Twitter says its image crops weren’t very biased, but is phasing them out anyhow

Twitter is phasing out its image cropping algorithm in favor of showing larger previews, and the company has announced that it’s been looking into accusations that the system was biased. It found that the crops were slightly biased towards certain races and genders, but Twitter’s main gripe seems to be that the algorithm doesn’t give users control.

Starting in 2018, Twitter started using the algorithm to crop images so they fit in its image preview boxes. The algorithm tries to determine what will be most visually interesting to a user so it can include it can include it in the preview, but some have accused it of favoring white faces over black ones. To determine if that’s the case, Twitter had some of its researchers produce a report analyzing the choices it makes when determining crops.

When testing to see if the algorithm indeed had racial biases, the researchers’ found that the algorithm chose to crop to white women over black women 7 percent of the time, and white men over black men 2 percent of the time, with an overall 4 percent preference for white individuals.

Twitter also tested for gender bias, trying to see if the algorithm chose to crop around men more than women, and if it cropped pictures of women inappropriately (read: cropping around a woman’s chest or legs, rather than her face). It found that the algorithm favored women 8 percent of the time, but didn’t seem to crop them in a perv-y manner: in the approximately 3 percent of cases where it didn’t crop to a woman’s face, it was focused on things like a sports jersey’s number.

Image: Yee, Tantipongpipat, Mishra

Twitter admits that the algorithms could still be harmful in other ways it hasn’t tested, and that relying on machine learning to crop images takes a lot of agency away from users. The company said draws the conclusion “that not everything on Twitter is a good candidate for an algorithm, and in this case, how to crop an image is a decision best made by people.”

To let people make that decision, Twitter has made the image previews bigger in its iOS and Android apps, so the algorithm doesn’t have to make as many cropping decisions. When it does need to crop an image, like if the image is very tall or wide, the user will be shown a preview of what it will look like.

The cropping algorithm hasn’t fully gone away, especially in the case of Twitter for the web where image preview crops still abound, but Twitter seems to hope that in most cases machine learning will take a back seat to human intuition.