Live face recognition: independent technical test and ethics review

Original sources: National Physical Laboratory, commissioned test report, March 2023; London Policing Ethics Panel, Five Years On, June 2025 (published July); ICO overview of five consensual police-force audits, 18 August 2026. Set against the Met’s operational report in the state-power question.

Technical evidence and endpoint

NPL used July–August 2022 footage from five live policing deployments, seeded a cohort of subjects into the crowd, then replayed footage offline against their images and a large filler gallery. For a 10,000-person watchlist and similarity threshold 0.6, estimated true-positive identification was 89% for watchlisted passers-by and false-positive identification ~0.017% (roughly 1 per 6,000 non-watchlisted passers-by), with error sensitive to gallery size and threshold. Demographic variation in false positives was not statistically significant at ≥0.6, but became significantly higher for Black subjects at lower thresholds; missed matches for under-20s were more common, plausibly partly crowdedness/height. NPL did not test officers’ decisions, the legitimacy of a watchlist or a 2025-era static-camera configuration. Do not substitute NPL’s 0.6 results for the Met’s operational 2024–25 0.0003% reported false-alert rate without matching settings and populations.

Governance is a separate bottleneck

The London Policing Ethics Panel sampled confidential authorisation forms, observed three deployments and conducted an updated public survey. It credited time/geography limits, signed areas, human intervention and useful specific identifications, but noted average watchlist growth from ~1,200 to ~15,000, broad eligibility including suspects and monitored individuals, cumulative repeated deployment not yet assessed, and incomplete publicly evaluable cost/benefit data. It recommended monitoring repeat deployments by location, clearer avoidance signage, local community consultation, outside oversight, independent evaluation and clearer legislation. This is not a claim that present matching software inevitably fails. The Met’s 2025 first annual report answers some transparency requests with offence and outcome breakdowns, but still does not identify a no-LFR policing counterfactual.

The independent ICO’s August 2026 overview found five other consenting forces generally documented legal bases and LFR data minimisation; across forces it issued 107 accepted or partly accepted recommendations on records, senior oversight, testing and bias. It explicitly had not yet audited the Metropolitan Police; do not cite the five-force audit as a Met compliance certification. The ICO’s separate warning about a 2025 National Police Database retrospective face-search algorithm is not a result about the Met’s LFR false-alert rate.

Disagreement

Big Brother Watch calls scanning all passers-by a mass identity check irrespective of accurate matching, alleges protest chilling and fears that preexisting CCTV can become biometric checkpoints; it contests a broad police-authored eligibility policy. The ethics panel, unlike BBW, regards limited LFR as potentially defensible if net public benefit and safeguards are evidenced. Police self-report and the panel both document serious-offender apprehension; neither settles whether such benefits exceed privacy and institutional trust costs under large or repeated deployments. A 2019 panel survey measured stated event avoidance, not observed attendance changes; avoid claiming demonstrated protest suppression.