Welsh Police Facial Recognition Scans Ethnic Minorities at 4x Their Population Share

FOI data shows 24% of South Wales police face scans target ethnic minorities who make up 7% of the population — a disparity that persists even against arrest demographics.

South Wales police officers using smartphone-based facial recognition scanned ethnic minority individuals at roughly four times their share of the local population in 2025. Freedom of Information data covering a full year of the force's "officer-initiated facial recognition" (OIFR) programme shows that 24% of people selected for face scans were from ethnic minorities, against a population share of just 7%. Even measured against the force's own arrest data — where ethnic minorities account for 11% — the disparity is stark and unexplained.

The OIFR system works by having officers capture a person's face on a smartphone and run it through NEC's "Neoface" algorithm, matching against custody mugshots and other face databases. In 2025, officers initiated scans of 257 adults and 14 children, including two under-10s. More than half were suspected of crimes, being unlawfully at large, or breaching bail conditions. Where matches were made, arrests followed in 48% of cases; a third resulted in no further action. The case of Alvi Choudhury, 26, illustrates the downstream consequences. Arrested in February for a burglary in a city he had never visited, Choudhury was misidentified after the software confused him with another person of south Asian heritage in the custody database. His experience underlines the compounding risk: if ethnic minorities are scanned disproportionately, and the algorithm has higher error rates for non-white faces, the probability of wrongful arrest concentrates on exactly the communities least positioned to absorb it. The force acknowledged the figures "raise concerns about disproportionality" but defended deployments as individually justified and subject to supervisory scrutiny. It pointed to the small sample size and described the technology as being "in its infancy." The infancy defence cuts both ways — patterns established during rollout tend to calcify as institutional norms, not disappear with scale. The political context is split cleanly. Home Secretary Shabana Mahmood has championed facial recognition as the biggest policing breakthrough since fingerprints and DNA. Croydon's lamp-post cameras led to 173 arrests between October 2025 and March 2026, including rape and kidnap suspects. That genuine crime-fighting utility makes it harder — not easier — to demand proportionality, because the technology demonstrably works for the majority population while externalising its failure modes onto minorities. The Home Office announced a new facial recognition watchdog earlier this month with powers to shut down non-compliant forces. The police reform bill would also impose a statutory duty for face scans to be intelligence-led and proportionate. Whether "proportionate" is defined against population share, arrest rates, or some other denominator will determine whether oversight has teeth or merely provides institutional cover. The deeper structural question is who bears the cost of the learning curve. Small sample sizes don't reduce harm to the individuals scanned — they just make the statistics noisy enough for institutions to claim ambiguity. If 24% becomes the baseline ratio and is never corrected because each individual scan is deemed individually justified, the system will have encoded racial disproportionality as standard operating procedure, supervised and approved at every level.