Essex–Cambridge LFR study: identifications, watchlist composition, and a limited deterrence test

Original study: Matt Bland and Jacob Verrey, University of Cambridge, Live Facial Recognition: Accuracy, Watchlists, and Deterrence, final report 12 March 2026. Commissioned by Essex Police; first two studies use field volunteers and police-supplied records; third compares nearby crimes in periods around deployments. This is Essex, not London's Met, using a different system and watchlist scale. Connect to counterfactual question and NPL/ethics evidence.

What it directly observes

• Technical handoff: 188 volunteers staged passing operational cameras; at Essex's threshold 55 the system identified about half of watchlisted participants who crossed. Men and Black participants were more likely to be identified than respective comparators in that test. A single-day volunteer trial is not a population-wide catch rate. Lowering threshold to 40 was modeled to improve sensitivity by ~22 percentage points with no observed increase in false positives at typical watchlist sizes; small false-positive counts (six under extreme gallery/threshold conditions, four involving Black subjects) do not prove equal risk in larger rollouts. Do not transfer the settings to the Met's 0.64 threshold.
• Authorization inputs and outputs: Essex supplied deployment summaries, individual watchlist demographics/offence codes, and alert-intervention outcomes across 41 operations (August 2024–February 2025). Median watchlist 790, range 191–1,403; 32,677 person-listings represent 7,304 distinct people, 4,159 appearing at least twice and 141 on ≥half of lists. About one-third of offence classifications were ABH/common assault, ~one-fifth harassment, stalking or coercive control, ~11% shop theft. These are classifications linked to inclusion, not convictions. Roughly 1.32m faces scanned → 123 officer interventions → 48 arrests (one unrelated to watchlist), 44 other resolutions, 30 no further action, one false-positive intervention. No-further-action is not synonymous with misidentification or wrongful intervention. Researchers checked consistency but did not independently inspect source systems or legal justification per record, limiting any claim that criteria were audited and passed.
• Deterrence rather than identification: across 27 analyzable deployment windows the study compared recorded public-space crimes in the same vicinity on the day before, deployment day and day after. Mean 7.4, 8.0 and 9.2 per period, Friedman χ²=1.0, p=.61. Crime on the deployment day was not timed exactly against the few hours cameras were operating; drug, public-order and officer-assault records were excluded to reduce police-presence recording bias. No nearby untreated control area, randomized timing or power for modest effects. This does not establish zero deterrence, let alone zero capture benefit; authors say deterrence alone should not justify deployment.

Interpretation and strongest objections

The study challenges both lazy positions: operational false arrests are not evidently common in this sample, but reducing false positives does not answer whom police put on the lists, why, or whether benefits beat staffing and biometric exposure. The authors argue missed wanted-person identifications can also produce unfair under-policing; this presumes lawful, proportionate underlying watchlists and beneficial enforcement, which the test does not independently establish. Watchlist composition matters because an LFR system cannot match a subject absent from the list and a broad list increases who is eligible to be stopped. Contrast the Met’s 962 arrests and very low false-alert count with a modestly powered same-location before/during/after Essex deterrence test: they answer different questions.

Open

Replicate sensitivity for other devices/sites with known true watchlisted crossings; independently inspect original reasons for each listing, repeated exposure and officer stops; compare phased camera days versus matched high-visibility officer patrols without LFR, followed by serious-offender capture, prosecutions and crime over months. Do not claim this report is a net-benefit evaluation.