Twitter Image Algorithm Shows Bias Against Multiple Marginalized Groups
Researchers discover that Twitter's controversial image-cropping algorithm exhibits ageist, ableist, and Islamophobic biases alongside its previously documented racial discrimination. The social media company hosts a hacking contest to uncover these embedded flaws in its automated systems.
Researchers reveal that Twitter's automated image-cropping algorithm displays significant bias against multiple marginalized groups. Building on last year's discovery that the system favors white faces over Black faces, a new contest at the Def Con hacker conference uncovers that the same artificial intelligence routinely ignores older people, Muslims, and individuals with disabilities.
The algorithm actively learns to exclude people with white or gray hair, individuals wearing religious headscarves, and those who use wheelchairs. Twitter hosts this first-of-its-kind competition, offering cash prizes up to $3,500 to researchers who successfully identify new ways the automated system discriminates against specific demographics.
Twitter largely decommissions the problematic algorithm after it goes viral for repeatedly cropping out Black politicians in favor of white politicians. Experts note that this situation highlights a widespread issue in the tech industry, as artificial intelligence systems trained on existing societal data inevitably absorb and replicate those same underlying prejudices.