An algorithm that has to choose where photos are cut out in a tweet appears to focus mainly on white faces.
Twitter users have found a significant problem with the automatic cropping of photos on the social network: it appears that white faces are much more likely to be the focal point than the faces of people with dark skin.
The problem seems to be with an algorithm that automatically determines where the focus of a photo is. When you publish a photo in a tweet, it is automatically cropped to the correct size. You can then click through to see the full photo.
Twitter trained a neural network to choose how to do that cutting, with a focus on faces, among other things. But that neural network has failed, as demonstrated by various Twitter users.
With several photos featuring both a white and a black man, in the same suit and with similar poses, the algorithm almost always chooses the white face.
The series of informal tests was started by a user who tried to post a tweet about the automatic face recognition in Zoom.
He too had problems recognizing the face of a black colleague. When he published the tweet, Twitter appeared to prefer to show his white face than the face of the colleague. Various further tests have shown the algorithm to have a clear preference.
Twitter itself has already apologized and indicates that the system is being worked on. “It’s clear we need to do more analysis,” said Liz Kelley of Twitter’s communications service in a tweet. “We are going to make our work open source so that others can review and replicate it.”