What Did London Get From Half a Million Face Scans?
Key takeaways
- A British Transport Police trial at London railway stations recorded 500,000 face scans and zero direct arrests.
- The scan count measures activity and may include repeated scans of the same people.
- Zero direct arrests does not establish how accurately the system identified faces.
- A meaningful assessment must track confirmed matches, subsequent action, operating costs, and disruption to passengers.
Half a million face scans would make an impressive-looking usage chart. According to The Guardian’s September 29, 2026 report, a British Transport Police (BTP) trial at London railway stations recorded 500,000 face scans and zero direct arrests. For passengers walking past those cameras, the useful question is what all that activity actually achieved.
A scan count is an activity metric
Tech companies have a familiar problem with metrics: the easiest number to measure can become the number everyone celebrates. Police surveillance deserves a tougher scorecard.
The 500,000 figure tells us how much scanning took place. It does not necessarily represent 500,000 different people. Someone passing the cameras repeatedly could have contributed multiple scans.
Zero direct arrests describes an outcome. If the trial was intended to produce arrests, that result matters. A large volume of processing cannot, by itself, demonstrate effective policing.
It also leaves questions open. How many people on the watchlist actually passed the cameras? Did any information help an investigation without directly producing an arrest? The arrest count alone cannot answer either question.
The trial’s objective therefore matters. Demonstrating that equipment works in a busy station is one test. Demonstrating that it improves police results is another. An assessment needs to say which question it is answering.
An alert is the beginning of a check
Live facial recognition typically compares features extracted from camera images with faces on a watchlist. When the similarity exceeds a preset threshold, the system generates an alert.
That alert signals a possible match. It does not establish someone’s identity or, by itself, provide grounds for arrest. A person must check the match and determine what action is legally justified.
A false positive occurs when the system incorrectly flags someone as a possible match. A false negative occurs when it misses someone it should have detected. Neither rate can be inferred from the absence of direct arrests.
There is also a difference between a mistaken alert and a mistaken police intervention. An officer might dismiss an incorrect match immediately. If it leads to a stop, however, a passenger may have to interrupt their journey and explain who they are.
An arrest-only scorecard leaves that experience out. For the person trying to catch a train, the interruption still counts.
Small error rates can create plenty of work
Scale changes the practical meaning of an error rate.
Consider a hypothetical example: a system checks 100,000 faces belonging to people who are not on its watchlist. If it incorrectly flags 0.01% of those checks, it generates 10 false alerts.
Those are illustrative numbers, not measured results from the London trial.
Now consider how rarely an actual watchlist subject might appear. If genuine targets are scarce, false alerts can make up a substantial share of all alerts, even when the false-positive rate is low.
That makes two questions essential:
- How often does the system incorrectly flag someone who is not a target?
- Of all the alerts it generates, how many identify an actual target?
The first measures how often innocent passersby trigger an alert. The second tells officers how much confidence to place in an alert once it arrives.
Raising the matching threshold can reduce false alerts. It can also increase the chance of missing a real target. A single headline about “accuracy” conceals that tradeoff.
Show what the same budget buys
A useful evaluation would follow the entire sequence: scans, alerts, confirmed matches, and subsequent police action. That would show where the system contributed something useful and where it consumed time without producing a result.
Costs belong in that assessment too. Cameras and software are only part of the bill. Officers spend time reviewing alerts. Passengers may lose time to unnecessary stops.
The strongest comparison is with existing police work using the same staffing and budget. Did facial recognition help locate wanted people faster? Did it produce additional arrests? What extra result justified the resources and the intrusion?
Those are questions for the trial to answer, not benefits to assume.
Half a million scans and zero direct arrests put the burden on the people proposing these systems to explain their value. Before facial recognition becomes an ordinary feature of the commute, police should establish what measurable improvement earns it a place there.
Sources
Deepen your perspective
Comments
Loading comments...