Twitter first introduced the “auto crop” function in 2018. When changing the aspect ratio of the screen, it uses AI to determine the human “visual focus” and automatically decide where to place the center.However, at the end of last year, there were allegations that Twitter’s AI tailoring would be particularly biased towards light-skinned and females, although Twitter preliminary investigationIt shows that the cutting is not biased, but the complete survey released today shows that it is still a bit biased.
In comparison between men and women, Twitter’s AI algorithm is 8% female, while in skin color, it is 4% white. However, Twitter did not find a problem with “male perspective” (staring at the chest or butt). Of the 100 photos, only three were not focused on the face, and these three were not particularly looking at.
Twitter said that it will feel that the crop is looking at the chest, but it may actually have noticed the numbers on the jersey, but this is not an excuse. After all, machine learning prevents people from expressing their own personality and value, because whatever the machine sees, it may not be what users want. Therefore, Twitter’s conclusion is that image cropping itself is not something that should be left to the algorithm, or that it is better for humans to decide on their own.
As a countermeasure, Twitter will launch full-size image previews on Android and iOS, and will continue to improve the way the platform handles media.