Divergent patterns of engagement with partisan and low-quality news across seven social media platforms

Contributions

• Cross-platform analysis of 10.7 million news-sharing posts, bridging a gap in literature dominated by single-platform studies (primarily X/Twitter and Facebook).
• Distinction between heterogeneous patterns (political lean) and consistent patterns (news quality) in engagement across platforms, challenging the notion of a universal "right-wing advantage."
• Evidence that the quality-engagement penalty holds even on chronologically-ranked platforms without engagement-based algorithms (Mastodon), implicating user preference over algorithmic bias.
• Domain-level quality ratings aggregating six expert sources across 11,520 news domains, accounting for paywalls and non-English content.

Method

The authors collected all posts with links to rated news domains from January 2024 across seven platforms using a combination of platform-specific APIs and search features, yielding 10.96 million posts from unique users. News quality was operationalized as domain-level trustworthiness ratings (0–1 scale) from a "wisdom of experts" approach combining six independent rating sets (professional fact-checkers, journalists, academics) with imputation. Political lean was rated by GPT-4o on a −5 to +5 scale (Democrat/liberal to Republican/conservative).

Engagement (likes + reshares) was log-transformed and predicted via linear regression with user fixed effects, allowing comparison within-user: do the same user's followers engage more with their low-quality vs. high-quality posts? Meta-analyses pooled results across platforms. User fixed effects eliminate confounds from follower counts, posting frequency, and post quality (unobserved). Separate models tested political lean and news quality with and without mutual control.

Results

Political lean and platform composition. Platforms stratified by user composition: BlueSky, Mastodon, and LinkedIn skewed liberal; TruthSocial, Gab, and GETTR skewed conservative. Neutral-domain posts dominated overall but with substantial heterogeneity by platform.

Political engagement is platform-dependent. Conservative-leaning news received higher engagement on right-leaning platforms (TruthSocial, Gab, GETTR); liberal-leaning news received higher engagement on left-leaning platforms (BlueSky, Mastodon, LinkedIn). Meta-analytic slope for political lean without quality control: b = −0.001 (not significant overall; heterogeneity test Q(6) = 153.6, P < 0.0001). Across platforms, the strength of the political-engagement association correlated with platform political lean (r = 0.78, P = 0.037), consistent with an echo-chamber effect.

News quality engenders consistent penalties across all platforms. A given user's posts with lower-quality news links received significantly more engagement than their posts with higher-quality links across all seven platforms (meta-analytic effect: b = −0.011 to −0.013, P < 0.001 depending on controls). The negative effect of quality held even after removing paywalled outlets (a potential confound), excluding non-English posts, and analyzing reshares and likes separately. On X/Twitter alone, where impressions (views) data were available, lower-quality links received more impressions than higher-quality links, but the effect was smaller (~1/5 the size of the engagement effect), suggesting algorithmic or network structural factors partially moderate the relationship but do not fully explain it.

Quantifying the magnitude: posts with the lowest-quality domains received ~7% more engagement than posts with the highest-quality domains. When binned by quality quartile, this effect was consistent across platforms (SI Appendix, Fig. S17). The effect was robust to outlier domains; major high-quality outlets (New York Times, Wall Street Journal, Washington Post, Reuters, USA Today) showed particularly low engagement on a per-post basis, but removing them did not eliminate the effect.

Mastodon (algorithmic-free) showed the strongest quality penalty. The pattern was particularly pronounced on Mastodon, which ranks posts purely chronologically with no engagement-based algorithm, ruling out algorithmic preference for low-quality content as the sole explanation.

Notes

Strengths:

• Uniquely broad cross-platform scope with rigorous within-user comparisons (user fixed effects eliminate user-level confounds).
• Domain-level quality operationalization avoids the selective sampling problem of fact-checked claims, which tend to be controversial.
• Evidence against algorithmic bias (Mastodon result) is particularly valuable for policy discussions.
• Reconciles two seemingly contradictory findings: high-quality news dominates in volume but underperforms per-user.

Limitations:

• January 2024 snapshot; platforms evolve rapidly (particularly post-Elon X takeover and BlueSky growth).
• Excludes major platforms (Facebook, Instagram, TikTok) due to data-access constraints, limiting generalizability to visual/short-form media.
• Observational; cannot experimentally isolate causal drivers of engagement (though user fixed effects mitigate confounding from user characteristics).
• Political lean via GPT-4o: model-based ratings may diverge from human perception; authors partially validated against crowdsourced and expert baselines.

Open questions:

• Does the quality-engagement penalty reflect genuine user preference for novel/emotional/moralizing content, or are there unmeasured post-level confounds (tone, topic, recency, network structure)?
• How do these patterns evolve as platforms stabilize post-migration waves (e.g., BlueSky post-2024 US election)?
• Do private messaging and group-based platforms (Signal, Telegram, WhatsApp) exhibit similar penalties? (Authors explored Telegram in SI but noted signal is noisier.)