Bouchaud, P., & Ramaciotti, P. (2026). Community Notes undermoderate polarizing content by design creating risks in electoral processes. Science Advances, 12. https://doi.org/10.1126/sciadv.aee6932
Summary
This paper offers a large-scale empirical audit of X’s Community Notes, arguing that its bridging-based moderation algorithm — which surfaces notes attracting cross-partisan agreement along an inferred latent ideological dimension — is structurally biased against moderating polarizing content. Using the full corpus of ~1.9 million notes and ~135 million ratings through March 2025, cross-referenced with ideological scaling data from 13 countries, the authors show that the very content most in need of correction is the least likely to receive it. They frame this as a design feature, not a bug, with material consequences for electoral integrity as the system scales globally.
Key Contributions
- First large-scale, cross-national empirical audit of Community Notes spanning 13 countries with heterogeneous polarization structures.
- Identifies a structural limitation of bridging-based moderation: cross-partisan agreement requirements systematically exclude polarizing claims from correction.
- Connects platform-governance debates about crowdsourced moderation to electoral integrity risks, challenging the framing of bridging algorithms as neutral alternatives to top-down moderation.
Methods
- Analysis of the complete Community Notes dataset through March 2025 (~1.9M notes, ~135M ratings).
- Cross-referencing raters and notes with country-level ideological scaling datasets covering 13 countries.
- Comparative evaluation of algorithmic behavior across differing national polarization contexts.
Findings
- Notes on polarizing content are systematically less likely to be surfaced as “helpful” because they fail to attract cross-partisan ratings.
- Coverage and effectiveness of moderation vary substantially across countries as a function of their polarization structure.
- Undermoderation of polarizing material follows directly from the bridging design rather than from implementation gaps, implying it cannot be resolved by tuning alone.
Connections
This work speaks directly to ongoing empirical scrutiny of Community Notes and crowdsourced moderation, including Allen2025-ot, Renault2025-uh, Bak-Coleman2025-pm, and Murtfeldt2025-wu. Its cross-national, ideology-aware framing links it to broader platform-governance and polarization research such as Bouchaud2026-lr and Rieder2026-pp, and its concern with electoral integrity resonates with work on coordinated and polarizing content dynamics like Pierri2025-hm.
Podcast
A research-radio episode discusses this paper: 🎧 MP3 · Spotify