Did you ever find it annoying when Twitter crops your photo in the most random areas of the image you posted? Well, Twitter finally released the results of their automated image cropping algorithm that users found biased. The outcome led to the app’s decision to remove the said algorithm.
Twitter users said that the app’s image cropping algorithm didn’t serve everyone equally back in October 2020. Software Engineering Director Rumman Chowdhury addressed the issue in a blog post on Wednesday. He explained that the app’s team conducted an assessment that lasted for several months. They even shared materials for those who want to read and reproduce their analysis in technical detail.
Twitter started using a saliency algorithm in 2018 to improve the image’s consistency in size in users’ timelines and be easily viewable. While the algorithm intended to determine what will be visually appealing, some users found it preferring white faces over black ones, according to The Verge.
When Twitter’s team tested the past algorithm’s race-based biases, they discovered that it cropped white women over black women 7% of the time. In contrast, it chose to crop white men over black 2% of the time. It has a 4% percent preference for the white demographic.
When it comes to testing the app’s gender bias, the team tried to see if the algorithm opted to crop around men more than women. They found out that it favored women 8% of the time. However, the algorithm did not crop them inappropriately, meaning it did not center on a woman’s chest or legs. During the 3% of cases that it didn’t crop a woman’s face, it focused on other areas like clothes with a jersey number on them.
‘We considered the tradeoffs between the speed and consistency of automated cropping with the potential risks we saw in this research. One of our conclusions is 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,’ Chowdhury concluded.
The people behind the assessment were Kyra Yee and Tao Tao Tantipongpipat from Twitter’s ML Ethics, Transparency, and Accountability (META) team. Joining them is Shubhanshu Mishra, a member of Twitter’s Content Understanding Research team.