In what the company claims a first, Twitter is launching a new bounty program rewarding people who can successfully report signs of bias in the company's image cropping algorithm.
Twitter announced in April that it would investigate potential "unintentional harms" caused by its algorithms, beginning with the image-cropping one. In 2018, it started using the algorithm to focus on the most interesting sections of images in previews.
The algorithm in question is used to select which section of an image should be shown in Twitter's mobile apps, and the company announced in May that it had received complaints that the algorithm "didn't serve all people equitably" owing to "gender and race-based biases."
Some Twitter users criticized how the platform handled automated cropping, claiming that the algorithm favors lighter-skinned people in photos.
Twitter wrote in a blog post how it shared its method of identifying bias in its algorithm, emphasizing how it made its code available for others. The social media company says it wants to take this work a step further by allowing and motivating the community to help uncover possible impacts of the algorithm beyond what it has already identified.
Twitter claims this is the "industry's first algorithmic bias bounty competition," with cash awards of up to $3,500 offered.
Rumman Chowdhury, director of Twitter's Machine Learning Ethics, Transparency, and Accountability team, stated in a tweet that the company is holding the contest because it believes that individuals should be rewarded for uncovering these concerns, and that Twitter can't solve these problems on its own.
Ariel Herbert-Voss, Matt Mitchell, Peiter "Mudge" Zatko and Patrick Hall will be among the judges assisting the company in analyzing entries.
Machine learning algorithms, such as the one used by Twitter, rely on massive amounts of data. The resulting algorithm may show prejudice if these data sets are weighted in favor of a given race, gender, or other factor.
The new competition's winner will also be asked to present their work at Twitter's DEF CON AI Village workshop in August.
Successful entries, according to Twitter, will use both quantitative and qualitative methodologies in their approach.
