narrower intrusion than scoring individuals, since it directs attention to a location rather than labelling a person.\n\nBut that defence requires the predictions to work, and the strongest available evidence is that these did not. A model that performs at under half a percent, and produces no measurable change in arrest outcomes, does not improve on intuition — it launders it.\n\nWHAT IS NOT ESTABLISHED\n\nThe Markup's accuracy analysis rests on a single department over ten months. It is the most rigorous public assessment that exists, but Plainfield was the only force of 38 that cooperated, so whether that success rate was typical is unknown.\n\nWhy 37 departments declined to share arrest data is not established, and no comprehensive list of Geolitica's customers has been published.\n\nWhat happened to the prediction data already generated for those 38 jurisdictions, and whether it was deleted when the company wound down, is undocumented. The 2021 investigation established that at least some of it had been sitting on an unprotected server.\n\nWhether SoundThinking's successor products reproduce the same weaknesses is untested. The company states it did not acquire the prediction modelling, but both its platform and Geolitica's advertised incorporating risk terrain modelling, and no independent evaluation of the replacement has been identified.",
      "typical_vendor": "Geolitica (formerly PredPol); wound down 2023, assets absorbed by SoundThinking",
      "related_subject": "predictive-policing",
      "url": "https://www.policeainews.com/technologies/predpol-geolitica"
    },
    {
      "slug": "nec-neoface",
      "name": "NEC NeoFace",
      "category": "Facial recognition",
      "summary": "NEC's facial recognition algorithm, used by South Wales Police and the Metropolitan Police across live, retrospective and handheld deployments, and the most independently evaluated system in this catalogue.",
      "description": "NeoFace is NEC's facial recognition algorithm, and it is the software behind most of what the United Kingdom calls police facial recognition. South Wales Police and the Metropolitan Police both use it, across all three of their use cases: live facial recognition at deployments, retrospective searching of recorded images, and operator-initiated facial recognition on an officer's handheld device. Where UK reporting refers to \"the algorithm\" without naming it, this is usually the one meant.\n\nIt is sold as a family rather than a single product. NeoFace Watch handles real-time CCTV matching, NeoFace M40 is the algorithm version South Wales Police names in its own public material, NeoFace V4 is the version the National Physical Laboratory evaluated, and NeoFace Nexus is the current platform NEC markets for investigative use. Version matters when reading accuracy claims, because they are version-specific.\n\nHOW THE COMPANY DESCRIBES IT\n\nNEC's own framing leans heavily on independent benchmarking. Its material states the technology has been proven by the US National Institute of Standards and Technology as the fastest and most accurate in the world, and as one of the most consistently equitable across demographics. It describes NeoFace as designed to find a needle in a haystack — reducing a large problem to a smaller review process — and as tolerant of ageing, angles, headwear and lighting.\n\nThe company also markets governance features rather than only performance: built-in privacy safeguards, auditing and reporting, and automatic and manual data-retention tools. NEC states it is guided by long-established internal principles on biometrics and AI encouraging ethical, proportionate and targeted use.\n\nThe NIST claim is genuinely substantiated in a way most vendor accuracy claims are not. NEC's algorithms have repeatedly placed among the top performers in NIST's Face Recognition Technology Evaluation for one-to-many identification, which is an independent benchmark the company does not control.\n\nHOW IT IS USED IN POLICING\n\nSouth Wales Police was the first UK force to deploy it, from July 2017, and the first anywhere in the UK to use real-time facial recognition at a large-scale sporting event. Cameras were mounted on police vehicles, streaming to the algorithm for matching against authorised watchlists of wanted people, suspects, missing persons and vulnerable individuals. The same system cross-checks crime scene images against the force's custody image database, then around 500,000 photographs.\n\nNEC stated it had NeoFace Watch deployments in 47 countries at that point, a figure the company later put higher.\n\nThe three modes carry different intrusion profiles. Live facial recognition scans everyone passing a camera against a watchlist. Retrospective search takes an image after the fact and searches a custody database. Operator-initiated recognition lets an officer photograph a person in front of them and check that image. South Wales has used the handheld mode since 2024; the Met announced its own from December 2025.\n\nTHE ACCURACY EVIDENCE, AND WHAT IT ACTUALLY SHOWS\n\nThis is the best-evidenced algorithm in this catalogue, because the National Physical Laboratory published an operational evaluation in April 2023 rather than relying on laboratory benchmarks alone.\n\nFor retrospective and operator-initiated use the findings were strong: a true positive identification rate of 100%, with no significant difference by gender or ethnicity. Even critics of facial recognition policing have acknowledged this squarely — the campaign group StopWatch noted the finding addresses some of the concerns people hold about police use, while arguing accuracy is not the only question that matters.\n\nFor live facial recognition the picture is more conditional, and the condition is the threshold setting. NPL found that at the default 0.6 threshold the Met would see roughly one false alert in 6,000. At 0.64, the algorithm returned no false positives at all. But at 0.58 and 0.56, false positives rose — and crucially, they rose unevenly, with Black people incorrectly flagged as possible matches more often than people of other ethnicities.\n\nSouth Wales Police says it typically operates at 0.62. Since the Court of Appeal judgment in Bridges, the force states testing confirmed its use does not discriminate on grounds of gender, age or race, and it resumed deployments on 5 April 2023. It publishes deployment dates and locations in advance except in exceptional circumstances, and states biometric data of anyone not on a watchlist is deleted immediately and automatically.\n\nTHE CASE FOR IT\n\nThe independent evidence base is the strongest argument, and it is unusual. Most systems in this catalogue rest on vendor claims; this one has been evaluated by a national measurement institute in operational conditions, and the results for two of its three modes were unambiguous.\n\nThe governance features are also real rather than rhetorical. Advance publication of deployment locations, immediate deletion of non-matches, and audit logging are meaningful controls, and their existence is verifiable from force material rather than only from marketing.