500,000 faces, one false alarm: what the rail facial-recognition trial really tells us

The figures from British Transport Police's live facial-recognition trial, obtained by the civil-liberties group Liberty and reported this week, read like a punchline. Over six months, cameras at London railway stations scanned the faces of more than five hundred thousand people. Officers spent close to a hundred hours on the deployments. The bill came to more than £320,000. And the system produced one alert, which turned out to be a false positive, and no arrests. When the trial was later extended onto the Underground, it produced three further alerts, all of them people already known to the authorities and already complying with existing court orders. Half a million strangers examined; nobody new found.

It would be easy to write this up as a technology that does not work, and easy to write it up, from the other direction, as proof that the police were reckless with the public's most personal data. Both readings are available and both are being made loudly this week. I think both miss the more useful point, which is that the result was largely predictable before the first camera was switched on, for reasons that have nothing to do with whether the software is any good, and that the reasons are exactly the ones any business weighing up facial recognition, or any other biometric system, needs to understand before it spends the money. This is not really a policing story. It is a story about arithmetic, about what the law calls special category data, and about the difference between a capability and a justification.

What actually happened

The technology in question is live facial recognition, LFR, which is the real-time kind: cameras watch a crowd, the system builds a biometric template of each face that passes, and it compares that template against a watchlist of people the police are looking for. British Transport Police ran it, using NEC's NeoFace platform, at London railway stations over roughly six months, describing the deployments as intelligence-led and aimed at crime hotspots and high-harm offenders. The force has been clear about two points in its defence: the system cannot identify anyone who is not on the watchlist, and the biometric data of the people who do not match is deleted immediately rather than retained.

Both of those statements are true and both matter, and I will come back to them, because the gap between "the data is deleted immediately" and "the data was never processed at all" is where the entire disagreement lives. For now, hold the shape of the result in mind: an enormous number of people processed, a watchlist presumably numbering in the hundreds or low thousands, and almost nothing found.

Why "almost nothing found" was the likeliest outcome

Here is the part that gets lost in the noise. Suppose the software is excellent. Suppose it matches a face to a watchlist entry with, say, 99.9 per cent accuracy, which would be a very good system. Now point it at a London terminus, where the overwhelming majority of the half-million faces belong to commuters who are on no watchlist at all and never will be. The number of genuine watchlist members walking through is, in any given period, very small; most days it is probably zero. So the system spends essentially all of its time correctly deciding that person after person is not a wanted individual, which is not useful because we already knew that, and the rare occasions it has anything to say are dominated not by the vanishingly few real matches but by the small percentage of mistakes it makes on the enormous innocent majority.

This is the base-rate problem, and it is not a flaw you can engineer away with a better camera. When the thing you are searching for is genuinely rare in the population you are searching, even a very accurate test produces alerts that are mostly false, and a low total number of alerts overall. A single false positive from half a million scans is not evidence that the software is broken; if anything it suggests the threshold was set conservatively to avoid stopping innocent people. The uncomfortable truth for proponents is that the maths does not improve with better technology, because the limiting factor is not accuracy but rarity. The uncomfortable truth for opponents is that "it made no arrests" is therefore not, on its own, the killer argument it sounds like; a tool can be statistically almost guaranteed to find little and still be lawful and proportionate if the few things it finds matter enormously. The real questions are the ones underneath.

The data-protection questions the numbers force

A face is not an ordinary piece of data. Under UK data protection law, biometric data processed for the purpose of uniquely identifying someone is special category data, in the same tier as health records, religion and sexual orientation, and it may only be processed where a demanding set of conditions is met. That framing is the whole game, and it is why the "we delete non-matches immediately" defence does not settle the matter. Building a biometric template of a commuter's face in order to compare it against a watchlist is processing their special category data, whether or not the template survives the next second. The deletion is a safeguard applied after the processing has already happened, not a reason the processing did not happen. Fraser Sampson, until recently the UK's biometrics and surveillance camera commissioner, made exactly this point in response to the trial: the system is handling special category personal data on a mass scale, even though, as he put it, if you are not on the watchlist the cameras do not "see" you in any way that persists.

That is what campaigners are objecting to when Big Brother Watch describes ordinary travellers as caught in a "digital police line-up" and Privacy International points out that thousands of people have highly sensitive data processed in public, often without realising it. Whether one finds that objection decisive or overblown, the legal architecture takes it seriously, and the architecture is the same one that governs a private business. And there is a genuine, unresolved backdrop here: the United Kingdom still has no statutory framework written specifically for facial recognition. Its use by police has been built on a patchwork of general powers, data protection law and case law rather than a purpose-built Act, and deployments were paused earlier in 2026 while independent testing examined concerns about accuracy varying across demographic groups at certain thresholds. When the rules are this unsettled for the police, a private organisation contemplating the same technology is standing on even softer ground.

