Like-minded sources on Facebook are prevalent but not polarizing

The study investigates whether exposure to politically like-minded sources on Facebook (so-called echo chambers) is widespread and whether reducing such exposure can decrease political polarization. The context is ongoing concern that social media algorithms and network homophily reinforce partisan identities and limit exposure to counter-attitudinal information, potentially increasing ideological extremity and affective polarization. The authors aim to fill gaps due to limited platform data access and a lack of large-scale causal evidence by (1) measuring exposure to like-minded sources across the U.S. adult Facebook population and (2) experimentally testing whether reducing such exposure during the 2020 U.S. presidential election affects users’ political attitudes. They articulate three problems: the absence of systematic exposure measures on Facebook, the paucity of causal evidence distinguishing correlation from causation in echo-chamber effects, and uncertainty about whether reducing like-minded exposure increases cross-cutting exposure or produces unintended feed composition changes.

Prior work often relies on surveys or data from platforms like Twitter, which are used by a minority of the public, and has shown that relatively few Americans have highly skewed online news diets. Observational associations link polarized attitudes with consumption of congenial news, but such correlations may be spurious because people with extreme views are more likely to seek like-minded content. Experimental studies frequently use short, simulated exposures and raise questions about external validity, the prevalence and durability of polarizing effects, and whether such effects are concentrated among low-interest users. The authors note the limited behavioral measures of Facebook exposure and measurement error in self-reports. They also discuss the possibility that reducing like-minded exposure might primarily increase neutral rather than cross-cutting content due to the structure of social feeds (mostly followed accounts are non-political and few are cross-cutting), and potential unintended effects such as changes in tone (e.g., civility) or off-platform substitution.

The study has two components:

1. Descriptive measurement of exposure (26 June–23 September 2020): Using internal Facebook classifiers, the political leaning of users, Pages, and groups was estimated on a 0 (left) to 1 (right) scale. Users >0.5 were classified conservative, otherwise liberal. Sources were labeled liberal (≤0.4), conservative (≥0.6), and those in between as neither. Like-minded sources share the user’s leaning; cross-cutting sources do not. Exposure was measured for all U.S. adult monthly active users (231 million accessed monthly in Q3–Q4 2020) as the proportion of Feed content viewed from each source type. Additional classifiers labeled content as civic (political) and news, and identified uncivil content, slur words, and misinformation repeat offenders.

2. Field experiment (24 September–23 December 2020): 23,377 consenting U.S. adult Facebook users were recruited via survey invites atop Feeds (Aug–Sep 2020). Participants were randomized to treatment or control. Treatment downranked all content (not limited to news/politics) from friends, Pages, and groups predicted to share the participant’s political leaning, using the strongest demotion that avoided near-empty Feeds. The goal was solely to reduce exposure to like-minded sources. Five survey waves were administered: W1 (Aug 31–Sep 12) and W2 (Sep 8–23) pre-treatment; W3 (Oct 9–23), W4 (Nov 4–18), W5 (Dec 9–23) during treatment (post-election outcomes measured in W4/W5). On-platform exposure and engagement behaviors were continuously collected.

Outcomes and analysis: Primary exposure outcomes included share of views from like-minded, cross-cutting, and neither sources; civic/news shares; uncivil/slur/misinformation exposures. Engagement was measured as total engagements (passive: clicks, reactions, likes; active: comments, reshares) and engagement rate (probability of engagement conditional on exposure). Attitudinal outcomes included affective polarization, ideological extremity, ideological consistency (issue positions, group evaluations, vote choice/candidate evaluations), partisan-congenial beliefs about election misconduct/outcomes, views toward the electoral system, and respect for election norms. Analyses followed a preregistered plan (OSF: https://osf.io/3sjy2), using OLS with robust standard errors; P-values FDR-adjusted; survey weights used for population average treatment effects on attitudes. Equivalence bounds analyses assessed precision of near-zero effects and heterogeneous effects were explored across preregistered subgroups (ideology, sophistication, digital literacy, pre-treatment political/like-minded exposure).

Descriptive exposure (June–September 2020):

• Median user saw 50.4% of content from like-minded sources vs 14.7% from cross-cutting; remainder from neither.
• For civic content, median 55% of exposures from like-minded; for news, 47% from like-minded.
• Civic and news content comprised small shares of overall Feed (medians 6.9% and 6.7%).
• Distribution: 20.6% of users received >75% of exposures from like-minded sources; 30.6% received 50–75%; 25.6% 25–50%; 23.1% 0–25%. For cross-cutting exposures, only 32.2% of users had ≥25% of exposures from cross-cutting sources.
• Similar patterns among most-active daily users (53% like-minded; 14% cross-cutting; only 21.1% >75% like-minded).

