Twitter began its investigation after users claimed that the image cropping feature favored white faces.
The social network mobile app’s automated tools automatically crop images that are too large to fit on the screen and choose which part of the image to crop.
However, experiments by graduate programmers seemed to show racial prejudice.
To see what Twitter’s algorithm chooses, Tony Arsieri takes a headshot of Senator Republican leader Mitch McConnell and a photo of former US President Barack Obama separated by a white space below. I posted a long image featuring it.
In the second image, Obama’s headshot is placed at the top and McConnell is placed at the bottom.
In both cases, the former president was completely removed.
Following the “horrible experiment” after the image he posted cut out a black colleague, Arcieri wrote:
At the time of writing, his experiment has been retweeted 78,000 times.
Twitter has pledged to investigate the issue, but said in a statement:
“From these examples, it’s clear that more analysis is needed. We will continue to share what we have learned and the actions we have taken, making the analysis open source for review and replication by others.”
A Twitter representative also pointed out a study by scientists at Carnegie Mellon University who analyzed 92 images. In this experiment, the algorithm prioritized the black side 52 times.
The company said in 2018 that the tool is based on a “neural network” that uses artificial intelligence to predict which parts of a photo will be of interest to the user.
Meredith Whitaker, co-founder of the AI Now Institute, told the Thomson Reuters Foundation: