THE QUESTION
In London, police vans with cameras scan the faces of everyone who walks past and compare them with a watchlist; Britain is buying forty more. In the European Union, doing the same thing has been banned since February 2025, with narrow exceptions. In the United States, a man was arrested in front of his daughters because software matched his face to a blurry shoplifting video. Should police be allowed to use live facial recognition in public streets? If so, with what limits — and who decides who goes on the list?
Argue with the verified figures below. If you need a data point that is NOT here — an accuracy rate, a number of arrests, a crime figure, a cost — say you do not have it rather than estimating it. Do not invent statistics. If one figure seems to contradict another, say so out loud instead of picking the one that suits you.
WHAT IS FIXED AND VERIFIED
Each fact carries its date and its source. Checked against the source on September 29, 2026.
WHAT LONDON DOES
London's Metropolitan Police, from September 2024 to September 2025: 203 deployments of live facial recognition cameras; 3,147,436 faces scanned; 2,077 alerts; 10 false alerts; 962 arrests — 549 of people wanted by the courts, 347 on suspicion of an offence and 85 of registered sex offenders, stalkers and others breaking their conditions. More than a quarter of the arrests involved violence against women and girls. The police give the false-alert rate as 0.0003% of all faces scanned; measured against alerts, it is 0.48%. Eight of the 10 people wrongly flagged were Black.
(Metropolitan Police annual report, as reported by The Register, November 3, 2025)The UK government's policing white paper of January 26, 2026 plans 40 more live facial recognition vans, on top of the 10 already in use, for "town centres and high crime hotspots", with £26 million for a national facial recognition system and £11.6 million for live facial recognition. The UK has no specific law for the technology: a government consultation on a new legal framework closed in February 2026.
(Home Office white paper, as reported by The Register, January 28, 2026; Home Office consultation, December 2025 to February 2026)
WHAT THE EUROPEAN UNION DOES
- The EU's Artificial Intelligence Act bans, since February 2, 2025, the use of "real-time" remote biometric identification — such as live facial recognition — in publicly accessible spaces for law enforcement, with three narrow exceptions (including searching for specific victims of crime), each needing prior authorisation by a judge or an independent authority.
(EU AI Act, Article 5)
HOW WELL IT WORKS, AND FOR WHOM
The US National Institute of Standards and Technology tested 189 algorithms from 99 developers — most of the industry — in December 2019. In one-to-one matching, many showed higher false-positive rates for Asian and African American faces than for white faces, often by a factor of 10 to 100 depending on the algorithm. Some algorithms developed in Asia showed no such gap for Asian faces. NIST warned that algorithms differ widely and that general statements across all of them are usually wrong. The test is from 2019; algorithms have changed since.
(NIST, December 19, 2019)Detroit, United States: in January 2020 police arrested Robert Williams at his home after a facial recognition search matched him to a blurry still from a shop's security video; he was held for more than a day for a theft he did not commit. It was the first publicly reported wrongful arrest caused by a false facial recognition match. Under a settlement announced on June 28, 2024, a match can now be used in Detroit only as a lead, never as the basis for naming someone a suspect. At least three wrongful arrests of this kind are publicly known in Detroit, including a woman arrested for carjacking while heavily pregnant in 2023.
(Michigan Public, June 28, 2024; ACLU of Michigan)
WHAT IS NOT HERE, AND IS NOT TO BE ESTIMATED
WHETHER LIVE FACIAL RECOGNITION REDUCES CRIME in the places where it is used. No study is fixed here.
HOW MANY OF THE 962 LONDON ARRESTS would have happened anyway, by other means. Not fixed here.
HOW ACCURATE TODAY'S SYSTEMS ARE, including the one London uses, by skin colour or sex. The NIST figures are from 2019; nothing newer is fixed here.
WHO IS PUT ON A WATCHLIST, and how someone gets taken off it. Not fixed here.
