Study: Twitter algorithms ‘prefer’ young and slim women

The Twitter feed’s automatic photo-cropping algorithm favors slim women. This conclusion was reached by a researcher at EPFL during a competition to identify algorithmic bias in the social network.

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

— Twitter Engineering (@TwitterEng) August 9, 2021

PhD student in the computer security department Bogdan Kulinich used deepfake techniques to automatically create images of various faces. He then tested the photo-cropping algorithm to see how it would respond to them.

Kulinich found that the image-cropping tool highlights young and slim women, leaving other people and objects out of frame.

“Essentially, the slimmer, younger and more feminine the image, the more likely it is to be preferred,” said Patrick Hall, chief scientist at the consulting firm BNH and one of the judges.

OpenAI security researcher Ariel Herbert-Voss believes that the biases uncovered by the contestants reflect the prejudices of the people who supplied the data used to train the model. She says that a thorough analysis of algorithms can help product teams address issues with their AI models.

The first Twitter algorithmic-bias bounty ran from July 30 to August 6, 2021. The company announced the results on August 8 at the DEF CON AI Village conference.

The winner Bogdan Kulinich received a prize of $3,500 from the company.

In June, OpenAI announced that softened bias and toxicity in the language model GPT-3.

In April, a group of students complained about biases in an AI that monitors students to prevent cheating on exams.

Follow ForkLog’s news on Telegram: ForkLog AI — all the news from the AI world!