Amazon’s racially skewed face recognition misidentifies 28 members of congress as criminals in a £9 test
Amazon’s face recognition technology has falsely equated no less than 28 members of the US Congress with convicted criminals, a study has found. Rekognition, the controversial technology helmed by Jeff Bezos, exhibited findings that were not only alarmingly inaccurate but also racially skewed.
The trial was conducted by the American Civil Liberties Union (ACLU) of Northern California, which tasked Rekognition with matching 535 members of Congress with 25,000 publicly available mugshots. The endeavour resulted in over two dozen surprising – and falsified – matches between politicians and convicted criminals. The humiliating test cost Northern California’s ACLU just $12.33 (£9.40) to conduct.
More alarmingly, the ACLU’s test showed that 11 of the misidentified members of Congress were people of colour. This highlighted a troubling disparity between the technology’s apprehension of lighter and darker skinned people; in the Congress microcosm, the error rate was 5%. Among non-white members of Congress, the error rate jumped nearly 800%, coming in at a shocking 39% of inaccuracies.
In a country – indeed, a hemisphere – where institutionalised racism is so ingrained, the last thing anyone needs is an unwieldy and illiberal facial recognition tool entrenching that racism further. Particularly not when it’s backed by the richest man in the world.
It’s not even the mismatches alone which are troubling. As the ACLU points out in a blog post, officers could get “matches” for previous convictions, such as concealed-weapon arrests, a happening which could result in a pre-existing bias “before an encounter even begins”.
For its part, Amazon conceded that the technology was far from perfect, saying in a statement that Rekognition’s software “could probably be improved by following best practices”. It also posited a 95% confidence threshold for authorities deciding if there’s a match.
As such, Rekognition is joining the ranks of other arbitrary and inaccurate facial recognition technologies. Not least of all one trialled by London’s Metropolitan Police back in May, which wielded an almost impressively erroneous 98% inaccuracy rate. Sounds like there’s more than a little tinkering to be done on the facial recognition technology platform.
In the meantime, the fate of Rekognition remains in the lurch. Although it has to be said, any technology named like a mid-90s boy band doesn’t exactly have us brimming with confidence.