Twitter will fix the automatic image cropping algorithm after complaints about its racial bias
Twitter will change the automatic image cropping feature after users complained about its bias. They reported that when trying to crop images, the algorithm would highlight people with lighter skin regardless of where the original image's focus was. The version of users was confirmed even in experiments with cartoon characters and black and white dogs.
The company says it is working to "reduce the dependency" on automatic cropping, and users could ultimately control how their images are displayed on a tweet. “We hope that giving people more choices to crop images and preview how they will appear on a tweet can help reduce the risk of harm,” the company said in a blog post.
Twitter has pledged to strive for what you see is what you get. It is not yet clear when and how these changes can be implemented.
At the same time, the company said that it has conducted its own testing of the function. During the test, pairs were matched by demographic principles. Then two faces from the picture were combined into one image, randomizing their order, and tried to calculate a significance map. The maximum at this significance was attributed to a certain demographic category. The experiment was repeated 200 times for each pair. Twitter claims the experiment revealed no "racial or gender bias."
Separate testing of the function is planned with the participation of the general public.
Twitter introduced an image cropping feature in 2018, stating that it is designed to highlight "prominent" areas using neural networks.