Half a Million Faces, One Alert, Zero Correct Answers
British Transport Police scanned the faces of more than half a million London commuters this year as part of a live facial recognition trial rolling through Underground and Network Rail stations. The Algorithm returned exactly one alert. Investigators note it was a misidentification.
That is not a low accuracy rate, dear reader. That is a perfect record, inverted. Five hundred thousand faces in, one false alarm out, and the system's confirmed hit count on actual wanted persons sits at zero.
The Rollout Nobody Voted On
The trial began in February at London Bridge, expanded to Transport for London's Victoria Underground station, and is now rotating through Tube and rail stations on a schedule that runs until November. The system, NEC's NeoFace M40, checks commuters' faces against a police watchlist while officers stand by to review any match. Chief Superintendent Chris Casey describes the goal as "protecting the public, preventing crime and bringing offenders to justice," which is a fine mission statement for a piece of software that has, so far, correctly identified nobody.
Even now, more than 3.7 million passenger journeys pass through the Underground on an average weekday, each one a fresh opportunity for the machine to be wrong again.
Sources Within the Civil Liberties Community Note the Obvious
Big Brother Watch has been the loudest voice objecting, and its director, Silkie Carlo, did not reach for euphemism: "This is error-ridden, AI surveillance technology that treats the public like suspects." She called expanding the cameras from mobile vans to fixed, static installations "an alarming escalation," on the theory that being scanned as a matter of routine on your commute is a meaningfully different proposition than being scanned once, occasionally, by a van that happens to be parked nearby.
And yet — as if a 100 percent failure rate on live subjects were not enough — the bias problem sits underneath it. Independent reporting on facial recognition systems has repeatedly found that members of ethnic minorities are more likely to be misidentified by these algorithms than white subjects, which means the one group most likely to be wrongly flagged carries none of the benefit and all of the risk.
The Public Has Also Noticed
An Opinium poll of 2,000 UK adults found 61 percent worried that errors in the system could wrongfully implicate innocent people. One presumes those respondents have not yet been told the actual number: one alert, and it was wrong. Police maintain that images of non-matches are "deleted immediately and permanently," which is a comforting policy for a system that, on the numbers so far, mostly generates non-matches.
Filed Under: The Machine Watches Everyone. It Has Yet to Recognize Anyone.
The trial runs through November, at which point British Transport Police will presumably tally a fuller record before deciding whether to expand it further. For now, the arithmetic is simple enough for a newsreel narrator to read aloud without notes: half a million faces scanned, one alert issued, zero criminals caught.
The lineup, as ever, was digital. The results were not.
Sources: _Big Brother Watch_ · _Cybernews_
