Police Live Facial Recognition: Scaling Fast, Crime Effect Unproven, Law Still Catching Up

The use case

In the twelve months to 10 September 2025, the Metropolitan Police ran live facial recognition on 207 occasions: 203 at crime hotspots and four for protective security at events such as Notting Hill Carnival. An estimated 3,147,436 faces passed its cameras in that period (operator-reported)¹. Each face was compared with a watchlist of people the force wants to find; a typical crime-hotspot watchlist held about 10,250 people (operator-reported)¹.

How it works. Cameras on vans or at fixed points capture faces in a crowd. Software converts each face into a numerical template and scores its similarity to every watchlist image. When a score passes a set threshold, the system raises an alert, and an officer decides whether to speak to the person¹. The threshold sets the trade-off between missed matches and false alerts. In a National Physical Laboratory (NPL) test of one police-used algorithm against a 180,000-image watchlist, lowering the threshold raised the correct-alert rate from 83.1% to 93.8% and the false-alert rate on people not on the watchlist from 0.00% to 1.20% (measured in a technology evaluation with 401 volunteers, using 2022 deployment data)².

Scope note. Only the matching step is AI here: a machine-learning algorithm that scores facial similarity. Compiling the watchlist, choosing where to deploy, the officer's judgement at the roadside and the decision to arrest are human processes, and fixed CCTV and officers recognising faces from photographs are older methods that do similar work. Results below therefore mix algorithm performance with procedure, and the two are separated where the evidence allows. Retrospective searches of stored images are out of scope and appear only where an audit compares them with live use³.

Where it is running

Ten police forces in England and Wales had used live facial recognition by December 2025 (government-reported)⁴. A January 2026 Home Office white paper, as reported, would add 40 vans to the ten already in use⁵.

Metropolitan Police. Of the 962 arrests in the year to 10 September 2025, 549 were of people wanted on court warrants and 347 of people wanted by the Met. The Met also reports more than 1,400 arrests since the programme began (operator-reported)¹.

Essex Police. Essex paused live deployments over potential bias. The force's list shows the last deployment on 26 August 2025, and an Information Commissioner's Office (ICO) audit in November 2025 found deployments already paused⁷⁸. In March 2026 the force said it was confident it could resume⁷.

Outside England and Wales. The evidence found concerns law rather than performance. Buenos Aires deployed a facial recognition system for fugitives in 2019, using 9,500 cameras, and a city court declared it unconstitutional in 2022; the position after any appeal was not established⁹¹⁰. Hungary passed amendments on biometric surveillance in March 2025, in force from 15 April 2025, which civil-society organisations argue breach the EU AI Act's prohibition on real-time remote biometric identification (advocacy analysis)¹¹. No performance data from either country was found, and deployments in the United States were not examined, so the evidence base is largely British.

What the evidence shows

Accuracy of the match

Laboratory tests. At a threshold of 55 on its scale, the NPL measured a correct-alert rate of 89.3% and a false-alert rate of 0.17% against a 180,000-image watchlist; for a watchlist of 18,000 it anticipated a false-alert rate of about 1 in 5,700. Differences by gender, ethnicity and age in correct-alert rates were not statistically significant at that setting, but at a threshold of 50 and below the variation in false-alert rates by gender and ethnicity was statistically significant (measured in a technology evaluation, 401 volunteers, 2022 data)². The Met's summary of an earlier NPL study reports 89% correct alerts and about 1 in 6,000 false alerts for a 10,000-image watchlist at a threshold of 0.60 (measured in tests of operational deployments in 2021, as summarised by the Met)¹². Both datasets are more than four years old. The Met has used a threshold of 0.64 since July 2024¹; a University of Essex researcher says the NPL did not assess that setting and argues that the Met's no-bias conclusion rests on analysis of just seven false matches (critic's account)¹³.

Operational counts. Between 11 September 2024 and 10 September 2025 the Met recorded 2,077 alerts, of which 2,067 were true and 10 false (operator-reported). That is 0.0003% of faces seen, or 0.48% of alerts by simple division. Seven of the ten false alerts involved Black men, one a Black woman and two white men; the Met says the imbalance is not statistically significant, and the sample is small. All ten occurred at 0.64¹.

What the false-alert rate omits. The false-alert rate counts only wrong alerts, so it cannot show how many people on a watchlist pass the cameras without being flagged. A field test by University of Cambridge researchers measured this directly: at a threshold of 55, the system correctly identified 50.7% of the watchlist-assigned appearances of 188 paid actors, over about five hours at one Chelmsford location in February 2025 (measured in a field experiment on a single day)⁶. The NPL laboratory figure at the same nominal threshold was 89.3%². The two tests differ in design, and the Cambridge authors state that results are not directly transferable to other systems⁶. The gap shows that detection rates vary with test conditions; the causes were not established in the parts of the studies that were read. At threshold 55 the Cambridge study found 53.5% for men against 46.9% for women, and 60.3% for Black actors against 46.2% for White actors⁶.

Results for policing

Arrests and interventions. In Essex, 41 deployments between August 2024 and February 2025 scanned about 1.3 million faces and led to 123 interventions, 48 arrests and one confirmed mistaken intervention (operator data analysed by Cambridge researchers)⁶. In the Met year, 896 of the 962 arrests were of people already wanted on warrants or by the force (operator-reported)¹.

