By 
Unlike other companies, Twitter has taken an open approach to their algorithm bias research. This is in contrast to the responses from other tech companies when confronted with similar problems. Twitter is now combating these problems by opening their systems up to external scrutiny. This includes their first ever open competition held to find algorithmic bias in its photo-cropping system.
Twitter Algorithm Bias: Preferring young, beautiful, and light-skinned faces
In March, the company disabled automatic photo-cropping after experiments suggested it favored white faces over black faces. Since then they have made changes and held this competition which has confirmed these earlier findings.
Drumroll please! Our first-ever algorithmic bias bounty at #defcon wrapped this weekend, and the winners were announced at yesterday’s #defcon @aivillage_dc workshop. ICYMI, you can watch a recording of the presentation here https://t.co/0cCwSAzSAq
Start off your week with tech news recap and trending conversations in your inbox
— Twitter Engineering (@TwitterEng) August 9, 2021
The top entry shows that Twitter’s cropping algorithm favors faces that are “slim, young. Specifically those of light or warm skin color and smooth skin texture, and with stereotypically feminine facial traits.
The second and third-placed entries show that the system is biased against people with white or grey hair. This suggests age discrimination among many other issues. Looking past the issues they clearly aim to solve, the competition is a good way for companies to find out problems they may not see themselves.
The winner, Bogdan Kulynych, a graduate student at EPFL Switzerland, got a whooping KES 350,000 cash prize. He used an AI program called StyleGAN2. It helped him generate a large number of realistic faces which he varied by skin color, feminine versus masculine facial features, and slimness. He then fed these variants into Twitter’s photo-cropping algorithm to find which it preferred.
The third-place entry, by Roya Pakzad, founder of tech advocacy organization Taraaz, reveals that the biases extend to written features too. Pakzad’s work looks at memes using English and Arabic script. In the end, it shows that the algorithm regularly crops the image to highlight the English text.
“If you’re not finding your bugs, or bug bounties aren’t finding your bugs, then who is finding your bugs? Because you definitely have bugs.” – Patrick Hall, a judge in Twitter’s competition and an AI researcher working in algorithmic discrimination.
Comments