\n\nTHE CASE AGAINST\n\nThe single most important finding is one neither the vendor nor most reporting foregrounds: equitable performance is a function of a threshold setting, and no law or formal policy requires any particular one. The same algorithm is demographically fair at 0.62 and demonstrably not at 0.56. A force can therefore comply with every published assurance about accuracy while operating at a setting where the disparity reappears, and the Met's own facial recognition lead has suggested lower thresholds might be appropriate in emergencies such as an ongoing terrorist attack — precisely the circumstances in which a wrongful match would be most consequential.\n\nA 1-in-6,000 false alert rate also sounds small and is not, at scale. A deployment scanning tens of thousands of faces produces false alerts as a matter of routine, and each one is a person stopped by police on a machine's say-so.\n\nThere is a broader objection that accuracy improvements do not answer. A perfectly accurate system still enables identification of everyone passing a camera, and the question of whether that should happen is separate from whether the machine gets it right. StopWatch's argument is essentially this: fixing the algorithm addresses the least fundamental of the objections.\n\nFace occlusion also degrades performance — masks, scarves, sunglasses, oblique angles — which is relevant given that watchlist subjects have obvious reason to obscure their faces.\n\nWHAT IS NOT ESTABLISHED\n\nWhich threshold each UK force operates at, deployment by deployment, is not systematically published. South Wales states its usual figure; comprehensive disclosure across forces does not exist.\n\nThe NPL evaluation covers NeoFace V4. Forces have since referenced M40 and NEC now markets Nexus. Whether the published equitability findings carry across versions has not been independently re-established.\n\nNEC's current deployment count in policing specifically is unclear. Figures of 47 and later 70 countries refer to all customers including airports and commercial operators, not to police forces.\n\nNo published data was found showing what proportion of NeoFace alerts across UK deployments resulted in an arrest, a charge or a wrongful stop, which is the outcome measure that would let anyone judge whether the accuracy figures translate into proportionate policing.",
      "typical_vendor": "NEC Corporation / NEC Software Solutions UK",
      "related_subject": "facial-recognition",
      "url": "https://www.policeainews.com/technologies/nec-neoface"
    },
    {
      "slug": "idemia-biometric-identification",
      "name": "Idemia Biometric Identification",
      "category": "Biometrics",
      "summary": "French-supplied fingerprint and facial recognition systems embedded in national police identification infrastructure, including Interpol, Australia's NAFIS and a shared facial recognition network across Southern Ontario.",
      "description": "Idemia is a French biometrics company, headquartered at Courbevoie, whose public security division supplies fingerprint and facial recognition systems to police forces, national agencies and Interpol. It is less visible in public debate than Clearview AI or NEC, but is arguably more deeply embedded: its systems sit inside national identification infrastructure in several countries rather than being deployed as visible cameras on streets.\n\nIts policing footprint includes Interpol's Multibiometric Identification System, the Kansas Bureau of Investigation's automated biometric identification system, work with France's Gendarmerie Nationale, and Australia's NAFIS NextGen biometric identification system delivered with the Australian Criminal Intelligence Commission. In Canada it supplies the shared facial recognition network used by police across Southern Ontario.\n\nHOW THE COMPANY DESCRIBES IT\n\nIdemia's public positioning is notably more measured than most vendors in this catalogue, and that is worth recording accurately rather than dismissing.\n\nIts central accuracy claim is anchored in independent benchmarking: the company states that fairness is provided by design in its identification algorithms, as demonstrated by independent public evaluation of their performance by NIST. It publicises top rankings across NIST fingerprint and face benchmarks and a top position in the US Department of Homeland Security's RIVTD evaluation for biometric accuracy and fairness.\n\nMore unusually, the company publicly concedes limits. Its own material states that it would be irresponsible to ignore facial recognition criticisms and claim the technology is flawless, that it is not flawless, and that it should not be the sole means of identifying individuals without review by a properly trained examiner. It also makes a candid technical point about a constraint on the whole field: the US Privacy Act of 1974 prevents federal agencies sharing even anonymised extracts from operational systems, which limits the data available for training and testing algorithms.\n\nIdemia frames responsible deployment through the Biometrics Institute's Three Laws of Biometrics — policy first, process follows policy, technology guided by both — and states that concerns about bias and misidentification arise when systems are deployed without appropriate governance, testing and safeguards. That is a claim about deployment conditions rather than about the technology, and it puts the responsibility on the purchasing agency.\n\nTHE INDEPENDENT ACCURACY RECORD\n\nIdemia's NIST results are genuinely strong. In NIST's Face Recognition Technology Evaluation 1:N ranking published 8 November 2024, covering more than 150 algorithms, Idemia placed second in the mugshot category at a gallery size of 12 million. A later submission recorded the fourth-lowest false negative identification rate at a false positive identification rate of 0.003.\n\nFor a police force conducting one-to-many searches against a custody image database, those are the directly relevant measures, and they are produced by an independent US government body rather than the vendor.\n\nHOW POLICE USE IT\n\nThe Canadian deployments are the most transparently documented and illustrate the operating model. Halton Regional Police adopted Idemia's system alongside neighbouring Peel and York, and Ottawa Police Service Board approved onboarding to the same shared network in June 2026.\n\nHalton's Deputy Chief Jeff Hill described the intended discipline clearly: only legally obtained images are compared against mugshots held in the force's system, and any match is treated as an investigative lead rather than an identification. In his words, identifying someone does not mean they are automatically guilty, and a full background investigation must corroborate the result.\n\nCosts are unusually well documented for Ottawa: a CA$150,000 deposit in the first year against a one-time implementation cost of CA$472,631, then CA$155,000 in annual maintenance from 2027 rising to CA$161,477 by 2031, for a total of roughly CA$1.26 million.\n\nInterpol's use gives a sense of scale at the other end: having added Idemia facial recognition capability in 2016, the organisation reported matching and identifying more than 300 suspects in 2019 across its member countries.\n\nTHE CASE FOR IT\n\nThe accuracy evidence is independent, current and specific to the task police actually perform. That distinguishes it from vendors whose figures come from their own testing.\n\nThe stated operating discipline is the right one. Treating a match as an investigative lead requiring corroboration, rather than as an identification, is precisely the control that prevents the wrongful-arrest cases documented elsewhere in this catalogue. Where forces genuinely apply it, the risk profile is materially different from systems used as a basis for arrest.