Start with the numbers. Six months, 18 deployments, over 500,000 faces scanned, one alert — which was a false positive — and zero arrests. The bill: £320,786.
This is the half-year report card for the British Transport Police's (BTP) live facial recognition deployment at London railway stations, reported by The Guardian on September 29. The data came from a freedom-of-information request filed by the investigations unit of civil liberties group Liberty.
The Cost Breakdown
The trial ran from February to July this year, covering some of London's busiest railway stations across 18 deployments. Equipment rental plus officer costs totaled just over £320,000, with nearly 100 hours of police time invested. What the system does is simple: cameras capture faces in real time, compare them against a police watchlist, and trigger an alert on a match.
In six months, the alert fired exactly once. That person was not on the watchlist.
The Base Rate Math Doesn't Care About Feelings
Looking at this trial alone, the most striking thing isn't the false positive — it's the denominator. No matter how long the watchlist is, it's a tiny fraction of a flow of 500,000 people. Once detection becomes "finding a needle in a haystack," even a false positive rate of one in ten thousand would trigger dozens of alerts across 500,000 faces. System-level accuracy is never the same as demo-video accuracy — base rates pull everything back down to earth.
The extension period made things worse. BTP announced last month that the trial would be extended by four months, with deployments expanding into the London Underground. During the extension, the system confirmed three matches. The Guardian's wording deserves a pause: these three individuals were subsequently confirmed to be complying with sexual offence prevention orders or other court directives. The three people the system surfaced were all following the rules when checked. Measured against the goal of catching offenders in the act, the accuracy over more than half a year was zero.
Expanding Deployment, Late-Arriving Evidence
When the report card was published, the decision had already been made. The extension and expansion were decided the previous month. TfL's stated public rationale for supporting the expansion is crimes against women and girls, with station selection based on "maximum impact." In the same period, the Metropolitan Police and the Mayor's office announced fixed facial recognition cameras in London's West End before Christmas, also to be activated after the pedestrianization of Oxford Street.
The overall scale is even larger. According to Liberty Investigates, more than half of police forces in England and Wales have already deployed live facial recognition on the streets. The Joint Committee on Human Rights recently named this technology's rollout as a particularly clear example of risk, calling for legislation. Former Biometrics Commissioner Fraser Sampson made a pertinent point to the media: catching pickpockets in a supermarket and catching people in a public transport network are two different things — the technology's proportionality depends on location, time of day, watchlist composition, and probability of interception. Get one link wrong, and the whole thing fails.
Why This Matters
In computer vision papers, facial recognition benchmark scores have long been polished to the point of numbness. The value of this FOI report card is that it measures something else: what a "technically mature" system delivers when embedded in real crowds for six months. Half a million faces scanned, £320,000 spent, zero offenders caught in the act, one person misidentified. CV practitioners would do well to report "deployment accuracy" and "demo accuracy" as two separate metrics — public policy debate needs the former.
There's also an ordering problem worth remembering: the expansion decision was made before the report card was published. The data arrived afterward; scale moved first.
- Sources: The Guardian, September 29, 2026 (Phoebe Davies and Daniel Boffey; data obtained via FOI by Liberty Investigates); BTP extension statement and TfL support wording as cited in that report; Sampson and parliamentary statements from the same source
- Caveats: The 500,000 faces, £320,786 cost, and ~100 hours figures are per the FOI documents; this article could not verify BTP's original files. The claim that "all 3 matches during the extension were law-abiding" relies on The Guardian's wording (3 individuals confirmed complying with court directives); whether anyone later broke the law falls outside this dataset. This article does not evaluate the technology's raw recognition accuracy — only the deployment's system-level output.