Six months, more than half a million faces scanned, and a single alert on a police watchlist - which turned out to be wrong. That is the balance sheet from British Transport Police's live facial recognition (LFR) trial across some of London's busiest railway stations, run between February and July this year at a cost exceeding £320,000. No arrests resulted from the technology itself.
What the trial actually measured
The figures, obtained through a freedom of information request by Liberty Investigates and shared with the Guardian, cover 18 separate deployments and nearly 100 hours of officer time. Cameras compared live footage of passengers against a police watchlist of wanted offenders and people subject to court orders. Out of roughly 500,000 faces scanned, the system generated one match - and that match was a false positive, meaning an innocent person was wrongly flagged as a person of interest.
British Transport Police has pointed out that officers made other arrests during the deployment windows, for offences ranging from assault to possession of an offensive weapon. But those arrests arose from ordinary policing, not from LFR alerts, so they sit outside the technology's own performance data. Separating the two matters: it is the only honest way to judge whether the cameras themselves are doing the job they were bought to do.
Why the numbers matter for oversight
Facial recognition in public spaces processes biometric data, which UK data protection law treats as a special category requiring extra justification. Fraser Sampson, a former biometrics and surveillance camera commissioner, has argued that deployments must be shown to be proportionate - meaning location, timing, watchlist composition and the likelihood of a genuine match all need to line up. Measured against that bar, a single false alert across half a million scans raises obvious questions about whether the trial's design matched its stated aims.
Sampson's comparison with retail use of facial recognition is instructive. In a shop, success looks like deterrence - problem customers simply stop coming in, and that absence is itself evidence the system works. Policing a transport network is different: success requires the system to actually find people on a watchlist. Judged by that standard, the trial's results look thin.
A wider rollout despite the figures
None of this has slowed expansion. BTP extended the trial for a further four months and widened it to London Underground stations, where it reports three confirmed alerts linked to people complying with sexual harm prevention orders or other court conditions. Transport for London has framed the technology as a tool to tackle violence against women and girls on the network. Separately, the Metropolitan Police and the mayor's office have confirmed fixed LFR cameras for Oxford Street and the wider West End.
More than half of police forces in England and Wales have now used LFR on British streets, according to Liberty Investigates' analysis, making this far from an isolated experiment. Parliament's joint committee on human rights has flagged facial recognition rollout as a clear example of AI-related rights risk, and some MPs, including Bell Ribeiro-Addy, have called for a pause until a proper legal framework exists.
The underlying tension
- Watchlists determine everything: if a person isn't on one, the cameras do not store or act on their image in the way general CCTV might.
- Low match rates can reflect either a well-targeted watchlist or a system poorly suited to its environment - the data alone cannot distinguish the two.
- Cost and officer time are real trade-offs against conventional policing methods, and oversight bodies are expected to weigh that balance publicly.
BTP says it has refined deployment locations, operating procedures and watchlist construction as the pilot progressed, and that watchlist growth has been deliberate and governed. Whether that refinement produces materially better results - or simply more scans against an unchanged underlying problem - will only become clear as further data is published.