\n\nThe company's willingness to state publicly that its technology is not flawless and should not be used as a sole identifier is, on the evidence reviewed, unusual among biometric vendors and worth noting.\n\nTHE CASE AGAINST\n\nThe most significant issue is not accuracy but aggregation, and it is visible in the Canadian model. Because Halton, Peel, York and now Ottawa share one network, each force joining expands the pool of photographs every other force can search. Halton's own reporting noted this explicitly as a benefit. The consequence is that the searchable database grows through a series of separate local procurement decisions, with no single point at which anyone decides how large it should become. Halton's chief has said openly that a national database would be preferable.\n\nThat is a governance structure in which scope creep is the default rather than the exception, and it happens without the public debate a national biometric database would attract if proposed directly.\n\nThe lead-not-identification discipline also depends entirely on being followed. It is a policy commitment described by senior officers, not a technical constraint on the system, and no published audit was found testing whether corroboration actually occurs before action is taken.\n\nIdemia's framing that problems arise from deployment rather than technology is defensible, but it also transfers responsibility to purchasers who, as the UK Biometrics Commissioner has observed of forces generally, frequently acquire such systems without fully understanding their capabilities or implications.\n\nFinally, the training-data constraint the company itself identifies cuts both ways. If legal restrictions genuinely limit the data available to test algorithms against operational populations, then NIST benchmark performance may not transfer cleanly to the specific demographic mix of any given jurisdiction.\n\nWHAT IS NOT ESTABLISHED\n\nThe total size of the shared Southern Ontario photo pool, and how it grows as forces join, has not been published.\n\nNo independent audit was found testing whether the investigative-lead discipline is applied in practice in any Idemia deployment.\n\nWhich specific Idemia algorithm version each police customer runs is generally not disclosed, so the NIST rankings cannot be reliably mapped onto a given deployment.\n\nThe full extent of Idemia's police customer base is unclear. The examples above are those the company or its customers have publicised; no comprehensive list exists.",
      "typical_vendor": "Idemia Public Security (France)",
      "related_subject": "biometrics",
      "url": "https://www.policeainews.com/technologies/idemia-biometric-identification"
    },
    {
      "slug": "magnet-axiom",
      "name": "Magnet AXIOM",
      "category": "Digital forensics",
      "summary": "A digital forensics platform consolidating evidence from phones, computers, cloud accounts and vehicles into one searchable case file, including a machine learning layer that classifies and flags communications content.",
      "description": "Magnet AXIOM is a digital forensics platform used to recover, analyse and report on evidence pulled from phones, computers, cloud accounts, vehicles and IoT devices, bringing all of it into a single case file. It is made by Magnet Forensics, based in Waterloo, Ontario, and founded in 2011 by Jad Saliba, a former police officer who built the original tool because the job needed one.\n\nIt belongs on a catalogue of policing AI because of one specific component rather than the platform as a whole. Most of AXIOM is conventional forensic software: parsing file systems, carving deleted files, decrypting disks. The AI element is Magnet.AI, a machine learning layer that reads communications content and flags material of investigative interest — introduced for contextual chat analysis and marketed as an industry first when it launched.\n\nHOW THE COMPANY DESCRIBES IT\n\nMagnet Forensics positions AXIOM around consolidation and speed. Its material describes examining evidence from mobile, cloud, computer and vehicle sources in one case file, using analytical tools to surface case-relevant artefacts automatically. The company states its software is used by more than 4,000 customers in over 100 countries, and that its purpose is to help investigators fight crime, protect assets and guard national security.\n\nThe platform is split into two applications. AXIOM Process acquires and parses data, recovering material from zip archives and deleted files and sorting it by type. AXIOM Examine presents what Process found for review in a case database, searchable by date range, keyword and other filters.\n\nIts distinguishing design choice is what the company calls an artefact-first approach: prioritising the material investigators actually want — chats, images, browser history, location data — over a raw file system view. A Connections feature draws a graphical map showing how artefacts relate to one another across devices.\n\nMagnet's own framing of the problem is worth recording because it is accurate: a senior product manager described investigators previously having to toggle between different tools with different report formats to correlate artefacts manually, risking missed evidence and missed connections.\n\nWHERE IT SITS IN POLICE USE\n\nDeployments recorded on this site's tracker are listed below. The Czech Police names Magnet AXIOM among the tools it classifies as applying machine learning technologies, in its own freedom-of-information disclosure — one of the few instances where a force has publicly stated which of its forensic tools it considers AI.\n\nMagnet Forensics reports that roughly 65% of its customers are public sector and about 60% North American. It sits alongside rather than in place of extraction hardware: many labs image a phone with one tool and analyse it in AXIOM, and the company sells its own acquisition products, Graykey for device access and Verakey for consent-based extraction, into the same workflow.\n\nTHE CASE FOR IT\n\nThe underlying problem is real and growing. Digital evidence features in a large and rising share of investigations, and the volume on a single modern phone can exceed what an examiner can review manually within any reasonable time. Practitioner accounts of AXIOM consistently praise its filtering as the thing that makes large datasets tractable — one reviewer describing recursive filtering as what pares a trove of data down to the material that matters.\n\nThe consolidation argument is also sound. Correlating artefacts across a phone, a laptop and a cloud account by hand is error-prone, and a single case file with a visual map of relationships between artefacts reduces a genuine source of missed evidence.\n\nReporting quality matters more than it sounds. The ability to produce a portable case file that presents evidence coherently to lawyers and courts is a practical benefit in a field where the alternative is a set of incompatible exports.\n\nTHE CASE AGAINST\n\nThe core tension in mobile forensics is not specific to this product but is sharpened by it: an extraction captures everything, and analysis tools make everything searchable. A phone examined in connection with one offence yields years of messages, photographs, location history and cloud account contents belonging to the device owner and to everyone who ever messaged them. The artefact-first design that makes AXIOM effective is precisely what makes the whole of a person's digital life immediately legible.