When a city's cameras start deciding what matters, what do they actually achieve?

The Dutch call them focusflitsers: cameras built to spot a driver holding a phone. About forty are in use, moving every two months or so between 300 designated spots. In 2025 they found 30% of the phone offences recorded, and for every photograph an employee of the CJIB, the national fines-collection agency, decides whether there really is an offence. The chart shows how the number of phone fines moved.

Brussels runs a bigger network. About 363 number-plate cameras enforce its low-emission zone, and in 2023 they identified an average of 358,315 different vehicles a day. By mid-2024, 99.3% of the vehicles the rules target were compliant, exempt or carrying a day pass. The region’s evaluation also credits the changing fleet with cutting road traffic’s nitrogen oxide emissions by 36% between 2018 and 2023. That figure is modelled: the fleet the cameras see is “la seule variable”, while emission factors and total distance driven are held constant. The cameras measure what drives in, not the air.

From our library

Evaluation de la zone de basses émissions — Rapport 2023 (Bruxelles Environnement)

Analyzes 363 ANPR cameras monitoring 358k daily vehicles, achieving 99.3% compliance and emission reductions.

Read more →

Warsaw shows how much the rule itself decides. Its clean-transport zone has covered 37 km² since July 2024. By August 2025 the municipal guard’s four mobile devices had registered more than 222,000 vehicles that year, and 0.59% of them did not meet the standard. Yet the guard had found no offence to fine: an older car may enter four times a year, and residents, Warsaw firms and the over-70s are exempt until 2027. Campaigners want the city’s own CCTV to read plates too; city hall says there is no legal basis to fine drivers automatically from it.

Kraków’s zone opened on 1 January 2026, with a paid way in for older cars during a transition. Its number-plate cameras were not yet enforcing it, so the chart shows six weeks of hand checks by the municipal guard.

“Od wprowadzenia SCT w Krakowie minęło dopiero 1,5 miesiąca, dlatego za wcześnie na ocenę skutków jej funkcjonowania”, the city’s spokesman Piotr Subik told GLOBEnergia in February: six weeks in, it was too soon to judge. By August, according to SmogLab, the number-plate cameras had been accepted from the contractor and switched on, and the city was working out how to pass their data to the inspectors.

Counting crowds asks least of the people filmed. Amsterdam’s Public Eye counted heads on camera images from 2019 to 2021, needing about 70% accuracy and delivering about 90% on its training images; it is now out of use. The register entry for a newer camera system, in use since 2022, says its counts can switch crowd signs automatically, but traffic measures need a visual check on site first. It also concedes that at very low counts, combined with other data, a passer-by’s route could in theory be followed.

These cameras deliver when the rule is narrow, a person confirms each case and a law lets the authority act on the picture. France’s noise radars show the third condition at work: the legal frame has been in place since 2023, but no device has yet been certified to fine. In July 2026 the Métropole du Grand Paris’s sensors in 27 communes only displayed drivers’ noise (“ne dressent aucun procès-verbal”) while the radars meant to fine were still being tested.

France’s 2023 Olympics law allowed an experiment: software on CCTV that flags predetermined events such as an intrusion, a dense crowd, an abandoned bag or a fire. The police prefecture, the transport operators RATP and SNCF and Cannes used it on about 800 cameras. The state committed €885,626 for it. Wintics was in practice the only supplier the prefecture, RATP and SNCF tested; Cannes used Videtics. An evaluation committee set up under the law reported in January 2025.

Some functions worked. Intrusion and crowd-density alerts were broadly satisfactory. Fire detection mistook car lights, flashing beacons, illuminated signs and the reflections of sunrise and sunset for fires, and in Cannes the software took the pavement for a person lying on the ground. The chart shows RATP’s alerts during the Games; look at the third bar of each group.

More than 78% of RATP’s abandoned-object alerts were false. The committee counted one intervention due to the software at the police prefecture and four at SNCF during the Games. RATP, over the whole experiment, passed 62 cases to its security post, a few of which led to action; RATP itself counts all 62 as interventions based on an alert.

The committee blames the small numbers partly on the heavy police presence, and calls the system’s value “tout à la fois limité et réel”: limited by uneven performance, real because alerts let operators focus on likely risks and place staff better.

In Germany, Mannheim’s police and the Fraunhofer IOSB institute have run a version since 2018. On 46 of the city’s 70 public-space cameras, the software reduces passers-by to stick figures and sounds an alarm at blows, kicks or shoves. The city paid €860,000 for the cameras and the police €190,000; Fraunhofer owns the code. With little real footage available, officers staged fights to train it.

