The crop that chose a face

In September 2020, Colin Madland, a university administrator from the Vancouver area, called out on Twitter a Zoom virtual background that erased his Black colleague's head. He then noticed that Twitter's automatic preview kept only his own face. Thousands of users reproduced the test; one experiment seemed to show that the preview preferred the face of white senator Mitch McConnell to that of Barack Obama.

Twitter did not simply deny it. On May 19, 2021, its ethics team published an audit of the "saliency" algorithm that chose which part of the image to show. Between Black and white people, it departed from demographic parity by 4% in favour of white people; between men and women, by 8% in favour of women. The company replaced automatic cropping by showing the whole image, concluding that this choice belongs to people.

In August 2021, Twitter launched a public "bias bounty" contest. Participants found other blind spots: the tool tended to leave out older people with white or grey hair, to crop out people in wheelchairs and to prefer English text over Arabic text. A valuable lesson: a machine's biases are easier to see when many people look. What remains is to make this transparency the rule rather than the exception, before a tool decides who deserves to be seen.