\n\nCloud access extends this further. The platform can be used to log into cloud-based accounts and retrieve deleted data and location history, which reaches material that is not on the seized device at all. The civil liberties research project AFSC Investigate lists Magnet Forensics on that basis.\n\nThe AI component carries a distinct risk that the conventional forensics does not. Magnet.AI classifies conversations and flags them as potentially relevant — for instance as indicative of a particular category of offending. A machine judgement about what a conversation means is a different kind of claim from recovering a deleted file, and it is the kind of claim that needs to be testable in court. No independent evaluation of Magnet.AI's classification accuracy was identified in this research.\n\nThere is also an evidential dependency question. Where a tool's output is presented as evidence, the reliability of the tool becomes material to the case, and commercial forensic software is generally not open to independent inspection.\n\nWHAT IS NOT ESTABLISHED\n\nNo independent accuracy evaluation of Magnet.AI's chat classification was found. Its error rate, and whether performance varies by language or context, is undocumented publicly.\n\nWhether UK forces use the Magnet.AI component specifically, as opposed to AXIOM's conventional forensic functions, is not established by any source reviewed. Naming the platform does not establish which features are enabled.\n\nHow forces constrain the scope of an AXIOM examination — whether analysis is limited to material relevant to the offence under investigation, or whether the full extraction is searchable — is not documented in the deployments reviewed, and is the question that most determines how intrusive the tool is in practice.\n\nNo published data was found on how often digital forensic examinations using this or comparable platforms produce evidence that changes a case outcome.",
      "typical_vendor": "Magnet Forensics (Waterloo, Ontario)",
      "related_subject": "digital-forensics",
      "url": "https://www.policeainews.com/technologies/magnet-axiom"
    },
    {
      "slug": "phonexia-voice-biometrics",
      "name": "Phonexia Voice Biometrics",
      "category": "Biometrics",
      "summary": "Czech speech technology used by police and forensic units to identify speakers by voice across recordings, bundled with transcription, language identification, deepfake detection and contested emotion and age estimation.",
      "description": "Phonexia is a Czech speech technology company, based in Brno and founded in 2006 as a spin-off from Brno University of Technology. Its Speech Platform bundles fourteen modular technologies covering voice biometrics and speech recognition, and it sells primarily to government, law enforcement and forensic units, reporting a presence in more than 60 countries.\n\nIts core capability is speaker identification: determining whether the voice in two recordings belongs to the same person, independent of what was said, in what language, or over what channel.\n\nHOW THE COMPANY DESCRIBES IT\n\nPhonexia frames speaker identification around two distinct questions, and the distinction is a useful one. Identification asks \"whose voice is this?\", typically as a one-to-many or many-to-many comparison — the company gives the example of fake emergency calls, and states that large-scale automatic identification is used by law enforcement for database searches and ranking suspects. Forensic voice analysis is the narrower later-stage task, using smaller amounts of data and one-to-one comparison to evaluate evidence, establish the probability of a speaker's identity, and produce something usable in court.\n\nThe technical premise, in the company's own account, is that speech organs and speaking habits are more or less unique to a person, so the features of a captured speech signal are also more or less unique. The hedging in that phrasing is the vendor's own, and is more careful than most biometric marketing.\n\nPhonexia claims over 98% accuracy in evaluations, participates in the NIST Speaker Recognition Evaluations, and works with a NIST-recognised speech research group at Brno University of Technology. Its platform is built exclusively on deep neural networks, extracting around a hundred feature vectors per second of audio.\n\nDeployment is deliberately on-premises or private cloud, with the company stating that sensitive data never leaves the customer's servers. Phonexia also notes it is a private Czech company with a transparent ownership structure, governed by EU legislation — a positioning aimed at buyers wary of biometric suppliers from other jurisdictions.\n\nWHAT ELSE IS IN THE PLATFORM\n\nBeyond speaker identification, the suite includes speaker diarization (separating and labelling individual speakers in a single recording), language identification across 140 languages, transcription in more than 60, keyword and phoneme spotting, voice activity detection, speech quality measurement, voice deepfake detection, and estimation of a speaker's age, gender and emotional state.\n\nWHERE IT IS USED\n\nDeployments recorded on this site's tracker are listed below. The Czech Police names Phonexia among the tools it classifies as applying machine learning, in its own freedom-of-information disclosure.\n\nPhonexia was also the voice biometrics provider in ROXANNE, an EU Horizon 2020 project bringing together 25 partners across 15 countries, including 11 law enforcement agencies, to build a platform for investigating organised criminal networks. The resulting Autocrime platform combined speaker identification with multilingual speech recognition, gender identification, keyword and topic detection, named entity recognition and network analysis, and the consortium said it would be offered to European law enforcement under an open-source licence. The project also produced ROXSD, a synthetic but realistic dataset representing a fictional organised crime network, released for research use.\n\nTHE CASE FOR IT\n\nSpeaker identification addresses a genuine investigative problem. Intercepted or recorded audio frequently features unidentified voices, and comparing recordings by ear is slow and unreliable. A system that ranks candidates from a voiceprint database gives investigators a starting point rather than an answer.\n\nThe company's separation of investigative identification from forensic one-to-one comparison is methodologically sound, and it matters: using a database search result as a lead is defensible, whereas presenting the same output as courtroom evidence is not, and Phonexia's own documentation draws that line explicitly.\n\nOn-premises deployment is a real advantage for this class of data. Voice recordings from investigations are among the most sensitive material a force holds, and a model where audio never leaves the customer's infrastructure avoids an entire category of risk that cloud-hosted biometrics carry.\n\nThe synthetic ROXSD dataset deserves credit too: releasing realistic training data that contains no real people's communications is a better answer to the research-data problem than the alternatives.\n\nTHE CASE AGAINST\n\nTwo capabilities in the wider suite warrant more scrutiny than the speaker identification does.