Experimental treatment effects (Sep 24–Dec 23, 2020): Exposure:

• Like-minded share reduced from 53.7% (control) to 36.2% (treatment), Δ = −0.77 s.d. (95% CI: −0.80, −0.75).
• Cross-cutting share increased from 20.7% to 27.9%, Δ = +0.43 s.d. (95% CI: 0.40, 0.46).
• Neither-category share increased from 25.6% to 35.9%, Δ = +0.68 s.d. (95% CI: 0.65, 0.71).
• Total views/day decreased by 0.05 s.d. (95% CI: −0.08, −0.02); time spent not significantly affected (−0.02 s.d., 95% CI: −0.050, 0.004).
• Reduced exposure to: slur-word content (−0.04 s.d., 95% CI: −0.06, −0.02; from 0.034% to 0.030% of exposures; −0.01 views/day), uncivil content (−0.15 s.d., 95% CI: −0.18, −0.13; 3.15% to 2.81%; −1.24 views/day), and misinformation repeat offenders (−0.10 s.d., 95% CI: −0.13, −0.08; 0.76% to 0.55%; −0.62 views/day).
• Civic content exposure slightly decreased (−0.05 s.d., 95% CI: −0.08, −0.03); news content slightly increased (+0.05 s.d., 95% CI: 0.02, 0.07).

Engagement (totals):

• Total engagement with like-minded sources decreased: passive −0.24 s.d. (95% CI: −0.27, −0.22); active −0.12 s.d. (95% CI: −0.15, −0.10).
• Engagement with cross-cutting sources increased: passive +0.11 s.d. (95% CI: 0.08, 0.14); active +0.04 s.d. (95% CI: 0.01, 0.07).
• Engagement with misinformation repeat offenders decreased for passive (−0.07 s.d., 95% CI: −0.10, −0.04); no significant change for active (−0.02 s.d., 95% CI: −0.05, 0.01).

Engagement rate (conditional on exposure):

• With like-minded sources increased: passive +0.04 s.d. (95% CI: 0.02, 0.06); active +0.13 s.d. (95% CI: 0.08, 0.17).
• With cross-cutting sources decreased: passive −0.06 s.d. (95% CI: −0.07, −0.04); active −0.02 s.d. (95% CI: −0.04, −0.01).
• Views per active day decreased slightly (−0.05 s.d., 95% CI: −0.08, −0.02).

Attitudes:

• No measurable effects on eight preregistered outcomes: affective polarization, ideological extremity, ideological consistency (issue positions, group evaluations, vote choice/candidate evaluations), partisan-congenial beliefs about election misconduct/outcomes, views toward the electoral system, respect for election norms. Seven of eight point estimates within ±0.03 s.d.; exploratory equivalence bounds ±0.1 s.d.; minimum detectable effect for affective polarization 0.019 s.d. The eighth (ideologically consistent vote choice/candidate evaluations) was a less precise null (0.056 s.d., equivalence bounds 0.001, 0.111).
• No significant heterogeneous effects across preregistered subgroups (272 tests FDR-adjusted). Exploratory analyses also found no heterogeneity by age or years since joining Facebook.
• No evidence of off-platform substitution toward like-minded media outside Facebook (equivalence ±0.07 s.d.).

The findings show that while like-minded sources constitute a majority of Facebook content exposures for U.S. adult users, extreme echo-chamber exposure is uncommon and explicitly political/news content is a small part of total feeds. The field experiment demonstrates that algorithmic demotion can substantially reduce like-minded exposure and lower uncivil and misinformation-related content, and modestly increase cross-cutting exposure. However, these sizeable on-platform changes did not translate into changes in political attitudes, suggesting that either (a) the political content affected is too small a share of overall information diets, (b) persuasion and attitude change are generally difficult and short-lived, especially during an intense election period, or (c) effects of decreasing like-minded exposure are not symmetric to effects of increasing it. Users also displayed a tendency to engage more intensely with congenial information when encountered, indicating that preferences and identity-driven behaviors may limit the attitudinal impact of algorithmic adjustments. Overall, the results challenge narratives that social media echo chambers are a primary driver of polarization and that reducing like-minded exposure will straightforwardly reduce polarization.

This study provides population-level descriptive evidence and a large-scale field experiment on Facebook during the 2020 U.S. presidential election. It shows that like-minded sources are prevalent in users’ feeds but that political/news content is relatively scarce, and extreme echo-chamber exposure is limited to a minority. An intervention that reduced like-minded exposure by about one-third decreased exposure to uncivil content and misinformation repeat offenders and modestly increased cross-cutting exposure but did not measurably affect polarization-related attitudes across diverse measures and subgroups. The findings suggest that algorithmic changes alone are unlikely to meaningfully shift political attitudes and that identity-driven preferences shape engagement. Future research should examine long-term and cumulative effects, samples more representative of platform populations or oversampling potentially affected subgroups, mechanisms behind why some users experience high like-minded exposure, direct measurement of content slant (not just source leaning), and replication in other countries and contexts.

• Duration and timing: The experiment occurred over several months during a contentious election; effects of persuasion are often small and ephemeral. Longer-term or earlier-life exposure effects were not captured.
• Generalizability: The participant sample differs in some respects from the broader Facebook population; effects in other populations or countries may vary.
• Measurement scope: The study infers political congruity from source leaning rather than directly measuring content slant; political content and news are a small share of feeds.
• Asymmetry of effects: The intervention decreased like-minded exposure for ethical reasons; effects of increasing like-minded exposure may not be symmetric to decreasing it.
• Off-platform information diets: Most people obtain information from multiple sources; on-platform changes may represent a small share of overall exposure.
• Heterogeneity: Despite extensive tests, no subgroup effects were detected; nonetheless, small effects in niche groups cannot be entirely ruled out.
• Platform-specific context: Findings pertain to Facebook’s algorithms and user behavior during 2020 in the U.S.; other platforms or periods could yield different results.