British Bangladeshi man wrongfully arrested for theft after facial recognition error
A 26-year-old British software engineer was wrongfully arrested for a burglary 115 miles from his home after facial recognition software misidentified him as a suspect despite obvious physical differences and his solid alibi. The Home Office's commissioned research later revealed the Cognitec system produces error rates for Asian faces 100 times higher than for white faces at the police's chosen settings.
What happened
Software engineer Alvi Choudhury, 26, was arrested at his home in Southampton on 25 January 2026 after facial recognition software used by Thames Valley Police flagged him as a match for a burglary suspect in Milton Keynes, 115 miles away. He was held in custody for nearly 10 hours before being released.
The CCTV footage showed a noticeably younger man with different facial features, including no facial hair, a lighter complexion, and different eye shape and nose size. Choudhury had never visited Milton Keynes and had a sound alibi.
Reported impact
Choudhury experienced emotional distress and reputational damage in his neighbourhood following the public arrest at his home in front of neighbours and parents. He fears remaining in police databases could lead to future wrongful arrests and worries the incident may jeopardise his chances of obtaining security clearances for his career.
The case illustrates broader harms to public trust in both facial recognition tools and police deploying them, particularly among ethnic minority communities already experiencing disproportionate policing.
Index analysis
The Cognitec software, procured by the Home Office from the German company, runs approximately 25,000 monthly searches against 19 million police mugshots on the UK-wide police national database. Thames Valley Police claimed a human visual assessment confirmed the match, but officers later admitted Choudhury looked nothing like the CCTV footage.
Home Office commissioned research revealed that settings used by police produce a 4.0% error rate for Asian faces compared to 0.04% for white faces. Police reportedly admitted the arrest "may have been the result of bias within facial recognition technology." The police were using outdated algorithms from 2020, according to government assessment.
Governance controls
Facial matches are intended to be treated as intelligence rather than fact according to National Police Chiefs' Council guidance. However, automation bias appears to have led officers to rely on the software match over their own visual assessment.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.