Twitter published on Wednesday (19) the final report regarding complaints that its algorithm has a racist bias, in addition to other choices that interfere with the automatic cutting of images on the social network.
According to the study, which was signed by three researchers from the platform, the image cropping algorithm actually made choices that were exclusive.
The internal tests carried out indicate that 8% of the choices favored women instead of men, while the preference of white people over black women was 4% – percentages considered significant in the survey.
On the other hand, there was a hypothesis that AI also preferred body parts over faces, in a possible sexualization of images. This charge, however, has been refuted.
“Cutting based on machine learning is fundamentally flawed because it removes the user’s power and restricts the expression of his own identity and values, instead of imposing a normative look at which part of the image is considered the most interesting,” says report.
The platform cited several reasons for the algorithm’s bias, but none explains all cases. The problems involve eye color and confusion with the background colors of the images. Still, the company says that cropping images “is a better job done by people” and that “not every task on the network is good to be done by an algorithm”.
A lot of work ahead
The platform began investigating the issue in September 2020, when sets of messages on the social network suggested that automatic cropping of images was skewed, giving preference to white people and, in some cases, leaving faces outside to focus on other parts of the body.
As a partial fix so far, since May 2021 Twitter has fully displayed the images in the mobile app, avoiding the use of AI to mess with the file.