\n\nEmotion recognition from voice is scientifically contested. The claim that a machine can identify a speaker's emotional state from acoustic features is not settled science, and the EU AI Act treats emotion recognition as a restricted category in several contexts precisely because of that. Its inclusion in a law enforcement platform, alongside age and gender estimation, invites inferences about people that the underlying evidence may not support.\n\nThe detention use case is the more striking. Phonexia's own material describes detecting suspicious behaviour in detention centres including \"misuse of call allowances or use of prohibited languages\" to ensure inmates comply with phone rules. Automated detection of which language a detained person is speaking, in order to enforce a prohibition on speaking it, is a significant capability to see described as a routine product feature, and it sits close to the concerns raised elsewhere in this catalogue about continuous monitoring of prisoner communications.\n\nThe accuracy claim also needs reading carefully. Over 98% in evaluations does not specify which evaluations, at what dataset size, or under what audio conditions. Speaker identification performance degrades with short samples, background noise, poor channels and deliberate disguise, and none of those are addressed by a single headline figure.\n\nFinally, one-to-many identification depends on a voiceprint database existing. Building one means enrolling voices, and the governance around whose voice is enrolled, on what basis, and for how long, is not addressed by the technology and is rarely addressed publicly by the forces using it.\n\nWHAT IS NOT ESTABLISHED\n\nWhich specific evaluations produce the 98% figure, and under what conditions, is not published.\n\nNo independent assessment of the emotion recognition or age estimation components was identified, nor any statement of whether police customers enable them.\n\nHow voiceprint databases are governed in any deployment reviewed — enrolment criteria, retention, deletion — is undocumented.\n\nWhether the Autocrime platform from ROXANNE was adopted operationally by any of the 11 participating law enforcement agencies after the project concluded is not established by the sources reviewed.",
      "typical_vendor": "Phonexia (Brno, Czech Republic)",
      "related_subject": "biometrics",
      "url": "https://www.policeainews.com/technologies/phonexia-voice-biometrics"
    },
    {
      "slug": "smartview-hong-kong-cctv",
      "name": "SmartView (Hong Kong territory-wide CCTV)",
      "category": "Surveillance",
      "summary": "Hong Kong's police-run CCTV programme, already carrying AI crowd-size analysis and number plate recognition across about 11,000 connected cameras, and funded to reach roughly 60,000 by 2028.",
      "description": "SmartView is the Hong Kong Police Force's territory-wide closed-circuit television programme. Cameras are mounted on lampposts, smart lampposts, government premises and police vehicles, and feed a central system that officers can search. Camera sites are chosen using crime rates and district data, pedestrian flow, risk and the public order situation.\n\nWhat distinguishes it from an ordinary municipal camera network is that software, not only officers, watches the footage.\n\nWHAT THE SYSTEM DOES TODAY\n\nThe Legislative Council Panel on Security paper of 24 February 2026 states that the system 'is currently equipped with artificial intelligence (AI)-powered functions, such as crowd size analysis and automatic number plate recognition'. Those two functions are the established present-day capability. Everything else described below is planned.\n\nTHE SCALE AND THE MONEY\n\nThe same paper puts the network at about 5,000 cameras installed by the police, with about 6,000 cameras belonging to other government departments and organisations connected to it. Over 20,000 cameras are to be added annually, reaching about 60,000 across the city in 2028, with about 6,500 more for newly developed areas by 2031.\n\nNon-recurrent cost is given as HK$4.066 billion across 2026-27 to 2031-32, split between HK$893.2 million of hardware, HK$1.311 billion of communications network and HK$956.4 million of implementation services. Recurrent costs start at HK$72.9 million a year and are projected to reach HK$243.1 million by 2036-37.\n\nThe paper also records that the programme has assisted police in detecting 899 cases, including serious crimes such as murder and robbery. That figure is the force's own, is not broken down by offence or by which capability contributed, and has not been independently audited.\n\nTHE ANALYTICS TOOLBOX IS THE PART THAT MATTERS\n\nThe paper sets out a plan to 'build an AI-powered video analytics toolbox to analyse the CCTV footages', so that police 'will be able to track and search for target persons and vehicles in a speedier and more precise manner'. It also states an intention to introduce more AI-powered video analytic functions for crowd management.\n\nThat is a change in kind rather than degree. Searching recorded footage for a named person across a city-wide network is a different capability from counting a crowd, and it is the point at which the number of cameras stops being the main constraint on what the system can do.\n\nFACIAL RECOGNITION IS ANNOUNCED, NOT ESTABLISHED\n\nSenior officers have been reported during 2025 and 2026 as intending to add facial recognition to the network. Facial recognition appears nowhere in the February 2026 LegCo paper and nowhere on the Hong Kong Police Force's own SmartView page, both of which were read directly. Until it appears in a government document or a force publication, it should be treated as a stated intention.\n\nTHE CASE FOR IT\n\nHong Kong is dense, vertical and heavily built, and camera coverage in public space is uncontroversial in much of the world. Automatic number plate recognition and crowd-size estimation are among the better-understood forms of video analytics: the first is a mature technology with a long enforcement record, and the second does not identify individuals at all. The 899 detected cases, whatever their composition, represent investigative work that officers would otherwise have done more slowly or not at all. Concentrating cameras by crime rate and pedestrian flow is a defensible allocation method, and connecting existing departmental cameras avoids duplicating public spending.\n\nLEGAL BASIS AND OVERSIGHT\n\nThere is no dedicated statute. The force states that the scheme operates under the auspices of the Task Force on District Governance, that cameras capture public places only, and that it complies with the relevant laws of Hong Kong. It says it maintains close liaison with the Office of the Privacy Commissioner for Personal Data and commissioned independent consultancies to carry out Privacy Impact and Security Risk Assessments, and that it has internal guidelines for monitoring standards.\n\nThe February 2026 paper budgets for a Privacy Impact Assessment but names no independent body with oversight of the analytics themselves. Procurement uses a direct engagement approach rather than competitive tendering, which the paper attributes to national security and supply chain considerations.\n\nWHAT IS NOT KNOWN\n\nThe accuracy of the deployed crowd-analysis and number plate functions has not been published. Neither has the retention period for footage, the threshold at which analytics are run against a recording, who may authorise a search of the network, or how many such searches have been made. No published evaluation exists of the 899-case figure. Whether the analytics toolbox will include facial recognition, and under what authorisation, is unresolved.",