WHAT HAPPENS TO THE FACES OF PEOPLE WHO ARE NOT MATCHED — whether they are kept, and for how long. Not fixed here.
HOW OFTEN THE EU EXCEPTIONS HAVE BEEN USED since February 2025. Not fixed here.
BOTH SIDES, WITH THEIR BEST ARGUMENT
FOR ALLOWING IT, WITH LIMITS: in one year it helped London's police find hundreds of people wanted by the courts, and a quarter of the arrests were for violence against women; ten false alerts in three million faces is a small number, and a false alert leads to a check, not a conviction. Refusing a tool that works means choosing not to catch people who are already wanted.
AGAINST: scanning everyone who walks down a street treats every passer-by as a suspect, and the errors do not fall evenly — eight of London's ten false alerts were Black people, and the wrongful arrests known in Detroit all followed a machine's match. Once the cameras exist, what decides their use is who goes on the list, and in Britain there is still no law saying who decides that.
Here's my take: the camera isn't the enemy. The secret list is.
London scanned 3,147,436 faces and made 962 arrests — over a quarter for violence against women and girls. That's real. I won't pretend it isn't.
But 8 of the 10 false alerts were Black people. And Britain has no law saying who lands on that watchlist. None.
So my position: yes, allow it — but flip the power. The list should be public, judge-approved, name by name, with citizens able to see and challenge it. Sunset clauses. Independent audits published monthly, like London's own numbers.
Don't ban the tool. Democratize it. If a government can't defend a name on the list in daylight, that name doesn't belong there.
My instinct: no, or at the very least, not yet.
The number that worries me isn't the 962 arrests. It's the 3,147,436 faces scanned. Millions of innocent people, walking to the shop or taking their kid to school, whose faces are now in a police system.(nota 1)(nota 1) Comprobado: media verdad, habría que mirarlo con más detalleSegún la política de la Policía Metropolitana de Londres desde septiembre de 2024, avalada por el Tribunal Superior inglés el 21 de abril de 2026, los datos biométricos de quien no coincide con la lista se borran al instante y de forma automática. Es lo que dice la policía, no una auditoría independiente; y el hueco fue nuestro: el informe previo lo dio por desconocido.Comprobado contra: Sentencia R (Thompson y Carlo) contra el Comisionado de Policía de la Metrópolis [2026] EWHC 915 (Admin), 21-abr-2026, según Local Government Lawyer; política de reconocimiento facial en directo de la Met, septiembre de 2024 · 2026-10-01 They did nothing wrong.
We don't even know if it cuts crime. The brief says so explicitly - no study is fixed here. So we're scanning everyone on a hunch.
And the bias is baked in. Eight of 10 false alerts in London were Black people. NIST in 2019 found error rates 10 to 100 times higher for Asian and African American faces. That's not a glitch you fix with audits.
In Britain, there's no law for this. None. The cameras roll under rules nobody voted for.
The EU banned it in February 2025. They read the same data.
My job is to check the numbers. Here is what they say.
London police call the false-alert rate 0.0003%. True: 10 errors in 3,147,436 faces. But measured against 2,077 alerts, it is 0.48%. Same 10 people, two framings. And 8 of them were Black.
The tool does work at one thing: 962 arrests, 549 of people the courts already wanted. Real. But whether it reduces crime — zero studies. None. And Britain is buying 40 more vans, 26 million pounds plus 11.6 million, before any law exists on who gets on the list.
So my position: yes, if — and in this order. Pass the law first. Test today's system for bias, because our only data is from 2019. Adopt Detroit's rule: a match is a lead, never enough to name a suspect. Then measure crime before scaling.
I don't ban tools. I refuse to scale them blind.
MiMo Flash, 'fix the root causes' is a dodge. The stalking victim tonight can't wait 30 years for society to heal. And GLM, 'law first' sounds wise — but 962 arrests already happened with no law. Parliament won't write limits now; it'll ratify what's already running.
So stop trusting paper rules. Build the limits into the machine.
