Twitter announces 'stop automatic trimming to eliminate racism'



On May 19, 2021, Twitter announced the results of a survey that the ability to automatically crop images tended to 'prioritize whites'. In response, the company has decided to phase out the algorithm for the automatic trimming function and leave the trimming decisions to the user.

Image Cropping on Twitter: Fairness Metrics, their Limitations, and the Importance of Representation, Design, and Agency
(PDF file)

https://arxiv.org/pdf/2105.08667.pdf

Sharing learnings about our image cropping algorithm
https://blog.twitter.com/engineering/en_us/topics/insights/2021/sharing-learnings-about-our-image-cropping-algorithm.html

Twitter says its image crops weren't very biased, but is phasing them out anyhow --The Verge
https://www.theverge.com/2021/5/19/22444372/ twitter-image-crop-racial-gender-bias-research

In 2018, Twitter introduced the optimization of automatic cropping of images using AI technology. You can understand how this algorithm determines the trimming range by reading the following article.

Twitter introduces optimization of automatic image trimming by AI technology --GIGAZINE



With the introduction of the algorithm, when a person appears in the image, it has been improved so that the person is centered as much as possible and trimmed. There was a voice asking, 'Isn't there?'

So, when Twitter examined whether the algorithm was racially biased, the algorithm prioritized the face of white women over black women at a rate of 7%, and the face of white men over black men at a rate of 2%. It turns out that there is a tendency to prioritize. Overall, the algorithm prioritized whites at 4%.

In addition, when examining the bias between genders, it was found that there was a tendency to prioritize women over men at a rate of 8%, but there was no tendency to 'prioritize chest and legs over face'. is.



In response to the findings, Twitter's director of software engineering, Rumman Chowdhury, 'examined the trade-off between the speed of automatic cropping and the potential racist risk found in the findings. , We have come to the conclusion that it is best for people to decide how to crop the image. '

Prior to this announcement, Twitter made changes to the app in early May 2021 to prevent the image on the timeline from being cropped, which included 'displaying the image as large as possible and automatically cropping it with an algorithm. The aim was to 'minimize the impact.' In addition, when trimming is necessary, a preview is displayed and it is left to the user to judge the range to be cut.

Related Posts:

in Web Service, Posted by log1l_ks