Differences in misinformation sharing can lead to politically asymmetric sanctions

Mohsen Mosleh; Qi Yang; Tauhid Zaman; Gordon Pennycook; David G. Rand

doi:10.1038/s41586-024-07942-8

In brief

When technology companies adopt politically neutral anti-misinformation policies, enforcement outcomes can still appear asymmetric across political groups—not because of bias in the policy, but because misinformation-sharing behavior itself is asymmetric. Analyzing Twitter suspensions after the 2020 election and testing on seven datasets, the authors show that conservatives shared significantly more low-quality news links and false claims than liberals, explaining observed suspension disparities without requiring intentional platform bias.

Contributions

Demonstrates that political asymmetries in enforcement outcomes can emerge naturally from asymmetric misinformation sharing patterns, without requiring partisan bias in policy implementation.
Introduces a methodology to disentangle platform bias from behavioral asymmetry using both expert and politically balanced crowdsourced fact-checker ratings to evaluate content quality.
Provides evidence across nine datasets spanning Twitter, Facebook, and survey experiments from 2016–2023 and 16 countries, showing the pattern is robust and generalizable.
Shows that conservatives' higher estimated likelihood of suspension stems from their greater propensity to share low-quality news and false claims, not from discriminatory policy enforcement.

Method

The researchers analyzed the 2020 US presidential election on Twitter, specifically examining users who had shared hashtags related to the election. They randomly selected 4,500 Twitter users who shared at least one #VoteBidenHarris2020 hashtag and an equivalent sample of #Trump2020 users. They collected the users' shared links and evaluated news quality using two approaches: (1) ratings from eight professional fact-checkers, and (2) politically balanced crowdsourced ratings from laypeople representing both Democrats and Republicans.

The paper evaluates the association between political orientation (assessed via user sharing behavior and social networks), low-quality news sharing, and Twitter suspension rates. They then test whether this pattern replicates across seven additional datasets on Facebook, Twitter, and survey-based sharing experiments from 2016 to 2023, spanning 16 countries.

Key analyses include:

Correlation analysis between average news quality shared and political orientation
Logistic regression models predicting suspension likelihood based on low-quality news sharing, controlling for confounds
Cross-dataset validation using both professional fact-checker and politically balanced crowdsourced quality ratings
Simulations of hypothetical politically neutral enforcement policies to assess whether asymmetric suspension rates would persist

Results

The headline finding is striking: conservative Twitter users during the 2020 election were 4.4 times more likely to have been subsequently suspended than liberal users who shared #VoteBidenHarris2020 hashtags. However, this disparity directly correlates with their sharing behavior:

Conservatives shared significantly more links to low-quality news sites (using fact-checker ratings: r = 0.88, p < 0.0001; using politically balanced layerson ratings: r = 0.82, p < 0.0001).
The correlation between sharing low-quality news and suspension was strong (using fact-checker ratings: r(57) = 0.68, p < 0.0001; using politically balanced ratings: r(57) = 0.74, p < 0.0001).
When the researchers simulated a politically neutral suspension policy based purely on low-quality news sharing, it predicted the observed asymmetry in actual suspensions.

Across the seven replication datasets (spanning Facebook, Twitter, and survey experiments from 2016–2023 and 16 countries), the pattern held: conservatives were estimated to be more susceptible to suspension because they shared more low-quality information than liberals, regardless of which method (professional fact-checkers or politically balanced layperson ratings) was used to evaluate content quality.

Critically, the authors show that the asymmetry in suspension rates cannot be easily attributed to platform bias. When using unbiased measures of content quality and simulating politically neutral enforcement, the observed disparities emerge naturally.

Notes

This paper makes an important methodological contribution: it shows how to disentangle platform bias from behavioral asymmetry in enforcement outcomes. The use of both professional fact-checkers and politically balanced crowdsourced ratings is particularly valuable, as it demonstrates that the finding is robust to the method of evaluating content quality. The cross-dataset replication spanning countries, platforms, and time periods strengthens confidence in the generalizability of the result.

The paper does not claim that all platform enforcement asymmetries lack a bias component, nor does it argue that politically neutral policies are sufficient for fairness. Rather, it demonstrates that observed asymmetries in enforcement can be explained by asymmetric misinformation sharing patterns alone, which has important implications for policy design: platforms implementing unbiased policies should still expect to observe politically asymmetric enforcement outcomes if misinformation sharing is itself asymmetric.