      "typical_vendor": "",
      "related_subject": "surveillance",
      "url": "https://www.policeainews.com/technologies/smartview-hong-kong-cctv"
    },
    {
      "slug": "deepfake-detection-forensics",
      "name": "Police Deepfake and Synthetic Media Detection",
      "category": "Digital forensics",
      "summary": "Classifiers that judge whether a seized video is synthetic, used by police to direct investigations rather than to prove them; the best-documented system reports about 80 per cent accuracy.",
      "description": "A deepfake detection tool takes a video file and returns a judgement on whether it was generated or manipulated. It is a classifier, not a proof: it produces a probability, and the probability is wrong some of the time in both directions.\n\nHOW DETECTION ACTUALLY WORKS\n\nThree families of signal are used in the systems documented so far. The first is artefacts of face generation and replacement, the seams left where a synthetic face is composited onto a real head. The second is physiological: a real face recorded on video carries faint periodic colour changes from blood flow, and blink patterns that generators reproduce imperfectly. The third is file-level, examining metadata and encoding history for traces of the tools used.\n\nA separate and increasingly important step is attribution rather than detection: working out which generative program produced a file, and which source images it was built from, so that a distribution chain can be reconstructed.\n\nTHE KOREAN SYSTEM IS THE CLEAREST DOCUMENTED CASE\n\nThe Korean National Police Agency's National Office of Investigation put a detection tool into use in March 2024. Korean reporting of the announcement gives a training corpus of 5.2 million data points from 5,400 people, including one million Korean samples and 130,000 from other Asian populations, and an analysis time of five to ten minutes per video depending on quality and length.\n\nThe deliberate weighting of the corpus towards Korean faces is worth noting. Detection accuracy, like face recognition accuracy, varies by the population a model was trained on, and building a national tool on a national corpus is a reasonable response to that.\n\nWHAT IT HAS BEEN USED FOR\n\nBy August 2026 the tool had been applied in 1,636 investigations, including 72 election crime cases during 2026. Korean police reported 419 arrests for deepfake offences between January and July 2026, with 17 people detained. Those arrests are the product of whole investigations and not a measure of the software's performance.\n\nThe agency budgeted 27 billion won for 2025 within a programme of 91 billion won running to 2027 to develop detection of both synthetic video and synthetic voice.\n\nACCURACY IS THE WHOLE PROBLEM\n\nThe reported detection rate is about 80 per cent. Officials said explicitly at launch that because it is not complete, results would guide the direction of an investigation rather than serve as evidence. That is the right framing and an unusually candid one.\n\nIt also sets the limit of the technology. At 80 per cent, roughly one judgement in five is wrong, and the two failure modes are not equivalent. A false negative lets a fabricated video pass as real. A false positive tells investigators that a genuine recording, which may be evidence of an actual offence against an actual person, is a fabrication. In the sexual-image cases that dominate the Korean caseload, both errors fall on complainants.\n\nDetection is also adversarial in a way that most forensic tools are not. Generators improve continuously, and a classifier trained on last year's outputs degrades against this year's. A published accuracy figure is a measurement of one moment, not a property of the tool.\n\nTHE CASE FOR IT\n\nThe alternative to a classifier is not certainty; it is a specialist examining every file by eye, which does not scale to the volume of material now in circulation. Used as triage, a tool that sorts a large seizure into probably-real and probably-synthetic lets scarce forensic expertise go where it is most needed, and the Korean model routes ambiguous results to human examiners rather than accepting the software's answer. Where synthetic sexual imagery is generated at volume by cheap consumer tools, no purely manual response is realistic.\n\nWHO OVERSEES IT\n\nNo published oversight framework for police deepfake detection was identified. There is no known independent evaluation of any deployed system's accuracy, no published standard for when a detection result may be put before a court, and no established disclosure obligation requiring a defendant to be told that a classifier was used to direct the investigation against them.\n\nWHAT IS NOT KNOWN\n\nWhether the reported 80 per cent holds against current generators is unknown, as is whether accuracy varies across skin tone, age or video quality. Whether the Korean tool was built wholly in-house is unresolved: police describe it as domestically developed and name no supplier, while a commercial vendor separately announced a detection solution developed in collaboration with the agency, and no source read establishes whether these are the same system. Deployments outside South Korea are not well documented.",
      "typical_vendor": "",
      "related_subject": "deepfakes",
      "url": "https://www.policeainews.com/technologies/deepfake-detection-forensics"
    },
    {
      "slug": "oyoon-dubai-surveillance",
      "name": "Oyoon (Dubai city-wide AI surveillance network)",
      "category": "Surveillance",
      "summary": "Dubai Police's city-wide network of AI-analysed cameras, using facial recognition and plate reading to flag wanted people and vehicles, operating since 2018 under no dedicated AI legislation.",
      "description": "WHAT THE SYSTEM ACTUALLY DOES\n\nOyoon — Arabic for 'eyes' — is Dubai Police's programme for putting artificial intelligence behind cameras the emirate already had, rather than a single new camera network.\n\nFeeds from cameras at tourist sites, on the Metro and wider public transport, on the road network and at public buildings are pooled into a Dubai Police command centre. Software converts passing faces to templates and compares them against police records, reads vehicle registration plates, and analyses behaviour, raising an alert for a human operator when it believes it has found a wanted person or vehicle.\n\nDubai Police operate it with partner bodies. Gulf News reporting names the State Security Agency, Dubai Municipality and the Roads and Transport Authority among them, which is what allows cameras owned by transport and municipal authorities to feed a police system.\n\nTHE NUMBERS, AND WHERE EACH COMES FROM\n\nThe programme was launched in 2018. In March 2019 Gulf News quoted First Lieutenant Ali Al Shehi of the Dubai Police Criminal Investigation Department saying the cameras 'helped arrest 319 suspects in different areas in Dubai through their automatic facial recognition technology' over the preceding year. The same report gave 5,000 cameras in the Metro system alone.\n\nIn July 2021 Gulf News reported that Sheikh Mohammed bin Rashid Al Maktoum had been briefed on the programme alongside Lieutenant General Abdullah Khalifa Al Merri, Commander-in-Chief of Dubai Police, and quoted the ruler saying Dubai is monitored by more than 300,000 cameras. The same report attributed a fall in emergency response times, from 6.46 minutes to under three minutes, to the smart system.