STEM and Teens: An Algorithmic Bias on a Social Media

Abstract

Evidence of algorithmic bias has recently raised questions about their fairness. In a field experiment, we evaluate the algorithm of a social media platform by prompting it to reduce its advertising distribution bias against women. We ran advertising campaigns targeting high schools in France on behalf of an engineering school. One ad was gender-neutral (control) and the other ad had women-specific content (treatment). Women-specific content reduced the advertising distribution gap between women and men. However, the treatment also had a crowding-out effect, such that the ad with women-specific content was displayed fewer times to users overall. The treatment reduced the number of impressions of ad for high schools with more women in science track. The treatment also reduced the advertising distribution gap between low-income and high-income high schools.

Keywords: Advertising, Algorithm, Machine Learning, Field experiment

JEL Classification: J16, I24