Doctoral Dissertation Research in Economics: Algorithmic Bias and Dynamics of Hate Speech on Social Media

Brown University, Providence RI

Investigators

Abstract

This award supports research on social media algorithms that lead to hate speech and polarization on these platforms. It is thought that the sharp increase in hate speech around the world is partly caused by social media algorithms that amplifies such hate speech. However, researchers have not been able to test whether this is the case or not for lack of appropriate data. Working with one of the largest social media platform in the world, the researchers will study the effects of algorithmic recommendations on increased hate speech and the cumulative effects of past exposure to hateful content prompted by algorithms on current user engagement with such social media posts. The use of experimental methods will allow the researchers to disentangle the effects of user preferences for hate speech from the effects of algorithmic amplification of hate speech. The results of this research will provide important inputs into policies to reduce hateful speech on social media platforms and thus establish the US as a global leader in reducing hate speech on social media. Algorithmic recommendations are widely used to tailor content to users’ preferences on social media platforms leading to amplification of some messages, yet little is known about the causal effect of these algorithms on hateful speech. The algorithms expose different users to specific kinds of content based on their innate preferences over social content. These preferences are not observed by the researcher but are learned by the algorithm over time. This project investigates the influence of algorithmic recommendation systems on the amplification of engagement with hate speech. To accomplish this, the researchers will conduct a large-scale RCT in collaboration with one of the largest social media platforms in the world. In this experiment, content recommendations will be switched off for a random set of users. As a result, a large number of users will be exposed to content that is chosen randomly from the entire corpus of posts. The researchers hypothesize that the effect of past exposure on sharing of current content will cause algorithmic customization to be more polarizing than it would be in the absence of such dynamic effects. The results of this research will provide important inputs into policies to reduce hateful speech on social media platforms and thus establish the US as a global leader in reducing hate speech on social media. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.