Twitter’s image-cropping algorithm is even more biased than previously known, researchers say.
The social media giant largely stopped using its cropping tool this year after it went viral last September for automatically highlighting white people in photos that also included a Black person.
Researchers have now found that the algorithm for highlighting faces in photos also discriminated against Muslims, people with disabilities and older people.
The image-cropping system also favored thinner and younger-looking people, another researcher showed.
The findings were part of a contest hosted by Twitter at the Def Con hacker conference in Las Vegas over the weekend to find new ways that the algorithm coded bias.
Twitter gave cash prizes to the winners, including graduate student Bogdan Kulynych, who produced the research about favoring younger and thinner people and won $3,500.
A University of Toronto AI startup called Halt AI produced the research showing additional groups that were deprioritized by the cropping tool. That research placed second behind Kulynych’s.
“Building responsible algorithms takes a village, and we’re grateful for the community helping us uncover the unintended harms and avenues for exploitation that could exist in our cropping algorithm,” Twitter’s engineering account said while announcing the winners.