Reshares on social media amplify political news but do not detectably affect beliefs or opinions

Contributions

• Large-scale randomized controlled trial (N=23,402) measuring causal effects of social media content redistribution on political outcomes.
• Demonstrates that reshares—the core viral mechanism on Facebook—primarily amplify exposure to political news and engagement, not necessarily extreme or false content.
• Shows a disconnect between changes in what users see and changes in what they believe: feed curation shifted content substantially but did not shift political attitudes or polarization.
• Documents that removing reshares reduces exposure to both trustworthy and untrustworthy news sources approximately equally, maintaining a similar information balance.

Method

Researchers partnered with Meta/Facebook to recruit U.S. participants during the 2020 election (September–December 2020). Consenting users (N = 23,402 completed surveys; N = 3,781 had passive web tracking data) were randomly assigned to one of two conditions:

• Control: Normal Facebook feeds with reshares present (28% of feed views were reshared content).
• Treatment ("No Reshares"): Feeds stripped of all reshared content from friends, Groups, or Pages (5.8% of feed views were reshared content).

The researchers pre-registered their hypotheses and measured outcomes via surveys (affective polarization, issue polarization, news knowledge, factual discernment, election beliefs, trust in media, political efficacy, voting behavior) and passive behavioral tracking (clicks, reactions, off-platform news visits). They estimated both population average treatment effects (PATE, weighted to generalize to Facebook users) and sample average treatment effects (SATE, for the consenting sample).

Results

Feed content changes: Removing reshares substantially altered users' exposure:

• Political news content decreased by 60% (6.2% → 2.5% of feed).
• Content from untrustworthy sources fell by 30% (2.6% → 1.8% of feed).
• Relative proportion of friend-posted content dropped 10 percentage points; group content increased 8 points.
• Users saw less both like-minded (51.1% → 53.7% of feed) and cross-cutting (19.7% → 20.7%) ideological content, with neutral content increasing (22.6% → 26.2%).

Behavioral engagement: Removing reshares reduced:

• Time spent on Facebook (64% above baseline vs. 73% for control; p < 0.005).
• Click rate on content (7.5% vs. 8.3%; p = 0.04).
• Reaction rate (p < 0.01).

Political attitudes (primary hypotheses): No significant effects:

• Affective polarization: PATE p = 0.673.
• Issue polarization: PATE p = 0.418.
• No significant changes in perceived polarization, political efficacy, belief in election legitimacy, or support for political violence.

Knowledge (mixed results):

• News knowledge: Sample showed significant decline (SATE p < 0.01); population effect fell short of preregistered significance threshold (PATE p = 0.157) but estimates were similar in magnitude.
• Factual discernment: Modest decline, did not survive FDR correction for multiple comparisons.

Engagement with partisan news: Removing reshares did reduce clicks on partisan news sources (PATE −0.109 SD, p < 0.01), the only secondary outcome to survive multiple-comparison correction.

Notes

Strengths:

• Rigorous RCT design with large N and pre-registration, enabling strong causal inference.
• Multi-method approach: surveys, passive tracking, behavioral logs, and ecological validity (real platform, real election period).
• Transparent reporting of both PATE and SATE; effect-size estimates provided for null findings.
• Clear documentation of what changed (feed content, engagement) and what did not (attitudes, polarization).

Weaknesses and limitations:

• Sample attrition of ~19.5%, though not differential between treatment and control.
• Sample bias: participants more active on platform, older, whiter, wealthier, more college-educated, and more liberal than population-weighted estimates.
• Limited to late September–December 2020 (near election); generalizability to other periods or elections unclear.
• "General equilibrium" effects unmeasured: the analysis cannot predict effects if reshare removal were scaled to all users simultaneously (would change content creator incentives and demand).
• Exploratory analyses (Section S3) suggest a 62% decrease in mainstream news source exposure may explain reduced news knowledge, but these post-hoc analyses should be interpreted with caution.
• The disconnect between large feed changes and null attitude effects is substantively important but mechanistically unexplained.

Methodological notes:

• Study design, measures, and analysis preregistered at OSF.
• FDR-adjusted p-values applied to control for multiple comparisons.
• Treatment effect heterogeneity limited across subgroups; effects were most pronounced among high-activity users but generalizable.
• Weighting scheme accounts for activity level, predicted ideology, network size, and other user characteristics to generalize beyond the sample.