Like-minded sources on Facebook are prevalent but not polarizing

The study investigates whether exposure to politically like-minded content on Facebook creates echo chambers that contribute to political polarization and whether reducing such exposure can mitigate polarized attitudes. Motivated by widespread concerns after the 2016 US election and theories of selective exposure and identity reinforcement on social media, the authors note major gaps: limited behavioral measures of exposure within platforms, scarce causal evidence of long-term polarizing effects of like-minded content, and uncertainty about how interventions affect actual cross-cutting exposure and other feed characteristics. The research aims to quantify prevalence of like-minded exposure across US adult Facebook users and to causally estimate the attitudinal and behavioral effects of algorithmically reducing exposure to content from like-minded friends, Pages, and groups during the 2020 US presidential election.

The paper situates its inquiry within research on polarization, echo chambers, and selective exposure. Prior work highlights homophily and ideological segregation across media and social networks, but most behavioral exposure evidence comes from Twitter or limited Facebook samples, and many studies rely on self-reports prone to bias. Survey correlations show associations between like-minded media consumption and polarized attitudes, yet causal evidence is limited and potentially confounded by selection (those with extreme views consume more congenial content). Experimental tests often use brief, simulated content and raise questions about generalizability, durability of effects, and concentration among low-information or news-avoidant individuals. Concerns also exist that reducing like-minded exposure may not increase cross-cutting content due to network composition and could shift feeds toward non-political content or trigger off-platform compensatory information seeking.

The study comprises two components: descriptive measurement of exposure across the full population of US adult Facebook monthly active users and a large-scale field experiment with consenting users.

Population exposure measurement: Using internal Facebook classifiers, the authors estimate political leaning scores (0 left to 1 right) for users; Pages and groups are scored via their audiences. Friends/Pages/groups are classified as liberal (≤0.4), conservative (≥0.6), or neither. Like-minded sources share the user's leaning; cross-cutting sources have opposing leanings; those between thresholds are neither. Exposure in the Facebook Feed from 26 June to 23 September 2020 is analyzed for all content, civic (political) content, and news content using platform logging and classifiers.

Field experiment: 23,377 US adult Facebook users were recruited via in-feed invitations (August–September 2020), consented, and completed at least one post-election survey. Block randomization assigned participants to treatment or control. Treatment (24 September–23 December 2020) downranked all content from sources predicted to share the participant’s political leaning (friends, Pages, groups), using a demotion strength validated to avoid empty feeds. Control feeds were unaltered. The intervention targeted all content, not only political/news, acknowledging the small share of political content and the blurring of social/political identities.

Survey design: Five waves—W1 (31 Aug–12 Sep) and W2 (8–23 Sep) pre-treatment; W3 (9–23 Oct), W4 (4–18 Nov), W5 (9–23 Dec) during treatment. Covariates were measured pre-treatment (W1–W2). Attitudinal outcomes were measured post-election during treatment (W4 and/or W5). On-platform exposure and engagement behavior were continuously logged.

Measures and analysis: Exposure shares to source types and content categories; total views per day; content classified as uncivil, slur words, and misinformation repeat offenders via internal classifiers and external labeled datasets. Engagement metrics: total engagement (passive: clicks, reactions, likes; active: comments, reshares) and engagement rate conditional on exposure. Attitudes include affective polarization, ideological extremity, ideologically consistent issue positions, group evaluations, vote choice/candidate evaluations, party-congenial beliefs about election misconduct/outcomes, views toward the electoral system, and respect for election norms.

Statistical approach: Ordinary least squares (OLS) with robust standard errors following a preregistered analysis plan. Covariates selected via lasso adjustments. Population average treatment effects (PATEs) for survey outcomes use survey weights reflecting adult monthly active Facebook users eligible for recruitment, with false-discovery rate (FDR) adjustments for P values. Sample average treatment effects (SATEs) are also reported. Exploratory equivalence bounds are used to assess precision and rule out effects of specified magnitudes. Heterogeneous treatment effects were preregistered across ideology (direction/extremity), political sophistication, digital literacy, pre-treatment exposure to political content, and pre-treatment like-minded exposure levels.

Descriptive exposure patterns among US adult Facebook users (Q3–Q4 2020; 231M monthly active users):

• Median exposure shares: 50.4% like-minded sources; 14.7% cross-cutting; remainder neither. For civic content, median 55% like-minded; for news content, median 47% like-minded. Civic and news content constitute medians of 6.9% and 6.7% of exposures.
• Echo chambers are not ubiquitous: 20.6% of users receive >75% of exposures from like-minded sources; 30.6% receive 50–75%; 25.6% receive 25–50%; 23.1% receive 0–25%. Only 32.2% have ≥25% cross-cutting exposure.
• Among daily active users: 53% like-minded vs 14% cross-cutting; 21.1% receive >75% like-minded exposures.

