The Twitter application on a smartphone. – Matt Rourke / AP / SIPA
On social media, technology when it’s fully automated isn’t always good. Twitter has just noticed that artificial intelligence does not always rhyme with ethics. The company has indeed established that its algorithm that crops photos is slightly biased in favor of white people and women. The platform therefore prefers to do without it.
“Algorithms are not suitable for everything on Twitter,” the social network acknowledged on one of its blogs on Wednesday. “It’s better for users to decide how to frame an image.”
An algorithm in service since 2018
The algorithm, put into service in 2018, is designed to crop images according to what it considers most important, in order to reduce their size and less clutter the thread of tweets. “People on Twitter have observed instances where our program chose white individuals over black,” says Rumman Chowdhury, director of software. “We tested our system on a large database to determine if there was a problem.” The researchers found that there was a 4% difference in favor of white people in general, and 7% in favor of white women as opposed to black women. The comparison between men and women in general revealed an 8% gap in favor of women.
The San Francisco-based company also looked for a possible so-called “male gaze” bias, in the sense that the algorithm would choose a woman’s chest or legs instead of her face. But only 3% of the photos tested were cropped outside people’s heads, and on non-physical elements, like a number on a sports jersey. “We did not find any evidence of an objectification bias,” says Rumman Chowdhury.
Less use of “machine learning”
The social network tested in March a new way to display images without cropping them, and then extended the tool to everyone, to allow everyone to see what their tweet looks like before posting it. “It reduces our reliance on machine learning for a task our users are better able to perform,” added Rumman Chowdhury. Artificial intelligence therefore still has progress to make.