From our library

Verhaltensscanner in Mannheim: Hier wird die Überwachung getestet, die so viele Städte wollen (netzpolitik.org, 18 Oct 2025)

Features 46 AI‑CCTV cameras, €1.05 m spend, staged training fights, no independent evaluation, influencing Hamburg.

Read more →

Seven years on, the police could not or would not say how often the software had raised an alarm or what it had found. Its success, they wrote, lies in the system’s continuous development and “kann zum derzeitigen Projektstand nicht mit Kennzahlen dargestellt werden”: it cannot yet be shown in figures. Recorded street crime in the watched areas is back at its level of the year before video surveillance began, though recorded crime rises wherever police look harder, so the figure says little either way. No independent evaluation is planned.

The idea spreads all the same. Hamburg switched it on in September 2025 and Hesse’s police law allows it. Berlin’s police law of December 2025 does too, and the Greens and the Left have asked the state’s constitutional court to review it. On 10 July 2026 the Bundestag passed a federal police law that would let stations use video surveillance that identifies people and assesses their behaviour automatically; it still needs the consent of the Bundesrat, which rejected an earlier version in 2021.

Italy shows the other end. Trento put AI on its existing CCTV, with street microphones, for two EU-funded research projects. In January 2024 the data-protection authority fined the city €50,000: research is not a municipal function, the anonymisation was inadequate and the city could not show it had assessed the impact first. The cameras with the thinnest evidence are the ones parliaments keep extending.

On 23 February 2024, near London Bridge, a facial-recognition camera matched Shaun Thompson, a 39-year-old Black community worker, to a photograph of his brother on a police watchlist. Officers asked him to prove who he was, and when he declined to give his fingerprints he was threatened with arrest, the High Court’s summary records.

From our library

R (Thompson and Carlo) v Commissioner of Police of the Metropolis [2026] EWHC 915 (Admin) — press summary, 21 April 2026

London's live facial recognition gains broader legal clearance, extending surveillance despite prior privacy concerns.

Read more →

The Met publishes an annual report. Over the year to September 2025 its 203 deployments scanned 3,147,436 faces. The Met puts its false-alert rate at 0.0003% of faces scanned; counted per alert it is 0.48%. The chart shows what the alerts became.

The Home Office confirms the 962 arrests and says over a quarter involved violence against women and girls. Of the ten false alerts, seven involved Black men and one a Black woman. The Met’s report calls the imbalance not statistically significant; Big Brother Watch called it “disturbing”.

From our library

Consultation on a new legal framework for law enforcement use of biometrics, facial recognition and similar technologies (Home Office, December 2025)

UK proposes new biometric law, citing 962 arrests, 10 forces, 25,000+ monthly searches, oversight body.

Read more →

In the six months of the Croydon pilot, the Met says, its fixed cameras led to 173 arrests in 24 operations, with more than 470,000 people walking past and one false alert. It also reports crime in the area down 10.5% on the same period a year earlier, its own before-and-after figure with no comparison area.

In the United Kingdom, judges have set the limits so far. In 2020, in the Bridges case, the Court of Appeal ruled that South Wales Police, after scanning perhaps 500,000 faces, had left its officers too free a hand. Nobody could tell who might go on a watchlist, the judges wrote, or by what criteria a camera’s location was chosen.

From our library

R (Bridges) v Chief Constable of South Wales Police [2020] EWCA Civ 1058 — Judgment

Groundbreaking appellate decision deeming 500k facial scans across 50 deployments unlawful for discretion failures.

Read more →

Six years later the High Court dismissed the challenge by Thompson and Carlo: the Met’s policy, read as a whole, set “clear, interlocking and cumulative constraints”. The Home Office consulted that winter on a statute and a new oversight body; we found no law passed since. Meanwhile the Met plans fixed cameras in the West End and Soho by December.

Elsewhere the answer differs. Italy’s parliament has extended a moratorium on facial recognition in public places to 31 December 2027, with police use subject to the data-protection authority’s favourable opinion. In Hungary, three civil-society authors write, 2025 amendments extended facial recognition to all petty offences, and trucks fitted with biometric cameras were parked around that year’s Budapest Pride march; the authorities say their system is not “real-time”. In none of these cases was the decision a city council’s. City governments can still say no: Amsterdam’s new coalition agreed in 2026 not to use facial recognition at all.

Since 2 February 2025, Article 5 of the EU AI Act has banned real-time remote biometric identification in publicly accessible spaces for law enforcement. Three exceptions remain: searching for victims or missing people, an imminent threat to life or of a terrorist attack, and locating suspects of serious crimes. Each use needs a national law and authorisation from a judge or independent authority, beforehand or within 24 hours in an emergency. Identification after the fact is not banned but counts as high-risk, and since the AI Omnibus those duties start on 2 December 2027.