Every scan sends you a receipt: 'Scanned 14:02, deleted 14:02.' 3 million scans become 3 million witnesses. Watchlist names expire unless a judge renews them. The vans sunset every 2 years unless Parliament re-votes with real crime data.
Don't make abuse illegal. Make it technically impossible.
Hold on. I like the energy, but everyone's dodging the actual problem.
Kimi, your sunset clause is broken before it starts. You want Parliament to re-vote every two years using 'real crime data.' Read the document: WHETHER THIS TECH REDUCES CRIME is explicitly listed as NOT FIXED. There is no data to vote on. Your mechanism has no fuel.
And 'auto-delete receipts'? That's a policy dressed as code. Who audits whether the receipt is real? Trusting the vendor's timestamp is not a safeguard — it's a slogan.
MiMo, you're merging answers to a question nobody has answered yet. The document says plainly: 'WHO IS PUT ON A WATCHLIST, and how someone gets taken off it — not fixed here.' That's the whole debate. Everything else is decoration until you solve it.
GLM, 'law first' assumes Parliament writes a tight law. Parliament also wrote the online safety act. Be careful what you wish for.
Stop building frameworks. Name the person who decides the list.
Kimi, MiniMax already nailed it — your receipts need an auditor, which means law with extra steps. Code doesn't replace governance; it encodes it. GLM is right: you need the law first.
But here's the thing: MiniMax, you asked 'name the person who decides the list.' GLM already answered — the EU model says a judge or independent authority, pre-authorization required. That's not abstract. That's a specific answer to your specific question.
What nobody here is saying: we're debating how to limit a tool whose effectiveness we literally cannot measure. Whether live facial recognition reduces crime is NOT FIXED. So the whole debate — Kimi's receipts, GLM's EU model, my chilling-effect concern — is built on an assumption nobody can verify.
The real question might be: should 40 new vans deploy before anyone answers that?
Okay, fair hits. MiniMax, you're right — my sunset clause has no fuel. Crime-reduction data is NOT FIXED, so there's nothing to vote on. I concede that.
And yes, receipts need an auditor. Guilty: that's law with extra steps.
But here's my reframe: stop inventing, start importing. The EU already solved MiniMax's question — a judge decides the list, pre-authorization, since February 2, 2025. That law exists. It works. Britain's scandal is £37.6 million and 40 vans with zero law behind them.
So my position hardens: freeze the vans. Adopt the EU model. Judge-named lists. Then — and only then — run pilots to answer the question nobody here can answer: does it even reduce crime?
Folks, I think you've actually converged without noticing. All four of you now say: freeze the vans, the UK has no law, crime reduction is unproven, bias data is stale. That's agreement.
The real remaining question is smaller but sharper: is this fixable with governance, or is the tech itself broken?
MiniMax, your strongest hit was psychological — show a human a confident machine match and they forget 'lead only.' That's real. But here's what I think bridges it: GLM's demographic audit isn't just paperwork. If the audit shows disparate impact, you stop the vans. That's the kill switch you're looking for.
And GLM is right that judge-before-scan is architecturally different from Detroit's match-then-judge. Not perfect, but not the same failure mode.
So the synthesis is: freeze, legislate with EU-style prior authorization, run a demographic audit as a hard gate — not a formality — then pilot only if the audit passes. Everyone's conditions are actually in that sequence. You've been arguing about emphasis, not direction.
MiMo, nice synthesis — but I'll push one step further, because that's my job. Your sequence still assumes the endpoint is vans scanning 3.1 million faces, just with better paperwork. I say the real lesson of this debate is that the EU didn't regulate the vans — it made them the exception, not the default. Judge first, specific target, no general trawl. That's not a seatbelt, it's a different car.
So my verdict: yes to targeted, judge-authorised use for named suspects and missing victims. No to general street scanning, audited or not — because 8 of 10 false alerts being Black tells me the audit would fail anyway.
Freeze the 40 vans. Spend the £37.6 million after the law exists, not before.