\n\nThe Oxford Institute of Technology and Justice, in a country study current to March 2026, describes Oyoon as using facial recognition and behaviour analysis to identify suspects and track vehicles and licence plates across more than 5,000 cameras covering tourist sites, public transportation and traffic, and records the 2018 arrest figure.\n\nThese figures are not one series and should not be read as one. 5,000 is a Metro estate, 300,000 is a whole-emirate figure given by the ruler, and the two are five years apart. The 319 arrests are a single year, 2018, and no later arrest figure has been published.\n\nTHE LEGAL POSITION IS THAT THERE ISN'T ONE\n\nThe Oxford study states plainly that there is no dedicated legislation in the UAE targeting the use of AI in criminal proceedings. What exists instead is general policy: the Charter for the Development and Use of AI (2024), the AI Ethics Principles and Guidelines (2022), and the AI Adoption Guideline in Government Services (2023). None of these is a statutory authority for police biometric identification, and none creates a route of challenge for a person wrongly matched.\n\nNo oversight body specific to the system has been identified, no published watchlist criteria exist, no retention period for face templates or plate reads has been made public, and no accuracy or false-match rate has ever been released. That is a longer list of unknowns than this catalogue records for most comparable networks.\n\nTHE CASE FOR IT\n\nDubai's authorities make an argument that deserves stating rather than dismissing. The emirate has an unusually large transient population — tourists, transit passengers and short-stay workers — and a correspondingly thin base of local knowledge for conventional investigation. A centralised camera network with automated matching is a way of substituting technical reach for that. The published response-time improvement, from 6.46 minutes to under three, is a real operational claim about getting officers to incidents faster, and Dubai's low recorded crime rate is used domestically as evidence the approach works.\n\nIt is also a system its operators talk about openly. The ruler's own briefing was publicised; the arrest figures came from a named officer. Compared with jurisdictions that deny operating facial recognition at all, Dubai's willingness to describe the programme is a form of transparency, if a limited one.\n\nTHE CASE AGAINST IT\n\nThe transparency stops at the point where it would allow scrutiny. Publicity about capability is not the same as publication of the rules governing it, and none of the documents that would let anyone test the claims — watchlist policy, retention schedule, accuracy testing, an audit of the 319 arrests — has been released.\n\nThe absence of dedicated legislation means that a face scanned in a Dubai shopping centre is processed under general government AI guidance rather than any law setting conditions or remedies. In jurisdictions that have legislated, the argument has turned on exactly those conditions. Here there is nothing to argue about.\n\nThere is also the question of what a network of this scale does to expression and assembly, which in the UAE is already tightly constrained. A system built to identify wanted persons will identify anyone whose face is on a list, and the composition of that list is unpublished.\n\nWHAT IS NOT KNOWN\n\nThe supplier is not established; no source consulted names one. The current camera count is not established — the 300,000 figure is from 2021 and refers to the emirate rather than to Oyoon specifically. Whether facial recognition runs live across the whole network or only on selected feeds is not established. Whether any person has been wrongly identified is not established, because no mechanism exists that would make such a case public.\n\nEVIDENCE BASE\n\nNo Dubai Police publication was read directly for this record; dubaipolice.gov.ae could not be reached. The record rests on official statements relayed by reputable regional journalism, and on one academic country study. That is a weaker footing than this catalogue prefers, and the gap is recorded here rather than papered over.",
      "typical_vendor": "",
      "related_subject": "surveillance",
      "url": "https://www.policeainews.com/technologies/oyoon-dubai-surveillance"
    },
    {
      "slug": "precobs",
      "name": "Precobs (Pre Crime Observation System)",
      "category": "Predictive policing",
      "summary": "A place-based burglary forecasting system used by Swiss cantonal police since 2013, still running in Aargau but abandoned by other forces after evaluations found no measurable effect on burglary rates.",
      "description": "Precobs takes recorded burglary reports and looks for clusters in them. Where a burglary shows the characteristics its operators associate with travelling, professional offenders, the system flags a surrounding area as one where further offences are considered likely over a short following period. Patrols are then directed into that area. It does not name individuals, does not score people, and does not predict who will offend. It predicts where, on the basis of what has already been reported.\n\nThe underlying assumption is near-repeat victimisation: the observation, well documented in criminology, that a burglary in a street raises the short-term probability of another burglary nearby. Precobs is one of several commercial products built on that idea. Its Swiss vendor is LogObject AG, according to AlgorithmWatch CH's atlas of algorithms used in Swiss public administration.\n\nWHERE IT IS USED\n\nAlgorithmWatch CH records Precobs in use in Switzerland since 2013, across three cantons: Zurich, Aargau and Basel-Landschaft. Basel-Landschaft discontinued it in 2020, on the stated ground that it no longer met the increased requirements of crime analysis. The canton of Solothurn never adopted it and uses a different product, Picar.\n\nThe Kantonspolizei Aargau is the force most closely associated with it. SRF, the Swiss public broadcaster, reported on 27 April 2020 that the force had used the software for five years. Netzwoche reported on 1 July 2026 that Aargau still uses Precobs, alongside Picar, and that these systems process case data rather than personal data.\n\nThe picture in Zurich is less clear and is not settled here. Swiss reporting distinguishes the cantonal police from the city police, and coverage read while researching this record points in different directions about which body uses what. That is recorded as an open question rather than resolved.\n\nTHE CASE FOR IT\n\nThe argument its users make is modest and worth stating accurately, because it is more defensible than the marketing around predictive policing generally. Aargau's media spokesman Roland Pfister told SRF: \"Die Software fasst keine Einbrecher\" - the software does not catch burglars. His claim was that it helps officers be in the right place, and that its purpose is prevention rather than arrest.\n\nThat is a claim about patrol allocation, not about foreseeing crime. A force with finite patrol hours has to decide where to send them, and it will decide somehow - by intuition, by yesterday's incident list, or by a clustering algorithm run over the same incident list. Framed that way, the question is not whether the software is clairvoyant but whether it allocates better than the alternative.