London Facial Recognition Trial: £320k Spent, Zero Arrests
British Transport Police spent £320,786 on a six-month facial recognition trial at London stations. The result: 500,000 faces scanned, one false alert, zero arrests.
👁 118 reads
British Transport Police concluded a six-month trial of real-time facial recognition at London’s busiest stations in July. A document obtained through freedom of information by Liberty Investigates and shared with The Guardian reveals the outcome: £320,786 spent on equipment rental and police staffing, 18 camera deployments, nearly 100 hours of officer time, over half a million faces scanned, and a single match alert. The alert was false. Arrests: zero.
This isn’t a story about technology failure. The system worked exactly as designed - it compared faces of passersby against a watchlist of wanted individuals and those subject to court orders. The problem lies elsewhere and is arithmetic. That’s why we sat down with a calculator.
Cost per face scanned and per alert
Let’s start with the conversion. The pound sterling exchange rate on October 1, 2026 is approximately $1.30 USD, which gives $417,022 USD spent from British taxpayer money on six months of piloting. This amount generated one result - a false positive match. The unit cost of one alert, regardless of its accuracy, is therefore $417,022. The cost of one actual arrest is impossible to calculate, because dividing by zero has no mathematical meaning - and that’s the most powerful number in this story.
The second calculation is more interesting because it shows the scale of data processing. £320,786 divided by 500,000 scanned faces gives 0.64 pence per face, equivalent to $0.008 USD. For that amount, the state processed biometric data of a person who, in the vast majority of cases, didn’t even know their face had been converted into a mathematical vector and compared against a database of suspects. Half a million is stated as “over,” so the real price per scan is even lower - and this is precisely the argument operators of such systems use in defense. Cheap, so why not.
The third calculation concerns people, not machines. Nearly 100 hours of officer work spread across 18 deployments gives an average of 5.5 police hours per camera deployment. At £320,786 total cost, that works out to £3,208 per hour of police presence at these cameras, or $4,170 per hour. Of course, this figure mainly includes equipment rental, not salaries - but that’s precisely what shows the expenditure structure. They paid for technology, not people, and the technology delivered nothing.
Former biometric surveillance commissioner: “pilot does not seem fruitful”
Fraser Sampson, until recently the UK’s Biometrics and Surveillance Camera Commissioner and now a board member of Facewatch, a company implementing facial recognition in shops, assessed the result without mincing words. He pointed out something fundamental - success of the system in retail and success in police enforcement are two different definitions.
We know the technology works, but using it in supermarkets to deter shoplifters and prevent attacks on staff is completely different from deploying it to catch people on the public transport network. Success in a shop means such people don’t enter at all. Success for police trying to catch someone means people get caught. In that regard, the pilot does not seem particularly fruitful.
Sampson told The Guardian something that’s key to legal assessment of such deployments: the system processes special category data, and the obligation to demonstrate it’s doing so in an appropriate, proportionate and necessary manner rests with the police. Not with the citizen, who must prove they’ve been harmed.
Despite these results, British Transport Police extended the pilot for another four months in August and expanded it to underground stations. Since the extension, three confirmed matches have been recorded - all three involved people who, upon verification, turned out to be complying with their court orders. So again: zero arrests, zero violations detected, three people legally moving around the city stopped for checks by an algorithm.
EU regulations and the AI Act ban
The legal difference between verification scenarios is enormous and stems from Article 9 of GDPR, which classifies biometric data used for unique identification as special category data - alongside information about health, political views, or sexual orientation. Processing it is prohibited unless one of the exceptions listed in the regulation applies. For police activity, Directive 2016/680 applies, requiring a statutory basis, not an operational decision by a commander.
Then came the AI Act. The AI Act in Article 5 explicitly prohibits real-time remote biometric identification in public spaces for law enforcement purposes, allowing exceptions only when searching for kidnapping victims, preventing a specific terrorist attack, or locating perpetrators of enumerated, most serious crimes - and only after obtaining court approval or authorization from an independent administrative body. The prohibitions in Article 5 have been in force since February 2, 2025. The London pilot takes place outside the EU, as the UK is not covered by it - an identical deployment in the EU would require demonstrating it falls within one of those narrow exceptions.
| Parameter | London (Feb-July 2026) | EU - legal status |
| Real-time crowd scanning | Permitted by police decision | Prohibited, three narrow exceptions + court approval |
| Pilot cost | £320,786 = $417,022 USD | No equivalent for comparison |
| Number of alerts | 1 (false) | - |
| Arrests | 0 | - |
| Cost per alert | $417,022 USD | - |
Why zero hits doesn’t mean the system is broken - and why that’s worse news
Sampson calculated the variables that must align for such a system to find anything: location, deployment hours, composition of the watchlist, and the probability that someone from that list actually walks past the camera. All at once. With 18 deployments over six months - on average once every ten days - the probability of encountering a specific wanted person in a multi-million city is minimal. The technology recognized half a million faces, of which it was wrong once. From an engineering standpoint, that’s a decent result. From a public spending standpoint, it’s a purchase that had no chance of paying off because it was placed where there was nothing to find.
The worse news is this: the ineffectiveness argument is a weak argument because it’s reversible. You just need to expand the watchlist, increase the number of deployments, and extend the hours for matches to appear - and that’s exactly what British Transport Police is doing by entering underground stations. The first three matches after the extension involved people complying with the law, which means recognition effectiveness is growing faster than enforcement effectiveness. The cost calculation will then cease to be an argument, leaving only the question of proportionality - the same one Sampson posed to the police at the end. And numbers won’t answer that.
Facts verified October 1, 2026. Sources: The Guardian - Trial of live facial recognition in London stations leads to a false positive and no arrests, document obtained through FOI by Liberty Investigates, Regulation (EU) 2024/1689 - AI Act, Art. 5, GDPR, Art. 9 and Art. 15. Calculations based on exchange rate of $1.30 USD per pound sterling.
More on this from QUANTSLANT
news·Fable 5.1 vs Mythos 5.1: One Model, Two Rule SetsAnthropic's Fable 5.1 launch isn't about performance - it's about access control. The same model now ships in two versions with different restrictions.