Contributions
- Large-scale empirical study (1.2M users, 271K posts over 4 years) of how information consumption patterns differ between science and conspiracy narratives on Facebook
- Identification of two polarized user populations: 255K users focused on scientific content vs. 791K on conspiracy content
- Demonstration that polarized conspiracy users are more committed to diffusion (higher correlation between likes, comments, and shares)
- Empirical evidence that exposure to parodic false claims receives high engagement from conspiracy-focused communities, supporting cognitive closure hypothesis
- Network analysis showing distinct modularity between science and conspiracy communities
Method
Data collection: Public Facebook pages categorized into science (73 total) and conspiracy news (73 total), covering 2010–2014. Dataset includes 271K posts created by 1.2M individuals, analyzed via Facebook Graph API for user interactions (likes, shares, comments).
User polarization: Users classified as polarized if >95% of their likes on one category exceeded total like activity on the other. This threshold identified 255K polarized science users and 791K polarized conspiracy users.
Community detection: Three approaches applied to page networks:
- Pages self-categorized by description
- Multi-level modularity optimization algorithm (Blondel et al.)
- Maximum modularity score algorithm (Clauset et al.)
Interaction analysis: Pearson correlation computed between couples of user actions (likes, comments, shares) on each post type. Consumption patterns characterized via empirical cumulative distribution functions (CCDFs) of likes, comments, and shares across both communities.
False information response: 4,709 intentionally false claims (parodic/satirical content from a deliberately deceptive page) presented to both communities to measure tolerance and engagement.
Results
Consumption patterns: Conspiracy and science posts exhibit similar post lifetimes (heavy-tailed distributions), but distinct user interaction patterns. Conspiracy users show higher correlation coefficients between actions (Likes/Shares: 0.816 vs. 0.218 for science; Comments/Shares: 0.658 vs. 0.522). This indicates conspiracy users are more committed to diffusion—engagement is coordinated across all interaction types.
Polarization characteristics: Polarized conspiracy users (791K, 91.5% of conspiracy sample) remain highly focused within their community: 99.08% of comments on conspiracy posts, 80.86% of likes. Polarized science users show less commitment: 90.29% of comments within their community, 76.79% of likes.
False information response: Exposure to 4,709 false claims (from a parodic page) found 77.92% of likes and 80.86% of comments from users usually consuming conspiracy content. Polarized conspiracy users engaged with intentionally false memes at high rates, demonstrating that community membership—not content veracity—drives engagement.
Network modularity: Modularity analysis confirms distinct separation between conspiracy and science communities across all three detection methods.
Notes
This is a foundational empirical study demonstrating that polarized communities around misinformation show higher engagement and commitment to diffusion than science communities. The key insight—that conspiracy narratives reduce cognitive complexity and appeal to users seeking closure—explains both their spread and users' resistance to correction. The paper's strength is its large scale (1.2M users over 4 years) and direct test of tolerance for false information, avoiding reliance on self-reported beliefs.
The modularity finding is significant: conspiracy and science communities are structurally distinct, making cross-community exposure unlikely without deliberate intervention. This has implications for inoculation and correction strategies—users unlikely to encounter counter-narratives naturally.
A limitation is the Italian Facebook context, which may have specific political or cultural dynamics not generalizable to other regions. Additionally, the paper studies only aggregated likes/comments; it does not analyze actual belief change or long-term persistence of user engagement with corrective information.