Twitter Inc's image cropping algorithm has a problematic bias that excludes dark-skinned people and men, the company said in a new study Wednesday, adding that "how to crop an image is a decision best made by humans".

The study, conducted by three machine learning researchers, follows criticism from users last year that image previews in messages excluded the faces of dark-skinned people, the firm reported in a blog post.

Women and white faces favored

She found that the difference with demographic parity in favor of women was 8% and in favor of white faces 4%.

The article cites several possible reasons, including background issues and eye color, but says none of them are excuses.

"Machine learning-based cropping is fundamentally flawed because it disempowers users and limits the expression of their own identity and values, instead imposing a normative view of which part of an image is considered most interesting", the researchers write.


To address this problem, Twitter recently began displaying photos entirely with a standard aspect ratio - without cropping - in its mobile apps and is trying to expand this practice.

Crop the algorithm...

The researchers also investigated whether cropping favored female bodies over heads, reflecting the so-called "male gaze," but found that it did not.

The results are another example of the uneven impact of artificial intelligence systems, including the demographic biases found in face recognition and text analysis, the paper said.

Work by researchers at Microsoft Corp and the Massachusetts Institute of Technology in 2018 and a more recent U.S. government study found that facial analysis systems misidentify people of color more often than white people.

In 2018, Amazon Inc rejected an AI recruitment tool that showed a bias against women.