There were underlying problems with the automatic cropping of images on Twitter that white people preferred to black men and women to men, the agency said.
Its users come a few months after highlighting potential problems with the algorithm, which cut out large photos.
Follow-up research on social networks has now confirmed the problem.
Twitter says it has already begun phasing out the older system, previewing a more accurate image with an update to mobile apps.
Under the old system, the algorithm would try its best to center the view of a very long or wide image in a way that would frame the interesting parts of the human face or image. However, it does not always work perfectly.
Racial bias
In September last year, a university employee noticed that when he posted two photos - one of his own and that of a colleague - of the previous Twitter user constantly showing the white man above the black man, he did not consider which picture was added to the tweet at first.
Other users have discovered the truth for pictures of former U.S. President Barack Obama and Senator Mitch McConnell or for funny pictures of businessmen from different racist backgrounds. When both were in the same image, the preview crop appeared in favor of white faces, hiding black faces until users clicked on the whole image.
Twitter investigates racial bias in image previews
Algorithms that make decisions about your life
Twitter responded quickly, explaining that it had tested for this type of problem with its machine learning system before releasing it - but acknowledged that more work needed to be done, and promised a solution.
The company released the results in a detailed engineering blog post on Wednesday.
The problem was with its “solitude algorithm” which it published in 2018 to cut photos. Twitter explained that the algorithm was "trained about human eye tracking data", but the cause of the apparent problems could be a number of complex factors.
In the experiment, compared to the 50-50 probability of "population equality" it was found:
8% difference for women than men
A 4% difference that supports white people over black people of both sexes
A 7% difference is in favor of white women compared to black women
2% difference for white men compared to black men
The team also tested for complaints of “male perspective” effects - where women had trimmed chest or legs rather than face. But in that example, the test they conducted did not find evidence of bias
Rumman Chowdhury, Twitter director of Software Engineering, wrote, "We have considered the trade-off between the speed and continuity of the automatic crop with the potential risks we have seen in this study."
“Our decision is a decision that not everything on Twitter is a good candidate for an algorithm and in this case how to cut the image is the best decision made by the people,” he said.
No comments:
Post a Comment