Facial recognition software misidentifies man in wrongful arrest
Thames Valley Police arrested a software engineer for a burglary he did not commit after facial recognition software mismatched him with a suspect 100 miles away. Home Office research has since confirmed the algorithm produces significantly higher false positive rates for Asian and Black faces than for white faces.
What happened
Thames Valley Police arrested Alvi Choudhury, 26, a software engineer, in January for a £3,000 burglary in Milton Keynes after facial recognition software matched him with CCTV footage of a suspect 100 miles away in Southampton, where Choudhury was working at his parents' home.
Choudhury was handcuffed, held in custody for nearly 10 hours, and released at 2am. The CCTV footage showed a noticeably younger man with different features: lighter skin, appearing around 18 years old, a larger nose, no facial hair, and smaller lips, according to documents shared with Liberty Investigates.
Reported impact
Choudhury stated he was left confused and very angry about the arrest. He is claiming damages against Thames Valley Police.
Home Office-commissioned research revealed the facial recognition algorithm produces false positive rates of 5.5% for black faces and 4.0% for Asian faces compared to 0.04% for white faces at certain settings, raising concerns about in-built bias.
Index analysis
UK police forces use an algorithm procured by the Home Office from Cognitec, a German company, running approximately 25,000 monthly searches against around 19 million police mugshots on the UK-wide police national database.
Thames Valley Police stated the decision to arrest Choudhury was made after a human visual assessment as well as the algorithmic match. Police and crime commissioners warned of 'concerning in-built bias' and stated that 'there is no evidence of adverse impact in any individual case, that is more by luck than design.'
Governance controls
Facial matches should be treated as intelligence, not fact, according to the National Police Chiefs' Council.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.