The company said in its latest study on Wednesday that Twitter’s image cropping algorithm has problematic biases in excluding blacks and men, adding: “How to crop an image is the best decision for people.”
The study conducted by its three machine learning researchers was conducted after users criticized the preview of images in posts other than black faces last year.
It found Compared with demographic equality, the 8% difference is in favor of women and 4% in whites.
The paper lists several possible reasons, including image background and eye color problems, but none of them have any excuses.
The researchers wrote: “Machine learning-based cropping is fundamentally flawed, because it eliminates user agents and limits the user’s expression of their identity and values, but rather on which part of the image is considered the most interesting The issue of has been paid a standard attention.”
In order to solve this problem, Twitter recently started to display standard aspect ratio photos in its mobile app, but without any cropping, and is working to expand this effort.
The researchers also assessed whether crops favor the female body, which reflects the so-called “male gaze”, but this is not the case.
The paper says that these findings are another example of the different effects of artificial intelligence systems, including demographic biases found in facial recognition and text analysis.
The work done by researchers at Microsoft and the Massachusetts Institute of Technology in 2018 and later US government research found that facial analysis systems are more likely to misidentify people of color than whites.
Amazon cancelled an AI recruitment tool in 2018 that showed prejudice against women.