The first time we heard the idea that Twitter's image-cropping algorithm may have some biases was in September 2020. Cryptographic engineer Tony Arcieri uploaded two images featuring Senate Majority Leader Mitch McConnell and former President Barack Obama, and the AI cropped to McConnell's face both times.
When the controversy eventually reached the folks over at Twitter, the company promised to analyze its model. Finally, we get to read through those findings.
Twitter Shares the Results of Its Algorithm Analysis
In early May, Twitter finally rolled out full-size image previews on its mobile timelines. A post to the Twitter blog has revealed that the app axed its image cropping algorithm after an analysis showed it had race and gender-based biases.
The platform started using a saliency algorithm in 2018 to crop its images. Twitter says that saliency models are trained on human eye-tracking data, which is then used by the algorithm to estimate what subjects/elements in the image it thinks is most likely to be important to most people.
The algorithm then would come up with a saliency score for all parts of the image before choosing the point with the highest score to be the center of the crop.
After running a quantitative test to check for potential biases, the team found that the saliency algorithm seemed to favor women over men, and white individuals over black individuals.
Twitter also tested the algorithm for the "male gaze," but no evidence of objectification bias was found.
Even if the saliency algorithm were adjusted to reflect perfect equality across race and gender subgroups, we [Twitter] are concerned by the representational harm of the automated algorithm when people aren't allowed to represent themselves as they wish on the platform. Saliency also holds other potential harms beyond the scope of this analysis, including insensitivities to cultural nuances.
In considering the tradeoffs between the speed and consistency of automated cropping and its risks, Twitter says its team realized that how to crop an image is a decision best made by people—not algorithms.
Twitter Gets Rid of Automatic Image Cropping
Twitter no longer uses the saliency algorithm, and now displays standard aspect ratio photos in full on both iOS and Android. When crafting a tweet, you'll find that the new update also includes a true preview of uploaded images, so that you know how everything will look once published.
If you're interested in reading the study's academic paper, you can find it on arXiv.org. You can also find Twitter's code for the analysis on GitHub.