\n\nIt is also, by the standards of this field, unusually contained. It works on offence records rather than on personal data, it produces areas rather than names, and nothing read suggests it feeds individual risk scores or watchlists. Many of the objections that attach to person-based predictive policing do not reach it.\n\nTHE EVIDENCE IS CONTESTED\n\nThe difficulty is that prevention is close to unmeasurable. SRF put this plainly: the Kantonspolizei cannot prove the software's effect, because crimes that did not happen cannot be counted. The force pointed to burglaries falling by around a third since introduction. SRF noted in the same report that neighbouring Solothurn, which does not use Precobs, saw a comparable decline.\n\nAlgorithmWatch CH goes further, stating that the three cantons using Precobs are nowhere near the best performers and that researchers found no correlation between the system and falling burglary numbers, citing a 2019 University of Hamburg report. It also notes that Swiss burglary rates fell until 2010 - before Precobs was introduced anywhere - and rose again in 2022 and 2023.\n\nThe honest summary is that burglary rates moved in ways that do not track adoption of the software in either direction, and that no evaluation read here establishes an effect.\n\nWHY SOME FORCES DROPPED IT\n\nBasel-Landschaft's stated reason, per AlgorithmWatch CH, was that the tool no longer met what the force needed from crime analysis - a capability judgement rather than a finding that it was harmful or unlawful. Solothurn declined to adopt it. Aargau has kept it while acknowledging it cannot demonstrate its value. Those three responses to broadly the same evidence are the most interesting thing about the record.\n\nTHE LEGAL POSITION\n\nNo statute, regulation or court ruling specific to Precobs was identified for this record. Swiss cantonal police law governs it in general terms, and the canton of Zurich has been debating amendments to its police law since early 2026 covering the use of intelligent analysis systems, which suggests the legal framework is regarded as incomplete rather than settled. No data protection authority finding on Precobs was found.\n\nWHAT IS NOT KNOWN\n\nNo cantonal police publication describing Precobs was located for this record; the account above rests on the public broadcaster, a research NGO's atlas and a specialist trade publication. Current usage figures, the number of alerts generated, and any internal evaluation are all unknown. The date Aargau began using it is uncertain: SRF's 2020 report implies about 2015, while AlgorithmWatch dates Precobs in Switzerland to 2013. Whether the Zurich cantonal police still use it is not established. Neither is the position in Germany, where the product has also been deployed and, on reporting not verified for this record, withdrawn in at least one state.",
      "typical_vendor": "LogObject AG",
      "related_subject": "predictive-policing",
      "url": "https://www.policeainews.com/technologies/precobs"
    },
    {
      "slug": "ai-social-media-monitoring",
      "name": "AI Social Media Monitoring",
      "category": "Surveillance",
      "summary": "Software that collects public social media posts at scale and uses AI to search, cluster and score them for investigators and intelligence analysts.",
      "description": "AI social media monitoring tools gather large volumes of public posts and let analysts search them by keyword, place, name, phone number or account, with machine-learning features such as sentiment scoring and automatic alerts.\n\nThe questions they raise are about scope rather than accuracy: what counts as public, how long material is kept, whether sentiment or topic scoring drifts into monitoring people for their views, and whether local surveillance rules require the purchase to be disclosed.\n\nRecords on this tracker are listed one body at a time. Vendor claims about the number of posts analysed are recorded as claims.",
      "typical_vendor": "LifeRaft; Zignal Labs; others",
      "related_subject": "",
      "url": "https://www.policeainews.com/technologies/ai-social-media-monitoring"
    },
    {
      "slug": "image-geolocation",
      "name": "AI Image Geolocation",
      "category": "Investigative analytics",
      "summary": "AI that estimates where a photograph was taken, down to a specific property, by matching visual clues against mapped and listed images.",
      "description": "Image geolocation tools take a single photograph, including one taken indoors, and match its visual details against large collections of street imagery and property listings to suggest where it was taken.\n\nPolice use is investigative: locating a victim, a suspect or a scene from an image posted online. The risks are misidentification of a property, and the ease with which the same capability can locate anyone who posts a photo of their home.\n\nGraylark Technologies markets Raven to police. Reporting in October 2026 named the Los Angeles Police Department and the Miami-Dade Sheriff's Office as having bought access.",
      "typical_vendor": "Graylark Technologies (Raven, formerly GeoSpy)",
      "related_subject": "",
      "url": "https://www.policeainews.com/technologies/image-geolocation"
    },
    {
      "slug": "ai-call-taker-training",
      "name": "AI Call-Taker Hiring and Training Simulation",
      "category": "Control rooms",
      "summary": "AI that simulates emergency calls to test job applicants and train new 911 and 999 call-takers, and reviews past calls for quality assurance.",
      "description": "Control rooms short of staff are using AI to rehearse calls: a synthetic caller plays out an emergency and the system scores how the candidate or trainee handles it. The same tools are increasingly offered for quality assurance, reviewing recorded calls against protocol.\n\nThis is distinct from AI that answers or triages live calls, which the tracker records under AI Emergency and Non-Emergency Call Triage.",
      "typical_vendor": "GovWorx; Aurelian; others",
      "related_subject": "",
      "url": "https://www.policeainews.com/technologies/ai-call-taker-training"
    },
    {
      "slug": "ai-assisted-redaction",
      "name": "AI-Assisted Redaction",
      "category": "Investigative analytics",
      "summary": "Software that uses machine learning to find and redact or anonymise personal information in police video, audio and documents.",
      "description": "AI-assisted redaction tools detect faces, number plates, names and other personal information in video, images, audio or documents and mark them for removal before material is disclosed to courts, the public or other agencies.\n\nIn English policing, Devon and Cornwall Police and Dorset Police are piloting a tool called Secure Redact, and Avon and Somerset Police used Pimloc SecureRedact until 31 March 2026, according to force Freedom of Information responses. A Microsoft customer case study of September 2025 (a company claim) describes a document redaction application built on Azure AI for Cleveland Police.\n\nThe tracker records such tools where a force or other public body is shown to be using them. The main questions are accuracy of detection, human checking of the output, and what happens to the material processed.",
      "typical_vendor": "",
      "related_subject": "",
      "url": "https://www.policeainews.com/technologies/ai-assisted-redaction"
    }
  ]
}