on the 2026–2027 academic job market

CV (PDF)

I run field experiments on social media platforms to study how toxic content, advertising, and ranking shape engagement, political behavior, welfare, and mental health.

Applied microeconomics · Political economics · Behavioral economics

Job Market Paper

Social Media Toxicity and Mental Health

What happens to mental health when toxic content is removed from social media?

We randomize a cross-platform browser extension that hides toxic posts and comments for six weeks (N = 664). Contrary to intuition, filtering toxicity worsens mental health: a 0.15σ rise in a GAD-7/PHQ-8 symptom index, with loneliness and a diminished relative moral self-view as candidate channels.

with Francesco Capozza, Kylan Rutherford, and Mateusz Stalinski

  • presented at the 2026 NBER Summer Institute (Digital Economics and AI)
paper (PDF)
Abstract

We study whether content-level moderation that reduces exposure to social media toxicity, holding platform access constant, affects mental health. We deploy, for randomly assigned users, a cross-platform browser extension that hides posts and comments exceeding a prespecified toxicity threshold, and follow N = 664 desktop users over a six-week intervention across two waves – one during the 2024 U.S. election campaign and one outside of it. Contrary to basic intuition, filtering toxicity worsens mental health: treatment raises a standardized index of GAD-7 and PHQ-8 symptoms by 0.15σ, and the share of respondents above a moderate-symptom screening cutoff rises by about 12 pp. A QALY-based calibration implies a welfare cost of about $158 per treated participant. Heightened loneliness and a diminished relative moral self-view emerge as potential channels. We conclude that curbing toxic content can impose unintended psychological costs, even where it yields other social benefits.

Publications and Conditionally Accepted Papers

  • Toxic Content and User Engagement on Social Media: Evidence from a Field Experiment American Economic Review, conditionally accepted with Rafael Jiménez-Durán, Mateusz Stalinski, and Jesse McCrosky

    Randomly hiding toxic content for six weeks on Facebook, Twitter, and YouTube reduces time spent, ad impressions, and the toxicity of users’ own posts. A survey experiment provides evidence that toxicity triggers curiosity. Together, the results imply a trade-off for platforms between curbing toxicity and engagement.

    paper
    Abstract

    Most social media users have encountered harassment online, but there is scarce evidence of how this type of toxic content impacts engagement. In a pre-registered browser extension field experiment, we randomly hid toxic content for six weeks on Facebook, Twitter, and YouTube. Lowering exposure to toxicity reduced advertising impressions, time spent, and other measures of engagement, and reduced the toxicity of user-generated content. A survey experiment provides evidence that toxicity triggers curiosity and that engagement and welfare are not necessarily aligned. Taken together, our results suggest that platforms face a trade-off between curbing toxicity and increasing engagement.

  • To the Depths of the Sunk Cost: Experiments Revisiting the Elusive Effect Journal of Economic Behavior and Organization, 241 (2026): 107352 with Sota Ichiba and Mateusz Stalinski

    Randomizing the price of a ticket to a real-effort task produces a moderate sunk cost effect: 0.12 SD, or 1.3 minutes more play time. Randomizing how long a YouTube pre-roll ad stays unskippable produces no statistically significant overall change in viewing, as more users leave before the video starts.

    paper
    Abstract

    Despite being often discussed both in practice and academic circles, the sunk cost effect remains empirically elusive. Our model based on reference point dependence suggests that the traditional way of testing it—by assigning discounts—may not produce the desired effect. Motivated by this, we evaluate it across the gain-loss divide in two pre-registered experiments. In an online study with N=1,806, we randomize the price (low, medium, or high) of a ticket to enter a real-effort task, and observe its effect on play time. Our intervention, which varies the sunk cost by $2 for a 14-minute task, results in a moderate sunk cost effect (0.12 SD or 1.3 minutes). We further explore the economic applications of the effect in a field experiment on YouTube with N = 11,328 videos in which we randomize whether the time until a pre-video ad becomes skippable is shortened (0 s), default (5 s), or extended (10 s). The intervention has an overall insignificant effect on video engagement. This is driven by a sizable negative effect on the extensive margin, a channel which is not present in the online study. Specifically, more users leave before the video starts in the extended treatment (5.2 pp or 28% more relative to the shortened treatment). Taking the results of both studies together, we offer evidence of the sunk cost effect in a controlled environment, but its application in policy settings may prove challenging.

