Job market paper

The Economics of Content Moderation: Theory and Experimental Evidence from Hate Speech on Twitter [paper]

Social media platforms ban users and remove posts to moderate their content. This "speech policing" remains controversial because little is known about its consequences and the costs and benefits for different individuals. I conduct two field experiments on Twitter to examine the effect of moderating hate speech on user behavior and welfare. Randomly reporting posts for violating the rules against hateful conduct increases the likelihood that Twitter removes them. Reporting does not affect the activity on the platform of the posts' authors or their likelihood of reposting hate, but it does increase the activity of those attacked by the posts. These results are consistent with a model in which content moderation is a quality decision for platforms that increases user engagement and hence advertising revenue. The second experiment shows that changing users' perceived content removal does not change their willingness to pause using social media, a measure of consumer surplus. My results imply that content moderation does not necessarily moderate users, but it marginally increases advertising revenue. It can be consistent with both profit- and welfare-maximization if out-of-platform externalities are small.



Publications

Cash: A Blessing or a Curse? [paper]

With Fernando Alvarez, David Argente, and Francesco Lippi

Journal of Monetary Economics, Forthcoming


The Fountainhead: Analyzing the Impact of Intraday Liquidity on Payment Behavior [paper (gated)]

With Aldo Marini and Javier Pérez Estrada

Journal of Financial Market Infrastructures. 3(2): 41-75 (2014)




Working papers

Estimating Repugnance toward Price Gouging with Incentivized Consumer Reports [paper]

With Justin Holz and Eduardo Laguna Müggenburg

R&R at the American Economic Journal: Applied Economics

Thirty-four states prohibit price increases during emergencies and many individuals take costly actions to report violators. We use an experiment to measure the willingness to pay to report sellers who increase prices of personal protective equipment. Over 75% of subjects pay to report even if others are willing to purchase at those prices. The willingness to pay is polarized and increases with price. We argue that reports contain information about repugnance—a desire to prevent third-party transactions at increased prices. The mechanism driving reports varies by good: we find a distaste for profits for hand sanitizers but not for face masks.



Work in progress


A Causal Measure of Hate Speech

A major challenge in the algorithmic detection of hate speech on social media is the low external validity of state-of-the-art measures when applied to different datasets from the ones they are trained on. One explanation is that existing machine-learning algorithms are trained to detect correlations, not causal relationships, between textual features and hate speech labels. I address this by exogenously varying the textual features that algorithms use to predict the labels. I randomly delete words from social media posts and ask crowdworkers to label multiple versions of each post with different words, which gives exogenous variation on observable features and allows controlling for unobservables at the post level. I train different machine-learning algorithms using this procedure and compare the performance gains relative to the standard approach.


Competition and Content Moderation

Would more competition in Big Tech solve the problem of hate speech? This paper introduces a model of perfect competition between interoperable social media platforms that differentiate themselves by their size, their content moderation policy, and their prevalence of hate speech. If hate speech is contagious and harms non-hateful users, there can only be a separating or “echo-chamber” equilibrium, in which some platforms cater exclusively to haters and some to non-haters, without having to moderate. This separation occurs because platforms do not internalize network effects that bring users together. Perfect competition might increase the total amount of hate speech relative to a monopoly if this type of content is sufficiently contagious or if network effects are small.