4-year study of X's Community Notes reveals moderation patterns and biases
Researchers analyzed 4 years of X's community-based moderation data — here's what they found.
Researchers from multiple institutions published a comprehensive study of X's Community Notes program over its first four years (2021–2025). They parsed the full set of notes and ratings, performed language detection across all notes, extracted embedded URLs from English-language notes, and identified discussion topics. Monthly interaction networks among contributors were constructed to map collaboration patterns. The analysis reveals that despite the program's global scope, English-language notes dominate, and contributor activity is highly skewed toward a small core group. Sourcing practices (e.g., linking to external articles) vary significantly by topic and language.
The team also released a curated dataset and accompanying source code to enable replication and further research. This is the first large-scale descriptive analysis of Community Notes, which started as Birdwatch on Twitter in 2021 and has since been adopted by other platforms like YouTube and Meta. The study provides a baseline for evaluating the effectiveness, biases, and scalability of crowdsourced content moderation. Key findings around linguistic diversity and contributor interaction networks suggest that without deliberate design changes, such systems may replicate offline biases and power imbalances.
- Analyzed 4 years of Community Notes data including ratings, URLs, and contributor networks.
- English-language notes dominate the dataset, raising questions about multilingual moderation effectiveness.
- Released open-source dataset and code to encourage further research on community-based content moderation.
Why It Matters
As crowdsourced moderation spreads to major platforms, understanding its biases and dynamics is critical for building fairer systems.