DEF CON Contest Reveals Twitter's Image Algorithm Favors Youth and Fair Skin
A recent algorithmic bias bounty at DEF CON shows that Twitter's automated image cropping tool favors younger, thinner, and lighter-skinned individuals while actively cropping out wheelchair users.
A DEF CON algorithmic bias bounty contest reveals that Twitter's automated image-cropping algorithm favors individuals who appear younger, thinner, and have lighter skin tones. The saliency algorithm, designed to focus on the most engaging parts of an image, shows a clear preference for able-bodied people and those with stereotypically feminine facial traits as users scroll through their feeds.
This competition builds on previous discoveries from last year where netizens noticed the tool preferred women over men and lighter skin over darker skin. To uncover additional flaws, Twitter's ML Ethics, Transparency, and Accountability (META) team sponsors this ongoing research, which now exposes further discrimination against Arabic text and a tendency to completely crop out people who use wheelchairs.
Contest winner Bogdan Kulynych earns $3,500 for generating fake faces to prove the algorithm ranks slim, young, and smooth-skinned individuals highest in saliency scores. Meanwhile, the second-place team from Halt AI identifies dangerous spatial biases, showing the tool frequently crops out wheelchair users who sit lower in photographs compared to standing individuals.