Contributions
- Novel methodology reducing evaluator-bias concerns. Leverages X's Community Notes bridging algorithm, which requires agreement from politically diverse raters, to assess partisan asymmetries in misinformation sharing without relying on potentially biased professional fact-checkers.
- Large-scale empirical quantification of partisan misinformation asymmetry. Analyzes 218,382 proposed Community Notes on 162,228 tweets from 39,140 unique users between January 2023 and June 2024, finding 69.79% of flagged posts written by Republicans vs. 30.21% by Democrats among helpful notes.
- Ruling out confounding explanations. Demonstrates that the partisan asymmetry cannot be attributed to (1) overrepresentation of Republicans on X, using random API samples showing Democrats still outnumber Republicans; or (2) overrepresentation of Democratic raters in Community Notes, since the bridging algorithm should suppress such bias on published notes, yet the asymmetry is starker among helpful notes.
- Topic-level analysis of partisan misinformation patterns. Shows substantial partisan differences across content categories: Republicans most overrepresented in Health (81.9% of flagged posts), Politics (73.3%), and Science (68.8%) misinformation.
Method
Partisanship inference. The authors infer user political affiliation using three complementary methods: (1) social connections (Mosleh & Rand 2022), (2) ideology scores from follower networks (Barberá 2015), and (3) large language model analysis of tweet content (GPT-4o mini). Users are included only if they receive consistent classifications from at least two of the three methods.
Data and analysis. The analysis covers English-language Community Notes flagged as "misinformed or potentially misleading," written between January 2023 and June 2024. For each proposed note, the study tracks whether it was rated helpful by the bridging algorithm—indicating agreement from users with diverse perspectives that the flagged post is misleading. The key outcome variable is note helpfulness, not mere proposal, to ensure only notes with genuine cross-partisan agreement are counted.
Controls. Logistic regression accounts for user-level characteristics (verified status, follower count, total tweet volume) and tweet topic (Politics, Science, Health, Economy, Other).
Results
Among all 218,382 proposed Community Notes, 60.05% targeted Republican posts vs. 39.95% for Democrat posts. However, this proposal imbalance could reflect ideological conflict rather than genuine misinformation asymmetry.
Key finding: Among helpful notes, the partisan asymmetry becomes stark. 69.79% of flagged posts (13,658 posts) were written by Republicans, compared to 30.21% by Democrats (5,911 posts)—a ratio of 2.3:1. In logistic regression controlling for user and tweet characteristics, being a Republican poster increased the probability of a proposed note achieving helpful status by 63.49%.
Robustness across inference methods: The asymmetry holds when using only the follower-network method (1.7 times as many Republican posts), or only the ideology-score method (2.1 times as many).
Topic variation: Partisan overrepresentation of Republicans among helpful notes varies by topic:
- Health: 81.9% Republican
- Politics: 73.3% Republican
- Science: 68.8% Republican
- Other topics: 66.9% Republican
- Economy: 63.7% Republican
Confound testing: A random sample of 474,394 tweets containing "the" from January 2014 to November 2024 shows Democrats still outnumber Republicans on X overall (though Republican share increased post-Musk acquisition). Thus, base-rate overrepresentation of Republicans cannot explain the flagging pattern. Additionally, the bridging algorithm's requirement for cross-partisan agreement makes Democratic rater overrepresentation an unlikely driver of the effect; if anything, the more asymmetric pattern among helpful notes (vs. proposed notes) contradicts a Democratic-rater-bias hypothesis.
Notes
Strengths:
- Addresses a key criticism of prior work. Uses crowdsourced, politically diverse raters rather than professional fact-checkers, directly countering Musk and Zuckerberg's claim that observed partisan asymmetries reflect evaluator bias.
- Rigorous confound testing. The authors systematically rule out base-rate explanations with API data and acknowledge the bridging algorithm's role in suppressing rater-composition bias.
- Large, real-world dataset. Analyzes nearly a quarter-million notes from actual platform data, not experiments or surveys.
- Multiple partisanship inference methods. Cross-validates political affiliation using three independent approaches.
Limitations:
- Partisanship inference imprecision. Classification relies on social connections and LLM analysis; edge cases and borderline accounts may be misclassified, though the use of convergence across three methods mitigates this.
- Temporal snapshot. Data covers January 2023 to June 2024; long-term trends and pre-2023 patterns are not captured.
- Topic classification. Uses a topic classifier (Chuai et al. 2024) that may misclassify or miss nuanced topics; cross-topic generalization of the partisan asymmetry is not directly tested.
- Causal mechanisms unclear. The paper establishes that Republicans share more flagged content, but does not directly test whether this reflects ideological differences in belief formation, differential susceptibility to falsehoods, or other mechanisms.
Implications: The findings provide empirical evidence that platforms moving from professional fact-checking to crowdsourced systems (e.g., X Community Notes, Meta's Oversight Board) will not eliminate observed partisan asymmetries in enforcement, because the asymmetries reflect genuine differences in misinformation sharing, not evaluator bias. This undercuts a key justification for eliminating professional fact-checkers.