Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H., Brown, T., Crespo-Tenorio, A., Dimmery, D., Freelon, D., Gentzkow, M., González-Bailón, S., Kennedy, E., Kim, Y. M., Lazer, D., Moehler, D., Nyhan, B., Rivera, C. V., Settle, J., Thomas, D. R., Thorson, E., Tromble, R., Wilkins, A., Wojcieszak, M., Xiong, B., de Jonge, C. K., Franco, A., Mason, W., Stroud, N. J., & Tucker, J. A. (2023). Reshares on social media amplify political news but do not detectably affect beliefs or opinions. Science, 381, 404–408. https://doi.org/10.1126/science.add8424
Summary
This paper reports one of the large-scale randomized experiments conducted through the 2020 US Facebook and Instagram Election Study (FIES), testing what happens when reshared content is stripped from users’ Facebook feeds for roughly three months during the presidential election. The intervention worked mechanically as intended—it sharply cut exposure to reshared posts, political news, and untrustworthy-source content, and reduced time on platform and engagement. Yet despite substantially reshaping the information environment, suppressing reshares produced no detectable change in affective or issue polarization or in nearly any political attitude or offline behavior, with only mixed, uncertain evidence for reduced news knowledge. The core argument is that a widely proposed platform reform can meaningfully alter what people see without measurably altering what they think.
Key Contributions
- Rigorous, large-scale causal (not correlational) evidence on how a specific platform feature—resharing—shapes political outcomes during a real election.
- Demonstrates that removing reshares changes content exposure and on-platform behavior but has limited detectable downstream attitudinal impact.
- Quantifies the composition of reshared content, showing it amplifies both mainstream political news and untrustworthy sources, while most reshared content does not go viral.
- Provides precise null/near-null estimates powered to rule out even moderately sized effects.
- Offers a model of preregistered, transparent industry–academic collaboration with archived design, data, and code.
Methods
A preregistered randomized controlled experiment embedded in the 2020 FIES, with recruitment invitations shown to ~14.6 million users and 193,880 consenting. Participants were assigned to a control condition (unchanged feed) or a “No Reshares” condition (feeds excluding reshared content from friends, Groups, and Pages) from 24 September to 23 December 2020. Analytic samples were N=23,402 (survey) and N=3,781 (web-tracking). Outcomes were drawn from five surveys, platform behavioral logs, and web-visit tracking; content was classified for political relevance, news, untrustworthy sources, incivility, slurs, and ideological alignment. The primary estimand was a population average treatment effect (PATE) weighted by ideology, friend count, pages followed, and activity, with unweighted SATE also reported, estimated via covariate-adjusted OLS with HC2 robust errors and sharpened FDR correction for multiple comparisons.
Findings
- Reshared content fell from 28% of views (control) to 5.8%; political news dropped from 6.2% to 2.5% and untrustworthy-source content from 2.6% to 1.8%.
- Uncivil and moderate/mixed-source content rose slightly; both like-minded and cross-cutting content fell modestly; slur exposure was unchanged.
- Treatment reduced daily time on Facebook, click rate (8.3% → 7.5%), and reaction rate, but did not change users’ own reshares, likes, or comments; friend content fell ~10 points while Groups (+8) and Pages (+2) rose.
- No significant effects on affective or issue polarization (FDR-adjusted p>0.8), media trust, institutional confidence, election legitimacy, factual accuracy perceptions, or support for political violence.
- News knowledge fell in-sample (SATE −0.069 SD) but the PATE (−0.053 SD) fell short of significance after FDR adjustment; election knowledge was unaffected.
- The one robust behavioral effect: reduced clicks on partisan news (PATE −0.109 SD); off-platform political news visits were unaffected.
- A 62% drop in mainstream news exposure emerged as a plausible driver of the suggestive knowledge effect.
Connections
This is part of the same 2020 FIES collaboration as Guess2023-ur and Nyhan2023-gb, and its null attitudinal findings echo the deactivation-experiment tradition of Allcott2019-gn and Bail2018-fk, reinforcing a recurring pattern in which large exposure changes do not translate into measurable belief or polarization shifts. Its analysis of virality and the amplification of untrustworthy content speaks to work on how resharing spreads news and misinformation, including Vosoughi2018-at, Gonzalez-Bailon2023-uy, and DeVerna2025-dl. The affective polarization outcomes connect to Iyengar2019-jj, while the characterization of misinformation’s limited reach and impact relates to Allen2020-nj and Grinberg2019-ua.