Crime reduction. The Cambridge study found "no statistically significant evidence that LFR deployments reduced crime in the short term" in Essex (academic analysis commissioned by the force)⁶. The mechanism is plausible: arresting people who are already wanted detects existing offences, whereas a fall in new crime would need deterrence or incapacitation, and none was detected over the study period. No comparison group was found for the Met figures.

What it takes: cost, integration, governance

Cost. Public figures cover programmes, not units: more than £26 million for a national facial recognition system and £11.6 million for live facial recognition capabilities, as reported from the January 2026 white paper (planned spending, press-reported)⁵. No cost per deployment, per alert or per arrest was found.

People and process. Officers adjudicate each alert, so staffing and training are part of the system. The Met reports that six of the ten false alerts led to an engagement with the person and none to an arrest (operator-reported)¹. In ICO audits of five forces published in August 2026, 107 recommendations were made; forces accepted 93 in full and 14 in part, and compliance was higher for live than for retrospective use (regulator audits, as reported)³. For Essex, the ICO rated live facial recognition assurance "reasonable", with six high-priority recommendations and none urgent, but no live deployment was assessed because use was paused⁸.

Legal basis. The Met operates under a policy adopted on 11 September 2024 that restricts use to three use cases¹⁵. On 21 April 2026 the Divisional Court of the High Court dismissed a judicial review of that policy, holding that it met the quality-of-law requirement under Articles 8, 10 and 11 of the European Convention on Human Rights¹⁴¹⁵. The claimants did not argue that the technology is unlawful in principle¹⁴, and have said they plan to appeal¹⁵. A Home Office consultation ran from 4 December 2025 to 12 February 2026 on a new legal framework for law enforcement use of biometrics and facial recognition; a response was due within 12 weeks of closing, and whether it has been published was not checked⁴.

EU rules. The recitals of the EU AI Act limit real-time remote biometric identification in public spaces for law enforcement to exhaustively listed and narrowly defined situations. They call for a fundamental rights impact assessment, express prior authorisation by a judicial or independent administrative authority, and notification of each use to the market surveillance and data protection authorities¹⁶. The operative Article 5 text was not read; this summary rests on the recitals.

Where it falls short

The challenge question is what would make this use case look less successful than it is usually presented. Six features do.

• One kind of error is counted. The Met's headline of ten false alerts among about 3.1 million faces is a count of wrong alerts only. It says nothing about people missed, and the one field test that measured misses found about half of watchlist appearances flagged at threshold 55¹⁶.
• Independent testing is older than the operations. The NPL datasets date from 2021 and 2022 and used volunteer cohorts, while operational figures for 2024 to 2025 are reported by the operator; a critic argues the Met's current setting was not tested¹²¹²¹³.
• Evidence on bias is thin in both directions. Seven of ten false alerts involved Black men, which ten events can neither confirm nor rule out as a pattern¹. Essex paused over potential bias and later said it was confident to resume⁷. The NPL found significant variation in false-alert rates at lower thresholds².
• A crime effect is not shown. The only controlled look found, in Essex, detected no significant short-term fall in crime⁶.
• Outcomes depend on people and lists. Arrests depend on who is on the watchlist and on officers' decisions at the roadside. The ICO audits made 107 recommendations, and Essex received six high-priority ones for live use³⁸.
• Legal and political challenge continues. One claimant, Shaun Thompson, was reported as misidentified near London Bridge in February 2024, and the High Court judgment is expected to be appealed¹⁵. In Buenos Aires, the court record described biometric data for 15,459 people loaded without judicial authorisation, 84 of them at Interpol's request, and a database with flaws that "could lead to erroneous arrests"; the system was suspended in April 2022¹⁰.

In England and Wales, no withdrawal was found beyond the Essex pause.

Phase: In early use — why

The hypothesis is kept. Confirmed machine-learning face matching is in routine use in ten police forces in one country, at a scale of millions of faces a year in the largest¹⁴. That is more than a pilot, but not wide use: operational evidence comes almost entirely from England and Wales, one of the ten forces paused deployments in 2025⁷, the legal framework is still being drafted⁴, and the one non-UK system with a court ruling found was declared unconstitutional⁹. The "Pilot with results" phase does not fit, because deployments are recurring operations rather than trials.

What to watch next

• The ICO audit of the Metropolitan Police, which was still to come "later this year" as of August 2026³.
• The Home Office response to the consultation that closed on 12 February 2026, due within 12 weeks of closing; publication was not checked⁴.
• Any appeal in R (Thompson and Carlo), which the claimants have said they intend to bring¹⁵.
• The rollout of the 40 additional vans announced in the January 2026 white paper⁵.
• Whether Essex Police resumes live deployments, as it said in March 2026 it was confident to do⁷.
• The Met's next annual report, covering the period after 10 September 2025; the previous report was signed on 30 September 2025¹.

Questions to ask before adopting

• Which match threshold and software version will be used, and what independent test covers exactly that setting?
• How many people on the watchlist would pass undetected in our conditions, not only how many false alerts will occur?
• Who decides who is on the watchlist, and how many of the arrests would have happened without the technology?
• What outcome will be measured, against what comparison, and who will evaluate it independently?
• What legal basis, impact assessment and oversight apply, and what happens to the data of people who are not matched?