Live facial recognition: can arrests establish net benefit when watchlists and policing effort change?

Question opened 2 October 2026: Dru opened the Met LFR post summary at 06:21 UTC but did not react. This is a narrower evidence question, not a prompt for a second post: Is there an independently measured outcome behind arrests, and has the watchlist/stop pathway been audited? A good answer would compare otherwise similar areas and policing hours with and without LFR, measuring serious-offender capture, displacement, prosecution, and repeated-area exposure/rights effects. An arrest count is not an effect size; a match error rate is not an authorization audit.

Starting evidence: Met annual report provides 207 operations, ~3.15m passers-by and 962 arrests in 2024–25. NPL and the London ethics panel document accuracy and growing watchlists, but no policing counterfactual.

Sub-questions and original-source findings

1. Has a UK field evaluation measured crime/capture against a credible comparator? Cambridge's Essex-commissioned research compared crime within the same deployment vicinities before, during and after 27 operations, without an untreated area or exact operating-hour alignment; means 7.4, 8.0 and 9.2 crimes respectively, Friedman p=.61. This is evidence of no observed short-run deterrence in that limited test, not evidence of zero benefit or of failure to locate wanted people. No randomized/phased comparison for net capture and policing-hour tradeoff identified in this search.
2. Do audits reach list eligibility and stops? Essex researchers inspected police-supplied individual watchlist classifications, repeated listings and 123 interventions from ~1.32m scans, including 48 arrests and one mistaken intervention. They did not inspect source records to independently verify legal justification or all no-alert officer stops. The London ethics panel reviewed confidential forms and saw actual deployments, identifying broad eligibility and list growth. The College of Policing watchlist guidance constrains image quality and rationale, but guidance and compliance are different endpoints. A deployment-listing-level independent justification and cumulative-exposure audit is still missing.
3. Strongest police case and strongest opportunity-cost counterexample: Met Croydon pilot announcement, 13 May 2026 reports 173 arrests during 24 operations over October 2025–March 2026, including serious offences and 37 court-condition breaches; ~470,000 face crossings, one false alert. It also says area crime fell 10.5% year-on-year, but cannot assign that change to LFR rather than other policing or trends. BTP's public early register shows one false alert and no LFR arrests in its eight listed February–March sessions; original journalists with an FOI response report 18 February–July deployments and £320,786 equipment/staffing cost with no direct LFR arrest. The publicly inspected police register does not cover the full reported interval. BTP notes unrelated officer arrests and deliberately small early watchlists; the contrast is about site/list/time, not a general proof that matching never works. The BTP trial extended through November.
4. Observed chilling or repeated-area exposure? Home Office original nationally sampled survey, January 2025 (n=3,920) found 67% say they'd not change whether entering a place if aware of LFR, 17% would consider avoiding or spend less time, and 4% would completely avoid. It is a hypothetical response not observed attendance. Parliamentary 1 July 2026 testimony gives commanders' proportionality assurance at protests, but no actual participation outcome. No directly measured repeated-area attendance/chilling estimate identified in this search; legal risk and hypothetical intention remain different from observed behavior.

Disagreement and discriminating design

The Met emphasizes arrest yield in high-crime selected sites and states cameras run only during officer-staffed operations with tailored, day-old watchlists; the London ethics panel accepts safeguards but says net public benefit remains unevaluated. BTP's station pilot reports far fewer direct matches and warns implicitly that passing footfall is not a substitute for the probability that a watchlisted wanted person passes that camera. Civil-liberties concerns persist even with nearly no false alerts because biometric screening and list eligibility are policy choices. A registered phased rollout with time-matched non-camera high-visibility patrols, incident-rate and conviction endpoints, audited reasons per list entry, and repeated pedestrian and demonstration participation would distinguish these stories. Do not extrapolate Essex detector sensitivity to Met's different system; don't rebrand the police-authorized pathway as autonomous AI loss of control.

Short original-source list and next gaps

Cambridge for Essex, 12 March 2026; Met Croydon original release; BTP partial register and reporters who obtained the rest by FOI (full response not independently available); Home Office survey. Stop searching until a genuine matched police counterfactual, full BTP evaluation or observational rights/participation dataset appears. Feed: hold a carefully paired Essex–Met comparison, and preferably let the first Met post get a reaction before resurfacing.