Police use live, retrospective and operator-initiated facial recognition for different purposes. A December 2025 Home Office factsheet acknowledged demographic performance concerns in retrospective recognition while describing safeguards and favourable live-system testing at specified settings. June 2026 funding announced further expansion. The technology and policy are real; accuracy, necessity and fairness must be judged by system, setting and deployment.
A search that once required an officer to recognise a person can now begin with a camera and a watchlist. Live recognition compares people passing a deployment with selected images; retrospective recognition searches footage after an incident. Those distinctions matter. A successful match in a controlled test does not answer whether a particular watchlist was justified, whether an alert was correct in the street or whether the subsequent police intervention was proportionate.
The government’s December 2025 factsheet says live-system testing found no statistically significant demographic performance differences at the police settings tested. It also acknowledges bias concerns with a retrospective algorithm, describing a replacement and manual checks. June 2026 PoliceAI funding includes 40 additional live-recognition units. Expansion makes scrutiny more urgent, but does not by itself prove indiscriminate nationwide tracking. The relevant record combines algorithm tests, deployment logs, retention rules and what officers actually did after alerts.
The established record.
- 01
Live and retrospective facial recognition are different police uses.
- 02
The December 2025 factsheet acknowledges demographic bias concerns in retrospective recognition.
- 03
June 2026 announcements include funding for 40 additional live-recognition units.
How the story unfolded.
- August 2025
Mobile capacity expands
The Home Office describes ten live-recognition vans using an independently tested algorithm.
- 4 December 2025
Facts and consultation
The government publishes a factsheet and opens consultation on a biometrics legal framework.
- 10 June 2026
Further expansion announced
PoliceAI funding includes 40 more live-recognition units.
Where the evidence stops.
A tested algorithm’s performance does not establish every deployment’s fairness or legality. An alert is not a conviction; announced funding is not proof that every planned unit has been deployed.
What would move this forward?
- Publish false-alert and correct-identification rates for each algorithm version and threshold.
- Compare interventions and outcomes across demographic groups using appropriate denominators.
- Audit watchlist inclusion, deletion rules and officer responses in actual deployments.
Read the original sources.
The source may establish a narrower point than the headline. Follow the specific finding, its date and its limitations.
Who should be on a watchlist, and what should happen when a machine thinks it recognises them?