Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Prosocial Ranking Challenge: Reducing Polarization on Social Media without Sacrificing Engagement

Jonathan Stray, Ian Baker, George Beknazar-Yuzbashev, Ceren Budak, Julia Kamin, Kylan Rutherford, Mateusz Stalinski, Tin Acosta, Chris Bail, Michael Bernstein, Mark Brandt, Amy Bruckman, Anshuman Chhabra, Soham De, Kayla Duskin, Sara Fish, Beth Goldberg, Andy Guess, Dylan Hadfield-Menell, Muhammed Haroon, Safwan Hossain, Michael Inzlicht, Gauri Jain, Yanchen Jiang, Alexander P. Landry, Yph Lelkes, Hongfan Lu, Peter Mason, Jennifer McCoy, Smitha Milli, Paul Resnick, Emily Saltz, Martin Saveski, Lisa Schirch, Max Spohn, Siddarth Srinivasan, Alexis Tatore, Luke Thorburn, Joshua A. Tucker, Robb Willer, Magdalena Wojcieszak, Manuel Wüthrich, Sylvan Zheng

Abstract

We report the first direct comparisons of multiple alternative social media algorithms on multiple platforms on outcomes of societal interest. We used a browser extension to modify which posts were shown to desktop social media users, randomly assigning 9,386 users to a control group or one of five alternative ranking algorithms which simultaneously altered content across three platforms for six months during the US 2024 presidential election. This reduced our preregistered index of affective polarization by an average of 0.03 standard deviations (p < 0.05), including a 1.5 degree decrease in differences between the 100 point inparty and outparty feeling thermometers. We saw reductions in active use time for Facebook (-0.37 min/day) and Reddit (-0.2 min/day), but an increase of 0.32 min/day (p < 0.01) for X/Twitter. We saw an increase in reports of negative social media experiences but found no effects on well-being, news knowledge, outgroup empathy, perceptions of and support for partisan violence. This implies that bridging content can improve some societal outcomes without necessarily conflicting with the engagement-driven business model of social media.

The Prosocial Ranking Challenge: Reducing Polarization on Social Media without Sacrificing Engagement

Abstract

We report the first direct comparisons of multiple alternative social media algorithms on multiple platforms on outcomes of societal interest. We used a browser extension to modify which posts were shown to desktop social media users, randomly assigning 9,386 users to a control group or one of five alternative ranking algorithms which simultaneously altered content across three platforms for six months during the US 2024 presidential election. This reduced our preregistered index of affective polarization by an average of 0.03 standard deviations (p < 0.05), including a 1.5 degree decrease in differences between the 100 point inparty and outparty feeling thermometers. We saw reductions in active use time for Facebook (-0.37 min/day) and Reddit (-0.2 min/day), but an increase of 0.32 min/day (p < 0.01) for X/Twitter. We saw an increase in reports of negative social media experiences but found no effects on well-being, news knowledge, outgroup empathy, perceptions of and support for partisan violence. This implies that bridging content can improve some societal outcomes without necessarily conflicting with the engagement-driven business model of social media.
Paper Structure (96 sections, 3 equations, 13 figures, 22 tables)

This paper contains 96 sections, 3 equations, 13 figures, 22 tables.

Table of Contents

  1. Supplementary Materials
  2. Preregistration and Deviations
  3. Outcome Measures
  4. Sample Size
  5. Interventions
  6. Results on All Outcomes
  7. Survey Outcomes
  8. Preregistered Indexes
  9. Mental Health
  10. Support for Partisan Violence
  11. Meta-perceptions on Support for Partisan Violence
  12. Affective Polarization
  13. Intergroup empathy
  14. Political Knowledge
  15. Meaningful Connections and Experience
  16. ...and 81 more sections

Figures (13)

  • Figure 1: Design and timeline of our experiment. Top: Participants installed a custom browser extension that intercepted platform API calls to extract the top 30-50 items of content served by each of three platforms, rerouting them to our ranking server for real-time reordering, deletion, and addition. Bottom: Experiment timeline, with key dates and survey waves. Wave W1 was the recruitment period, with participants completing a baseline survey on signup. W2 (midline) was pre-election while W3 (endline) was pre-inauguration. Diverse Approval, Challenging Stereotypes and Add News interventions started later than Uprank Bridging and Uprank Bridging, Downrank Toxic due to technical delays.
  • Figure 2: Effects on polarization and time on platform. Left: treatment effects on affective polarization index, and inparty and outparty components of this index, pooled cross all arms and per-arm. All survey outcomes were rescaled to a 0..1 range. Right: treatment effects on active time per platform, pooled across all arms and per-arm. "Bridging Pooled" is a preregistered aggregation of Uprank Bridging, Downrank Toxic and Uprank Bridging, not a separate treatment condition. (* $p<0.1$, ** $p<0.05$, *** $p<0.01$)
  • Figure S1: How in-feed surveys appeared to participants on X/Twitter. Similar UI was used for Facebook and Reddit.
  • Figure S2: Design of the Perspective ranking algorithms. The Uprank Bridging and Uprank Bridging, Downrank Toxic algorithms are identical except for different classifier weights, as listed in tables \ref{['tab:perspective_uprank_weights']} and \ref{['tab:perspective_uprank_downrank_weights']}.
  • Figure S3: Timeline of post rank change, posts added, and posts deleted for Uprank Bridging, 5 day moving average. Normalized rank change is the average numerical rank change for each item, divided by the length of the item slate.
  • ...and 8 more figures