The lesson for any organisation weighing biometrics

I am not going to tell you whether the police should use live facial recognition; that is a question for Parliament and the courts, and reasonable people hold firm views on both sides. What I can say with some confidence is what a business should take from this trial before it puts a face-matching camera in its lobby, its shop or its stadium, because the same forces that produced a £320,000 bill for one wrong alert will operate on you too, without the public-interest justification a police force can at least argue.

Start with necessity, which the law demands you establish before proportionality even arises. The question is not "would facial recognition be useful here" but "is there a real problem that facial recognition is the least intrusive effective way to solve". Most commercial deployments fail at this first hurdle: the loss-prevention or access-control goal can usually be met by means that do not involve processing everyone's biometrics, and if a less intrusive method would work, the biometric one is not necessary in the legal sense. Then confront the base rate honestly. If the people you actually want to catch are rare among the people you will scan, your system will spend its life correctly ignoring innocent people and occasionally accusing one of them, and you need to decide in advance what happens to the person your system gets wrong, because there will be one before there is a real match.

Treat the biometric templates as the special category data they are, which means a Data Protection Impact Assessment is not optional but a legal prerequisite, and it must be done before deployment, not written up afterwards to justify a decision already taken. That assessment has to name your lawful basis and your special category condition, both of which are far harder for a private company to satisfy than for a police force acting for law enforcement purposes. It has to define the watchlist: who goes on it, on whose authority, on what evidence, how someone comes off it, and how long entries last, because an ungoverned watchlist is where this technology does its real damage. It has to specify retention, ideally the immediate deletion of non-matches that BTP at least implemented, and it has to plan for transparency, because processing people's faces covertly removes the last shred of a proportionality argument. And it has to reckon with the demographic-accuracy problem: a system that is less accurate for some groups than others does not merely underperform, it discriminates, and that is a liability in its own right.

None of this is an argument against the technology in principle. It is an argument for doing the arithmetic and the paperwork before the procurement, rather than discovering afterwards that you have spent a large sum to process half a million people's most sensitive data in exchange for one mistake. The organisations that get biometrics right are the ones that could answer the necessity question honestly, sized the watchlist and the base rate before they bought anything, and wrote the impact assessment as a decision aid rather than an alibi. That is unglamorous work, and it is the difference between a defensible programme and a headline.

The behavioural footnote

There is one more thing worth saying, because it sits underneath the whole debate. The appeal of a face-scanning camera is not really operational; it is emotional. It feels like doing something. It projects capability and control, to the public, to a board, to oneself. The trial is a useful corrective precisely because it separates the feeling from the result: a great deal of visible activity, a great deal of money and public data expended, and almost nothing to show for it. Before adopting any surveillance technology, the honest question is whether you are buying an outcome or buying the appearance of one, and the way to tell the difference is to write down, in advance, what result would count as success and what you will do if, as here, it does not arrive. If you cannot answer that before switching the cameras on, the half-million faces were never going to be the point.

Frequently asked questions

Does "no arrests" mean facial recognition does not work? Not on its own. When the people on a watchlist are rare among the people being scanned, even an accurate system will generate very few alerts and a high proportion of the ones it does generate will be wrong. That is a property of the statistics, not necessarily of the software. The trial tells you the yield was low and the cost per useful result was very high; whether that makes it worthwhile is a proportionality judgement, not a purely technical one.

Is facial recognition legal in the UK? There is no statute written specifically for it. Its use is governed by a combination of general legal powers, the UK GDPR and Data Protection Act 2018, human rights law and case law. Biometric data used to identify people is special category data and attracts the highest level of protection, so any use, public or private, must clear a demanding necessity and proportionality test and, in practice, a Data Protection Impact Assessment.

Can a private business use live facial recognition, for example for shop security? Technically yes, but the legal bar is high and harder to meet than for the police, because a company cannot rely on law-enforcement purposes as its justification. It must establish necessity, identify a valid lawful basis and special category condition, complete a DPIA before deployment, govern its watchlist tightly, be transparent with the people being scanned, and address demographic accuracy. Many proposed commercial uses fail the necessity test because a less intrusive method would achieve the same goal.

What is the base-rate problem in plain terms? If you search a very large group for something very rare, most of your positive results will be false even when your test is accurate, simply because there are so many more innocent cases than genuine ones for the test to make mistakes on. It is the same reason a highly accurate medical screening test for a rare condition still produces mostly false alarms. In facial recognition it means scanning a crowd for a small watchlist yields little of value and disproportionate processing of everyone else.

We are considering biometrics at our site. What is the first step? Before anything else, write down the specific problem you are trying to solve and ask whether a less intrusive method would solve it; if one would, you do not have a lawful case for biometrics. If you still think you need it, commission a Data Protection Impact Assessment at the design stage, not after procurement. If that would help, it is the kind of work our data-protection and virtual-CISO practice does, and it is far cheaper than discovering the answer the way this trial did.