  • A Model of Harmful yet Engaging Content on Social Media AEA Papers and Proceedings, 114 (2024): 678–83 with Rafael Jiménez-Durán and Mateusz Stalinski

    A simple model shows that ad-funded platforms can profit from harmful content when it complements time spent. Under a subscription model the incentive disappears, absent network effects, since harm lowers willingness to pay. Rising engagement need not signal rising welfare.

    paper
    Abstract

    Why do social media users spend so much time consuming content that seemingly harms them? We build a simple model to argue that advertising-driven platforms can find it profitable to display content that harms users when it is complementary to their time spent on the platform. These incentives disappear, absent network effects, in the case of a subscription-based business model because harmful content reduces the willingness to pay for the platform. Our results warn against interpreting increases in engagement on social media as welfare increases.

  • Do Social Media Ads Matter for Political Behavior? A Field Experiment Journal of Public Economics, 214 (2022): 104735 with Mateusz Stalinski

    A preregistered experiment randomly hid political ads on Facebook before the 2020 U.S. elections. The reported ITT of political-ad exposure on turnout is 2.3 pp but imprecise (95% CI: −2.8 to 7.4), and may mask heterogeneity: Democrats slightly more motivated to vote, Republicans substantially less.

    paper
    Abstract

    We exploit Facebook’s introduction of a filter hiding ads from the feed as a unique opportunity to study the effects of online ads on political behavior. In a pre-registered experiment, we randomly assigned participants to hide political ads (treatment) or alcohol ads (control) for several weeks preceding the 2020 US elections. We report an insignificant intent-to-treat effect of political ads on turnout (2.3 pp.), but we cannot rule out a sizable positive effect, with 95% confidence interval of [-2.8, 7.4]. The result may mask important heterogeneity, with political ads making Democrats slightly more motivated to vote and Republicans – substantially less. We explore the reasons for this effect, such as natural variation in ad content: the majority of Facebook ads on users’ feeds skewed Democratic. Lastly, the effect on measures of affective polarization and informedness was negligible.

Working Papers

  • Social Media Advertising Loads as Prices Working paper, 2026 with Rafael Jiménez-Durán, Andrey Simonov, and Mateusz Stalinski
    • presented (by a coauthor) at the 2026 NBER Summer Institute (Digital Economics and AI / IO joint session)

    A Facebook field experiment (N = 1,810) shows users are highly insensitive to ad loads: hiding 82% of ads raises time on the platform by only 6%, while removing targeting sharply cuts ad clicks and long-run engagement. Counterfactual simulations indicate ad funding matches or exceeds subscription profits and yields higher consumer surplus.

    paper
    Abstract

    Most digital platforms are funded through advertising rather than direct payments. Why? We argue that three main factors could explain this prevalence: users are more sensitive to monetary prices than to ad loads, microtargeting improves the match quality between users and ads, and platforms can personalize ad loads and thus price discriminate. We conduct a field experiment on Facebook with 1,810 users who install a browser extension that (i) hides nearly all ads or (ii) replaces microtargeted ads with untargeted ones. We find that hiding 82% of ads increases time on the platform by only 6%, showing that users are highly insensitive to ad loads. Removing targeting sharply reduces ad clicks and long-run engagement, indicating that targeting increases the match quality between users and ads. Finally, two-thirds of ad-load variation occurs across users, consistent with ad-load discrimination. Counterfactual simulations indicate that an ad-funded model performs at least as well as a subscription model in terms of profits and delivers higher consumer surplus. The key mechanism is that users are much less sensitive to ad loads than to monetary prices, making advertising a relatively efficient revenue source.

  • The Prosocial Ranking Challenge: Reducing Polarization on Social Media without Sacrificing Engagement Working paper, 2026 with Jonathan Stray, Ian Baker, Mateusz Stalinski, and others

    Over six months around the 2024 U.S. election, 9,386 users were randomly assigned to a control arm or one of five feed-ranking algorithms across three platforms. The pooled DiD estimate on affective polarization is −0.03 SD (p < 0.05); time spent falls on Facebook and Reddit but rises on X.

    paper
    Abstract

    We report the first direct comparisons of multiple alternative social media algorithms on multiple platforms on outcomes of societal interest. We used a browser extension to modify which posts were shown to desktop social media users, randomly assigning 9,386 users to a control group or one of five alternative ranking algorithms which simultaneously altered content across three platforms for six months during the US 2024 presidential election. This reduced our preregistered index of affective polarization by an average of 0.03 standard deviations (p < 0.05), including a 1.5 degree decrease in differences between the 100 point inparty and outparty feeling thermometers. We saw reductions in active use time for Facebook (-0.37 min/day) and Reddit (-0.2 min/day), but an increase of 0.32 min/day (p < 0.01) for X/Twitter. We saw an increase in reports of negative social media experiences but found no effects on well-being, news knowledge, outgroup empathy, perceptions of and support for partisan violence. This implies that bridging content can improve some societal outcomes without necessarily conflicting with the engagement-driven business model of social media.

George Beknazar-Yuzbashev