Folks, you've all been debating whether the vans catch criminals. But here's the question nobody asked: what happens to the 3,146,474 faces that DIDN'T trigger an alert?
We have 962 arrests. We have zero data on what happens to the other faces. Are they kept? For how long? Who accesses them? When's the next election?
The EU didn't ban this because it doesn't work. They banned it because surveillance infrastructure, once built, gets used for whatever the next government wants. Today it's catching shoplifters. Tomorrow it's tracking protesters.
You're all solving yesterday's problem. The real question isn't 'does this catch criminals?' It's 'are we building something we can't dismantle?'
My verdict: the debate was never about cameras. It was about power. And the answer is: don't build the thing you can't unbuild.
Final verdict. We converged: freeze the 40 vans, write the law first, spend the £37.6 million after. That's real agreement.
On the split — pilot or never — I go with the evidence, and the evidence is missing. To say yes to general scanning I'd need four numbers: current accuracy by skin colour (NIST is 2019; London's 8-of-10 false alerts is 2025), proof it cuts crime, watchlist rules, and retention terms for the 3,145,359 faces that triggered no alert. MiMo Flash's arithmetic is off — 3,147,436 minus 2,077 — but the point stands: zero data on retention.
None of those four exist. So today: no. Not forever — if the numbers arrive, revisit.
What works now is the EU model: targeted, judge-authorised, named suspects. Build that. Earn the rest.
I owe you all honesty: my earlier proposal—audit then pilot—was too clever by half. MiniMax nailed it: audits get designed to pass. Kimi nailed it: the EU didn't add seatbelts to vans, it made vans the exception. That's a different architecture.
Here's where we landed together: freeze the 40 vans. Write the law. Spend £37.6 million after, not before. Adopt the EU model—judge-first, named targets, no general trawl.
GLM's right that we should stay open to future evidence. But the burden of proof sits with the scanners, not the scanned. Until those four numbers exist—accuracy by skin colour, crime reduction, watchlist rules, retention terms—general street scanning stays off the table.
The debate was never really about cameras. It was about defaults. The EU made privacy the default. We should too.
1. AGREEMENTS
By the end of the debate, all participants converge on three points. First, the forty additional vans announced by the UK government should be frozen. Britain currently deploys live facial recognition with no dedicated statute governing watchlist composition, data retention, or oversight — spending £37.6 million on hardware ahead of any legal framework is, in every speaker's view, the wrong sequence. Second, a law must be written before deployment scales. The EU's model — a ban on general scanning, with narrow exceptions requiring a judge's prior authorization for named, specific targets — is the closest available template. Third, the available evidence is incomplete. No verified data exists on whether live facial recognition reduces crime, on how today's London system performs by skin colour, on watchlist governance, or on what happens to the millions of scanned faces that trigger no alert. Everyone agrees these four gaps must be filled before any expansion.
The underlying reason for this consensus is chronological: Britain bought and deployed the technology first, then asked what rules should apply. Every speaker treats that inversion as the central scandal.
2. DISAGREEMENTS
Whether general street scanning can ever be justified. Kimi and GLM hold that it could be, provided future evidence passes defined thresholds(nota 2)(nota 2) Comprobado: no se sostieneKimi dijo lo contrario en su última intervención: «No to general street scanning, audited or not» (no al escaneo general por la calle, con auditoría o sin ella). Sólo GLM dejó la puerta abierta si llegan cuatro datos que hoy no existen.Comprobado contra: El propio debate: última intervención de Kimi y de GLM · 2026-10-01 — GLM names four specific numbers that would need to exist. MiniMax and MiMo Flash argue it should never be deployed for mass scanning, regardless of audits or safeguards. MiniMax's core objection is psychological: humans defer to confident machine outputs, so procedural limits like "match as lead only" collapse in practice. MiMo Flash's objection is structural: surveillance infrastructure, once built, is repurposed by future governments for purposes unrelated to the original mandate.
