The Receipts Room
You cross the concourse at King’s Cross with a coffee in one hand and your phone in the other. You do not stop, and nobody stops you. But for a fraction of a second, a camera has measured your face and compared it, in real time, against a watchlist of several hundred wanted people. (That walk is an illustration. The cameras are not.)
Between 11 February and 17 July 2026, British Transport Police set up live facial recognition cameras on the concourses of some of London’s busiest railway stations 18 times. By the force’s own estimate, the cameras looked at 537,697 faces.
In those six months, the system raised one alert. It came at King’s Cross on 24 February, and it was wrong. BTP’s register records it as a confirmed false alert, with no arrest and no other outcome (British Transport Police [BTP], 2026b, p. 1). Over the whole period, the cameras made no correct match at all.
Here is my proposition: a police AI that scans half a million commuters has to prove it finds the people it is looking for, and on BTP’s own published record, that proof does not yet exist. That is not just a disappointing statistic. BTP’s own policy makes effectiveness part of the legal case for using the technology at all.
Who pays? Every commuter who walks through the camera zone has their face scanned, whether or not they are on a list, even if, as BTP says, the data of people who are not flagged is deleted at once. The public pays for the deployments. Who benefits is the question the evidence has to answer.
How the cameras decide who to flag
BTP, which polices Britain’s railways, began its live facial recognition pilot at selected London stations on 11 February 2026. In August it expanded to some London Underground stations, and the pilot is due to run until November 2026. BTP says the pilot will evaluate how effective the technology is on the railways, its impact on public safety and how the public responds (BTP, 2026a).
The system works like this. Before each deployment, officers build a watchlist of people who are wanted by the police or the courts, or who are subject to court orders with conditions. Cameras on the concourse stream images to software built on NEC’s NeoFace M40 algorithm, which compares each face it can see against the watchlist. If the similarity score passes a set threshold, the system raises an alert, and an officer compares the camera image with the person in front of them and decides whether to speak to them (BTP, 2026a). BTP’s policy is explicit that “no action will be taken based only on an alert” (BTP, 2026c, p. 9).
BTP says the biometric data of people who do not trigger an alert is deleted automatically and immediately, that images linked to alerts are deleted within 24 hours, and that the ordinary camera footage is kept for 31 days (BTP, 2026a).
The receipt: BTP’s own spreadsheet
To its credit, BTP publishes a deployment register, one row per deployment, listing station, date, watchlist size, matching threshold, alerts, outcomes and estimated faces seen. BTP’s website blocked our access, so we read the register through the Internet Archive’s copy as it stood on 3 October 2026.
| Period (BTP register) | First six months: 11 Feb to 17 Jul | Extension to 24 Sep (incl. Underground) |
|---|---|---|
| Deployments | 18 | 9 |
| Estimated faces seen (our total of BTP’s figures) | 537,697 | 389,215 |
| Alerts | 1 (confirmed false) | 4 (recorded as true: 3 confirmed, 1 unconfirmed) |
| Recorded outcome | None | “No action” for the 3 confirmed alerts |
| Arrests recorded from alerts | 0 | 0 |
Source: BTP (2026b), pages 1 and 2. The face totals are our sums of BTP’s per-deployment estimates.
Two things stand out. First, the false alert rate was very low. One wrong alert in more than half a million faces is far below the 0.1 percent that BTP’s policy describes as a typical expectation (BTP, 2026c, p. 38). Second, the system almost never found anyone it was looking for. In the extension it correctly recognised four people, and the register records that the three confirmed matches ended in “no action”. The register does not say why.
Follow the money: £320,786 and almost 100 hours
BTP’s register does not include costs. Those come from a freedom of information request by the journalism nonprofit Liberty Investigates, reported with the Guardian on 29 September 2026. According to Liberty Investigates, the FOI document shows that equipment hire and police staffing across the 18 first-phase deployments cost £320,786, and the deployments took almost 100 hours of police time (Davis & Boffey, 2026). We have not seen the FOI response itself, so we attribute these figures to that reporting.
The other side: what BTP says
BTP’s response, as reported, is that officers working alongside the cameras did make arrests, for offences including assault, theft and weapons possession, and located people wanted by courts and other forces. BTP says those arrests did not come directly from a facial recognition alert, which is why they are excluded from its performance data. It also said the growth in its watchlists was “deliberate, controlled and aligned to BTP policy” (Davis & Boffey, 2026).
Key idea: Effectiveness evidence
Any AI system that intrudes on people should be able to show, with numbers set out in advance, that it does the job it was brought in to do. For facial recognition that means two measures, not one: how rarely it wrongly flags an innocent person, and how often it correctly finds someone it is looking for. A system can be excellent at the first and still deliver almost nothing on the second. Without both, “it’s accurate” tells you very little about whether scanning hundreds of thousands of faces is worth it.
Why “does it work?” is a legal question
Live facial recognition interferes with privacy, so police must show that each deployment is necessary and proportionate. BTP’s policy tells its authorising officers how to weigh that balance, and effectiveness is first on the list: “The more effective the deployment from a policing perspective, the more that weighs in favour of it striking a fair balance” (BTP, 2026c, p. 19). The same passage asks officers to weigh the number of people likely to experience the interference, and how the deployment compares with methods that do not use facial recognition.
Read that way, the register is evidence for, or against, the legal justification. A deployment that scans nearly 69,000 faces at Euston and finds no one, as one on 17 July did (BTP, 2026b, p. 2), weighs differently from one that finds a wanted violent offender. That does not mean the pilot was unlawful, and we make no such claim. It means the case for continuing has to rest on evidence that, so far, BTP has not published.
