← Back to feed

A six-month trial of live facial recognition cameras at London train stations scanned approximately 500,000 faces, produced one false positive, and resulted in zero arrests, according to documents obtained by Liberty Investigates through a freedom of information request and shared with The Guardian.

The trial cost £320,000 and consumed nearly 100 police hours. The single alert the system generated was incorrect.

What the Numbers Say

The headline ratio — half a million scans, zero legitimate alerts — creates a concrete performance baseline for a class of AI deployment that has mostly been evaluated in controlled conditions.

The false positive rate of 1-in-500,000 sounds low in isolation. But that single incorrect match still consumed police time and, more critically, the system failed to flag any genuine targets across the entire trial. A positive predictive value of zero — meaning every alert was wrong — is distinct from a low false positive rate. The technology produced no useful signal across six months of operation at one of the world’s most surveilled transit networks.

The cost breakdown is relevant context. At £320,000 for six months and roughly 100 police hours, the trial generated no arrests to set against that expenditure. Comparable manual spot-checking programmes typically generate actionable intelligence at lower cost per outcome.

The Deployment Gap

Lab evaluations of facial recognition frequently report high precision and recall on standard benchmark faces. Deployed systems at transit hubs operate under different conditions: varying lighting, crowd density, occlusions, hat-wearing passengers, and a base rate of genuine suspects that is very low. A system optimised for benchmark accuracy does not automatically transfer to operational accuracy in a high-throughput, low-target-density environment.

The UK trial result illustrates that gap with actual numbers rather than theoretical projections.

Policy Implications

The UK government has been expanding live facial recognition pilots since 2024. Critics have argued the technology encodes demographic bias and creates a de facto surveillance infrastructure without adequate legislative oversight. Supporters cite serious crime prevention as the operational rationale.

The trial data gives both sides harder ground to argue from. For critics: the system produced no arrests and one false positive in six months at significant cost. For supporters: a 1-in-500,000 false positive rate is technically low, and absence of arrests may reflect the absence of flagged suspects in that location during that period rather than a failure of the technology itself.

Parliament has not passed primary legislation governing live facial recognition deployment in public spaces. The Home Office currently treats it as a policing power that does not require specific statutory authority. The FOI data adds pressure to that position at a moment when the AI governance debate is becoming a legislative priority.