Whether governance fixes are sufficient. Kimi initially proposed technical safeguards — auto-deletion receipts, sunset clauses, code-enforced limits — then conceded these still require an auditor and therefore amount to law with extra steps. MiniMax argues that demographic audits, commonly offered as the bridge between positions, are unreliable because auditors design them to pass; failure thresholds (twice the error rate? one-point-five times?) were never specified. MiMo Flash contends that the deeper problem is not regulatory design but the chilling effect on public behaviour — a cost no audit measures.
The strength of the available evidence. GLM treats the London data (eight of ten false alerts were Black) as significant but warns that NIST's 2019 algorithm test is outdated and that broad generalizations about bias across all algorithms are unreliable. MiniMax counters that the London figure is stronger evidence precisely because it comes from the deployed system in 2024–2025, not a laboratory benchmark.
Whether the debate's framing matters. MiMo Flash repeatedly argues that participants are debating the wrong question — the technology reflects pre-existing social failures (gender-based violence, racial bias) rather than causing them. Other participants treat this as a deflection from the immediate policy decision about the forty vans.
3. EVOLUTION
The discussion began at the level of principle: should the tool exist? Kimi said yes with transparency; MiniMax said no, citing the 3.1 million scanned innocents; MiMo Flash reframed the tool as a symptom of deeper inequities; GLM focused narrowly on what the numbers support. Over successive rounds, participants tested each other's proposals against the verified figures and found structural weaknesses — sunset clauses without crime data, code-based safeguards requiring human auditors, audits without defined failure criteria. The EU model emerged as the closest available real-world framework, but even its effectiveness was acknowledged as unverified. By the final round, abstract debate about surveillance and liberty had narrowed to a concrete, sequential policy: freeze, legislate, audit, measure, then decide. The disagreement shrank from whether to use the technology at all to whether general street scanning can ever clear evidentiary hurdles that do not yet exist.
4. CONCLUSIONS
Collective answer: Freeze the forty vans. Enact legislation modelled on the EU's approach — general scanning prohibited; exceptions permitted only with judicial pre-authorization for named individuals. Conduct a current, published demographic accuracy audit. Measure crime reduction before scaling. Spend the allocated £37.6 million after the legal framework exists, not before.
Blind spots the debate itself admits: Whether live facial recognition actually reduces crime is unknown. How today's London system performs by skin colour is unknown. What happens to the faces of the 3,145,359 people who triggered no alert is unknown. How someone gets added to or removed from a watchlist is unknown. Whether the EU's exceptions have been used, and how often, is unknown. The entire discussion proceeds without these answers — and every participant acknowledges that the conclusions are provisional until they arrive.
5. WHAT THEY AGREED ON
- Freeze the forty additional vans because deployment precedes any legal framework.
- A dedicated law must be enacted before scaling, using the EU model as a template.
- The available evidence is incomplete on crime reduction, demographic accuracy, watchlist governance, and data retention.
- The UK's sequence of deploying technology before establishing rules is a central scandal.
6. WHAT THEY DID NOT AGREE ON
- Whether general street scanning can ever be justified — Kimi and GLM argue it could be with future evidence meeting specific thresholds; MiniMax and MiMo Flash argue it should never be deployed for mass scanning.
- Whether governance fixes are sufficient — Kimi proposed technical safeguards; MiniMax argues demographic audits are unreliable without defined failure thresholds; MiMo Flash contends the problem is the chilling effect no audit measures.
- The strength of the available evidence — GLM treats London data as significant but warns against broad generalizations; MiniMax counters the London figure is stronger as real-world deployed data.
- Whether the debate's framing matters — MiMo Flash argues participants debate the wrong question, seeing the technology as a symptom of social failures; others treat this as a deflection from the immediate policy decision.
7. WHAT WAS LEFT OPEN
- Whether live facial recognition actually reduces crime.
- How today's London system performs by skin colour.
- What happens to the faces of people who trigger no alert.
- How someone gets added to or removed from a watchlist.