BTP’s policy does set out the right measures. It says the true recognition rate should be tested by walking police volunteers on a test watchlist through the camera zone, that false alert rates should be recorded and reported, and that BTP will monitor “how many arrests and charges result directly from LFR alerts” (BTP, 2026c, pp. 37–39). That last measure, on the public record so far, stands at zero.
The list keeps getting longer
One more trend is visible in the register. On the pilot’s first day the watchlist held 381 people. By 14 September, at Victoria Underground station, it held 862 (BTP, 2026b, pp. 1–2). The policy’s version history shows that in July 2026 BTP changed its wording to cover adding people with ancillary court orders and court bail conditions to watchlists (BTP, 2026c, p. 47).
Bigger watchlists are not automatically wrong, and may help explain the extension’s first true alerts. But BTP’s own policy warns that the false alert rate is affected by the size of the watchlist, which “needs to be kept as small as possible, whilst still meeting operational objectives” (BTP, 2026c, p. 38). When a watchlist grows partly to include people who are complying with court conditions, a correct match may lead nowhere, which is consistent with the “no action” outcomes recorded in September.
The verdict
The cameras are good at not wrongly flagging people. On the published record, they have rarely found anyone, and no find has led to an arrest. One false alert in six months, four correct recognitions after that, three of them ending in “no action”, and no arrests recorded from alerts. BTP’s own policy says effectiveness weighs in the legal balance. So far, the scales hold very little on that side.
What a final evaluation should answer
BTP has promised a full assessment when the pilot ends (BTP, 2026a). It will land while the wider rules are still unsettled: the Home Office consulted on a new legal framework for police use of facial recognition from December 2025 to February 2026 and, as of mid-October 2026, says it is still analysing the feedback (Home Office, 2025). A useful evaluation would publish, at minimum:
- the success measures BTP set before the pilot began, and how the results compare;
- true recognition rates from the volunteer tests, by deployment and by demographic group;
- every alert, its outcome, and the reason for “no action”;
- the full cost of both phases, alongside what the same resources achieve without cameras;
- the public feedback BTP said it would collect at each deployment.
What the record does not show
Zero correct matches does not prove that the technology is inaccurate. It may mean that few people on the watchlists passed through those stations at those times. Nor does the register measure deterrence: wanted people may simply have avoided stations with visible cameras, which some would count as a benefit. We cannot see the arrests BTP says officers made during deployments, because BTP excludes them from its facial recognition data. The cost figure comes from an FOI document reported by Liberty Investigates, not from a document we have read. We read BTP’s register and policy through archived copies, and the register may have been updated since 3 October 2026. No final evaluation had been published when we wrote this.
Which leaves the question for November: when the pilot ends, will BTP publish the success measures it set in advance, and show whether the cameras met them?
What you can do, and what to ask
- If you travel through a station using facial recognition: BTP says it posts signs and offers alternative routes. If officers speak to you, BTP says they will explain why and give you a leaflet with contact details. You can also give feedback through the QR codes on the signs (BTP, 2026a).
- If you think you were wrongly stopped after an alert: ask for the reason in writing and keep the leaflet. Alert images are deleted within 24 hours, so raise it quickly.
- For police forces and public bodies: before any pilot starts, write down what success looks like, in numbers, and publish it. Then publish the result, even if it is zero.
- For anyone deploying AI that watches people: measure both failure modes, false positives and missed or useless matches, and do not let one stand in for the other.
Go further. Operational AI Readiness is FairByDesign’s free field manual for AI that is already live. Chapter 5, Human Oversight That Actually Works, covers what a meaningful human check on an alert looks like, and Chapter 6, Runtime Visibility and the Evidence Spine, covers how to record what a system did in a way that can answer the question “did it work?”. Read it free, or test one of your own systems with the free readiness assessment, paying close attention to domain 8, validation and failure readiness.
Next step: we will read BTP’s evaluation when it is published and report what it does and does not show. To follow the evidence on police AI closely, join Mercury Intelligence for $10 a month (or $100 a year; see what is included).
FairByDesign is reader-funded. If this helped, get the free Mercury Brief or see ways to support. Support never buys editorial control.
References
Bakos, V. (2026). Operational AI readiness: A practitioner guide to governing AI that is already live. FairByDesign. https://fairbydesign.org/operational-ai-readiness/
British Transport Police. (2026a). British Transport Police use of live facial recognition technology [Web page]. Retrieved via Internet Archive capture of 3 October 2026. https://web.archive.org/web/20261003230050/https://www.btp.police.uk/police-forces/british-transport-police/areas/about-us/about-us/facial-recognition-technology/
British Transport Police. (2026b). Live facial recognition deployments 2026 [Deployment register]. Retrieved via Internet Archive capture of 3 October 2026. https://web.archive.org/web/20261003230050/https://www.btp.police.uk/SysSiteAssets/media/images/british-transport-police/live-facial-recognition/lfr-deployment-register–1.pdf
British Transport Police. (2026c). Overt live facial recognition policy (Version 1.1). Retrieved via Internet Archive capture of 11 August 2026. https://web.archive.org/web/20260811180856/https://www.btp.police.uk/SysSiteAssets/media/images/british-transport-police/live-facial-recognition/lfr—policy.pdf
Davis, P., & Boffey, D. (2026, September 29). Facial recognition trial at railway stations cost £320k and had no arrests. Liberty Investigates. https://libertyinvestigates.org.uk/articles/live-facial-recognition-btp-trial-london-stations/
Home Office. (2025, December 4). Legal framework for using facial recognition in law enforcement [Consultation]. GOV.UK. https://www.gov.uk/government/consultations/legal-framework-for-using-facial-recognition-in-law-enforcement




Leave a Reply
You must be logged in to post a comment.