Contributions
- Large-scale randomized controlled experiment testing causal effects of feed algorithms on political attitudes and behavior during a real election campaign
- Demonstrates that removing algorithmic ranking reduces exposure to uncivil and divisive content
- Shows that exposure to platform-algorithmically-ranked content does not significantly drive changes in affective polarization, political knowledge, or voting behavior
- Provides empirical test of "filter bubble" and "algorithmic amplification" theories in a controlled setting
Method
The study recruited consenting Facebook and Instagram users in the United States and randomly assigned them to conditions during the 2020 presidential election campaign:
- Chronological Feed: Users saw posts in reverse-chronological order (newest first), as the authors initially described Facebook's past algorithm behavior
- Algorithmic Feed: Users saw the default feed ranked by engagement and relevance (control condition)
Participants were randomized starting 24 September 2020 through 23 December 2020. The researchers measured:
- Time spent on platform and engagement metrics
- Exposure to different content types (political, uncivil, partisan)
- Users' political attitudes (polarization, knowledge, beliefs)
- Self-reported political behavior (voting, participation)
Outcome measures were collected via surveys and platform analytics. To isolate the feed algorithm's role, the study used a diverse sample of nearly 3 million demographically representative US social media users.
Results
Exposure and engagement:
- Chronological feeds reduced time spent on Facebook by ~37% and on Instagram by ~18% relative to algorithmic feeds
- Chronological feeds decreased exposure to uncivil content on both platforms (e.g., ~31.7% reduction on Facebook, ~44.8% reduction on Instagram for certain content categories)
- Users saw less content from "cross-cutting" opposing political sources and more from ideologically aligned sources in both conditions, but differences between conditions were smaller on Instagram
Political attitudes and knowledge:
- No statistically significant differences in affective polarization, issue polarization, or political knowledge between chronological and algorithmic feed conditions
- No detectable effects on self-reported political participation, voting behavior, or offline political engagement
- Secondary hypotheses testing political beliefs, trust in institutions, and attitudes about election integrity also showed no significant effects
Content categories:
- Chronological feeds increased exposure to content from friends and Pages, and decreased exposure to content from Groups
- Political content was actually slightly higher in the Chronological Feed in some analyses, contrary to expectations
The effects on all outcomes were consistent across demographic subgroups and across both platforms.
Notes
The paper represents a significant empirical test of causal claims about algorithmic ranking and political polarization, using a real-world experimental design during an actual election. The finding that removing algorithmic ranking does not substantially reduce political polarization is important for policy debates about platform regulation and algorithmic transparency.
Two errata were published after the original 2023 publication: in December 2024 and March 2026, addressing issues with how Instagram carousel posts were counted in exposure metrics for untrustworthy sources. These corrections primarily affected Instagram-specific results on exposure to untrusworthy content; Facebook results and overall findings on polarization and political participation remained unaffected.