£320k, 500,000 faces, no arrests: BTP's face scan trial

British Transport Police (BTP) scanned more than 500,000 faces at nine London railway stations during a six-month live facial recognition trial. It cost £320,786 and produced exactly one alert, which was wrong. No arrests came directly from the technology. The figures were obtained through Freedom of Information requests by Liberty Investigates, published on 29 September 2026 with The Guardian and reported by The Register on 30 September.

BTP has extended the trial anyway, onto the London Underground.

Live facial recognition technologies are a serious invasion of our privacy. The government should be suspending their rollout until there are stronger safeguards in place, including a proper legal framework.
— Bell Ribeiro-Addy MP, quoted by Liberty Investigates

Half a million faces, one wrong match

Live facial recognition (LFR) works by pointing cameras at a crowd, turning each face into a biometric template and comparing it in real time against a police watchlist. When the software thinks it has found a match, it alerts officers, who then decide whether to stop the person.

According to Liberty Investigates, BTP ran 18 deployments between February and July 2026. The £320,786 covered equipment hire and police staffing, and the trial took about 100 hours of officer time. Out of more than half a million scans came a single alert, and that was a false positive. The Register reports that the system used is NEC's NeoFace M40.

BTP points out that officers did make arrests during the deployments, for offences including assault, theft and carrying offensive weapons. But those arrests didn't come from facial recognition alerts, so they aren't counted as results of the trial. BTP says the aim was "to understand how the technology could most effectively support the identification of wanted offenders and individuals who may pose a risk to passengers and staff".

The rules still haven't caught up

There is still no specific law in the UK covering police use of live facial recognition. Forces rely on a mix of human rights law, data protection law and their own policies. The Home Office ran a consultation on a new legal framework from 4 December 2025 to 12 February 2026. It proposed a single regulator to "set standards, investigate misuse and enforce compliance". Mind the Hype couldn't find a published government response yet, and new legislation is expected to take a couple of years.

In the meantime, deployments keep growing. The Metropolitan Police, South Wales Police and other forces already use LFR regularly, and BTP's extension takes it into one of the busiest transport networks in the world.

Fraser Sampson, the former Biometrics and Surveillance Camera Commissioner, told Liberty Investigates: "Success for the police trying to catch people means people being caught. In that respect the trial doesn't appear to have been very fruitful."

Ada's take: a test should be allowed to fail

The point of a trial is to find out whether something works. On the evidence published so far, this one didn't, at least not by the measure that matters most: finding wanted people. A single false alert from half a million faces is a strong result on accuracy in one sense. The system wasn't wrongly flagging lots of innocent commuters. But it also wasn't finding anyone the police were looking for.

That should make extending it a harder decision, not an automatic one. Every scan involves processing someone's biometric data, and the people being scanned are overwhelmingly just trying to catch a train. If the benefit is close to zero, the case for proportionality, the legal test police have to meet, gets very thin. And all this is happening while the legal framework that's supposed to govern it is still being written.

To be fair to BTP

Results depend heavily on who's on the watchlist and where the cameras go. Even Sampson noted that "there are many variables and all need to align", and a quiet six months isn't proof the technology can't help. Low false alerts matter too: earlier criticism of LFR centred on misidentifications, and this deployment didn't produce a stream of them. BTP may argue that a bigger network gives a fairer test. If so, it should publish its success criteria in advance so everyone can judge the extension on the same terms.

Mind the hype. The cameras saw half a million faces and didn't correctly spot a single person on the watchlist.

Sources: Liberty Investigates · The Register · techUK (Home Office consultation) · State of Surveillance