British Transport Police spent £320,000 on a six-month live facial recognition trial. The system scanned more than 500,000 faces across 18 deployments at London stations. It produced one alert — which was a false positive. Zero direct arrests.[1][2]

The math: £80,000 per “alert” — and that single alert was the wrong person. That is roughly what the BTP paid for a false positive.[1][2]

( the system is NEC NeoFace M40; non-match images are deleted within 24 hours; the trial has already been extended through November and now includes the London Underground, with Victoria station as the first deployment on August 11 [1][2] )

The figures were published September 30 via Liberty Investigates FOIA[4] and reported independently by The Guardian the day before[3]. The Metropolitan Police separately scanned 4.2 million faces in 2024 with 1,100 arrests reported — but that’s a different system, a different watchlist, and a different regulatory framework.[2] The useful comparison for any LATAM procurement isn’t “LFR works in London.” It is “LFR in London public transport, with an enterprise vendor and a curated watchlist, produced 1 alert in 500K faces, 0 arrests, and £80K for that single alert”.[1][2][3]

Every time I review a facial recognition deployment, the question that has the most impact on the buyer is not “how accurate is the AI?” It is “what is the cost per real arrest?” If the answer is “we don’t know — what we know is we paid $X for the system,” the vendor has already won before the buyer ever measured ROI. Fraser Sampson, the UK’s former Biometrics and Surveillance Camera Commissioner, put it bluntly to Liberty Investigates: “We know the technology works, but using it in supermarkets to deter shoplifters and prevent attacks on staff is very different from deploying it to catch people on a public transport network.”[3][4] Bell Ribeiro-Addy, Labour MP for Clapham and Brixton Hill, told The Guardian: “Tackling violence against women and girls on public transport must be an urgent priority, but that means investing in what actually works; not spending more than £320,000 on a measure which when piloted resulted in zero arrests and a single false alert.”[3][4]

The five clauses your LATAM facial recognition RFP needs to add in Q4 2026:

Clause 1 — cost-per-true-positive. Vendor documents the cost per correct alert with arrest resulting — not per face scanned, not per raw alert.[1][2]

Clause 2 — audit-log-of-all-alerts. The system records every alert with timestamp + location + officer decision-tree + outcome (true positive, false positive, compliance-only, rejected).[1][2]

Clause 3 — human-review-before-detention. Any detention based on an LFR alert requires documented human review before the detention — the officer signs.[1]

Clause 4 — false-positive indemnification. Vendor indemnifies the buyer for any downstream liability from a false-positive detention (including reputational damage + legal costs).[1]

Clause 5 — compliance-vs-arrest disclosure. Vendor documents the percentage of alerts that are people complying with court orders (like the 3 subsequent alerts in the BTP Underground extension, which were people on sexual-harm prevention conditions, not arrest targets [2]) versus actual arrests. Without this clause, the “ROI” the vendor reports may be 100% compliance alerts nobody needed.

P.D. How many real arrests has your facial recognition system generated in the last 12 months? If the answer is “we don’t track them separately from alerts,” today’s a good day to start. BTP showed that the “alerts to arrests” ratio can be 1 to 0 — and that £320K later, the real ROI is zero.[1][2][3][4]

📰 Source: The Register + BiometricUpdate (Sep 30 2026) + Liberty Investigates FOIA + Jedi internal analysis.

Sources

[1] The Register — UK rail cops’ £320K face-scanning spree nets zero matches (Lyons, Sep 30 2026) — https://www.theregister.com/security/2026/09/30/uk-rail-cops-320k-face-scanning-spree-nets-zero-matches/5299793 [2] BiometricUpdate — London expands LFR despite weak results from rail trial (Sep 30 2026) — https://www.biometricupdate.com/202609/london-expands-lfr-despite-weak-results-from-rail-trial [3] The Guardian — Trial of live facial recognition cameras at London stations produced false positive (Sep 29 2026) — https://www.theguardian.com/technology/2026/sep/29/trial-live-facial-recognition-cameras-london-stations-false-positive [4] Liberty Investigates — Live facial recognition BTP trial London stations (FOIA source) — https://libertyinvestigates.org.uk/articles/live-facial-recognition-btp-trial-london-stations/