Software that flags behaviour without identifying anyone falls outside that ban; national law sets its limits. In August 2026 France’s Conseil constitutionnel ruled on the Ripost law. It struck down provisions in eleven of the law’s articles but upheld those on cameras. They extend the experiment to 31 December 2030, and from events to buildings and places open to the public that are particularly exposed, permanently or for a time, to terrorism or serious attacks on people. For the permanently exposed places, the Council added, each authorisation may last no more than three months, renewable.

From our library

Décision n° 2026-915 DC du 14 août 2026 — Loi visant à offrir des réponses immédiates aux phénomènes troublant l'ordre public, la sécurité et la tranquillité de nos concitoyens (loi Ripost)

Extends surveillance authority to 2030, introduces three‑month data caps, and tests AI license‑plate analytics.

Read more →

The same law starts a three-year trial of an algorithm that combs number-plate data for vehicle movements suggesting organised crime or car theft. Its signal “ne peut fonder, par lui-même, aucune décision individuelle ni aucun acte de poursuite”: on its own it cannot ground a decision about anyone, or a prosecution.

Regulators are drawing the edges for cities. Poland’s data-protection office told a municipality in 2025 that it may not give the police a live feed from its CCTV: “brak jest jakiejkolwiek normy prawnej” for it, there is no legal provision at all. Police must ask case by case, in writing. Italy’s authority stopped Como’s facial-recognition cameras in 2020.

From our library

Gmina nie może przyznać policji stałego dostępu do swojego monitoringu (UODO, 30 Sept 2025)

Polish authority blocks municipal police access to live CCTV, demanding statutory backing for any surveillance sharing.

Read more →

Data also travels upwards. In 2025 Belgium connected 450 Brussels cameras, bought for the low-emission zone and speed checks, to ANPR@GPI, a national police number-plate platform, and offered municipalities €20 million to buy more. Two years earlier Flanders’ privacy supervisor had told councils to debate need and purpose before buying, because, it observed, purposes tend to be fixed after the network exists and a legal footing found later still, and cameras once up keep gaining new uses.

The same French decision can be read two ways. The Council itself wrote that algorithmic processing makes “une analyse systématique et automatisée” of the images, able to increase considerably how much can be drawn from them. That is why it demands particular guarantees. La Quadrature du Net, which campaigns against the technology, sees the law as one more move in a “stratégie des petits pas”: small steps that settle it into daily life without a real public debate.

The IMD Smart City Index asks 120 residents in each city whether they agree with a list of statements. Two concern cameras: whether CCTV has made residents feel safer, and whether they are comfortable with face recognition to lower crime. The chart sets the two answers side by side for the 24 European cities in the Hub’s scope. Each line joins two different questions, not two moments in time.

Kraków is the only exception; in Berlin the two answers run from 47.8% to 71.8%. One statement names a purpose and the other asks about a result, so this is a gap, not a contradiction.

France’s evaluation committee commissioned a survey of 1,005 people in December 2024. Of those asked, 81% favoured AI cameras that spot risky situations without identifying anyone, and 76% even cameras that recognise a specific person. Yet 62% thought it would make general surveillance normal. Asked what should come first for public safety, they put AI video well down the list.

Only one in five knew what the experiment actually was. In Britain, the Home Office’s own polling found two in three people support police use of facial recognition, fewer for live use against crime (59% said always or usually acceptable), while many worried about misuse and false identification.

Acceptance is high and conditional. That makes telling residents what a camera does part of the licence to run it, not a courtesy.

The evidence in this read turns into five questions an official can put to any proposal.

1. What rule does the camera check, and does a law let us act on what it sees? Warsaw’s guards fined no one in a year because the zone’s rules let almost every car in; the city also says it has no legal basis to fine automatically from its own CCTV.
2. Who confirms each alert before anyone acts? In the Netherlands a CJIB employee checks every phone photograph; Amsterdam checks on site before changing the traffic.
3. Is it in a public register, with its impact assessment? Amsterdam’s algorithm register entries say what each system does and where it can fail.
4. Is there an independent evaluation, with a comparison, published before the system grows? France’s law required an evaluation, though not of crime, and Parliament extended the trial after it; Mannheim has none planned.
5. Where does the data go next? Brussels’ clean-air cameras now answer national police searches.

The Smart Cities Hub is being built to map who runs which cameras under which rules, to connect cities facing the same decision and to gather evidence in the languages it is written in. It aims to support shared work, such as procurement clauses and impact assessments the next city can reuse.

If your city has a camera register entry, an evaluation or procurement terms to share, join the Hub’s community and describe them under your ideas for collaboration. The library holds the evaluations, rulings and regulators’ decisions this read draws on; the projects and organisations working on city data and AI are listed under Data Platforms, AI & Digital Twins.