Can We Curb Algorithmic Bias? An Experiment on a Social Network with Teens

While artificial intelligence (AI) algorithms are seen as a convincing new opportunity to reduce human bias, automated decision-making can also have unintended and unexpected results, especially in the context of online platforms. In the case of science, technology, engineering, and mathematics (STEM) career ads, there is evidence in the academic literature supporting the existence of algorithmic gender biases. We design field tests aiming to reduce these biases in the context of online STEM-related ads aimed at teens. We conduct online ad campaigns targeting young social network users with gender-unspecific, women-specific, and gender-neutral ads. We show that the ad which includes a gender-oriented message helps to reduce algorithmic gender bias but at the cost of lower overall ad display. We also find an unexpected effect of gender-oriented message in the case of minors for whom algorithmic gender bias is enhanced due to more restrictive platform regulation. We find further evidence that while text length reduces ad display significantly, gender-oriented message has a mixed effect depending on gender of the target audience