Experimental intervention effects (treatment vs control, 24 Sep–23 Dec 2020): Exposure (Fig. 3a; SATE):

• Like-minded exposure share reduced from 53.7% (control) to 36.2% (treatment), −0.77 s.d. (95% CI −0.80, −0.75).
• Cross-cutting exposure increased from 20.7% to 27.9%, +0.43 s.d. (95% CI 0.40, 0.46).
• Neither-category exposure increased from 25.6% to 35.9%, +0.68 s.d. (95% CI 0.65, 0.71).
• Views per day decreased −0.05 s.d. (95% CI −0.08, −0.02). Control average: 267 views/day (143 like-minded); treatment average: 255 views/day (92 like-minded).
• Uncivil content exposure reduced −0.15 s.d. (95% CI −0.18, −0.13); slur-word content −0.04 s.d. (95% CI −0.06, −0.02); misinformation repeat offenders −0.10 s.d. (95% CI −0.13, −0.08). Substantive reductions: slur content from 0.034% to 0.030% (−0.01 views/day), uncivil from 3.15% to 2.81% (−1.24 views/day), misinformation repeat offenders from 0.76% to 0.55% (−0.62 views/day).
• Civic content exposure decreased −0.05 s.d. (95% CI −0.08, −0.03); news content exposure increased +0.05 s.d. (95% CI 0.02, 0.07).

Total engagement (Fig. 3b; SATE):

• Time spent on Facebook: −0.02 s.d. (95% CI −0.050, 0.004), not significant.
• Passive engagement with like-minded sources: −0.24 s.d. (95% CI −0.27, −0.22).
• Active engagement with like-minded sources: −0.12 s.d. (95% CI −0.15, −0.10).
• Passive engagement with cross-cutting sources: +0.11 s.d. (95% CI 0.08, 0.14).
• Active engagement with cross-cutting sources: +0.04 s.d. (95% CI 0.01, 0.07).
• Passive engagement with misinformation repeat offenders: −0.07 s.d. (95% CI −0.10, −0.04); active engagement: −0.02 s.d. (95% CI −0.05, 0.01), not significant.

Engagement rate conditional on exposure (Fig. 3c; SATE):

• Passive engagement with like-minded sources: +0.04 s.d. (95% CI 0.02, 0.06).
• Active engagement with like-minded sources: +0.13 s.d. (95% CI 0.08, 0.17).
• Passive engagement with cross-cutting sources: −0.06 s.d. (95% CI −0.07, −0.04).
• Active engagement with cross-cutting sources: −0.02 s.d. (95% CI −0.04, −0.01).
• Views per active days decreased −0.05 s.d. (95% CI −0.08, −0.02).

Attitudinal outcomes (Fig. 3d; PATE, FDR-adjusted):

• No measurable effects across eight preregistered outcomes. Seven of eight point estimates are within ±0.03 s.d.; exploratory equivalence bounds allow ruling out effects of ±0.12 s.d. on primary outcomes. Affective polarization PATE −0.00 (95% CI −0.029, 0.022). Ideological extremity PATE −0.00 (95% CI −0.036, 0.028). Consistency in vote choice/candidate evaluations PATE 0.06 (95% CI 0.001, 0.110) is a less precise null.
• No evidence of heterogeneous effects across preregistered subgroups (272 estimates) or exploratory age/tenure analyses.

The findings challenge common narratives that pervasive social media echo chambers significantly drive polarization and that algorithmic interventions reducing like-minded content will lessen polarized attitudes. While like-minded sources constitute a majority of Facebook feed exposure, only a small fraction of content is political or news, and extreme echo chamber exposure is uncommon. The intervention effectively reduced like-minded exposure and concomitantly lowered exposure to uncivil content and repeated misinformation sources, yet it did not symmetrically boost cross-cutting exposure, instead increasing neutral sources. Importantly, substantial reductions in like-minded exposure over three months produced precisely estimated null effects on affective polarization, ideological extremity, issue positions, and election-related beliefs. The engagement results suggest users continue to preferentially interact with congenial content when encountered, indicating a persistent behavioral proclivity that algorithmic changes cannot fully counteract. These outcomes imply that social media’s role in shaping political attitudes may be limited relative to broader information diets and that persuasion effects are small and transient, especially in high-salience contexts like a presidential election.

The study offers comprehensive descriptive evidence of like-minded exposure on Facebook and a rigorous causal test of an algorithmic intervention to reduce it. Although congenial sources are prevalent, only a minority of users experience extreme echo chambers, and political/news content constitutes a small share of feeds. Reducing like-minded exposure by roughly one-third decreased uncivil and misinformation-related exposures but did not measurably change a wide range of political attitudes. The results suggest that platform-level algorithmic changes that decrease exposure to like-minded sources are unlikely to be a simple remedy for political polarization. Future research should examine longer-term and cumulative effects of social media use, consider asymmetric effects of increasing versus decreasing like-minded exposure, study more representative or targeted samples (e.g., those occupying echo chambers), directly measure content slant, and conduct replications in other countries to assess generalizability.

The experiment cannot capture prior or cumulative effects of years of social media use; attitude formation may have already occurred before the intervention. The intervention period coincided with a highly contentious presidential election, when persuasion effects tend to be small and short-lived. Political and news content comprise a small share of Facebook feeds and of overall information diets, potentially diluting any treatment impact. Effects of decreasing exposure may not be symmetric with effects of increasing exposure. The experimental sample differs from the overall Facebook population (e.g., heavier users, more like-minded exposure); although survey weights were applied, generalizability may be constrained. The study infers source leanings from audiences and users rather than measuring the ideological slant of the content itself. Cross-platform compensatory behavior was assessed and largely ruled out within specified bounds, but off-platform dynamics remain a consideration.