Misinformation and harmful language are interconnected, rather than distinct, challenges

Contributions

• Large-scale empirical demonstration that misinformation and harmful language co-occur on social media across distinct datasets and units of analysis
• Analysis of 8.7M Twitter posts from 6,832 users, complemented by 14,617 fact-checked headlines, showing consistent patterns
• Evidence that the association holds at post level, user level, and across news vs. non-news content
• Finding that despite statistical significance, the effect sizes are insufficient for harmful language to serve as a practical detection signal for information quality

Method

The study combines Twitter behavioral data with fact-checked headline analysis. For Twitter, the authors sampled N=6,832 users selected from those following major political figures, retrieved up to 3,200 posts per user (8.7M posts total), and measured news source quality via domain trustworthiness ratings from 4,767 news websites. Harmful language was detected using Google Jigsaw Perspective API (measuring toxicity, severe toxicity, identity attack, insult, profanity, threat) plus a hate-speech classifier, reduced via principal component analysis (PCA) to a single harmful language score.

For headlines, 14,617 headlines from Snopes (10,672 false, 3,945 true) and PolitiFact were analyzed with the same harmful-language approach. All analyses used linear regression with standardized coefficients, including controls for user characteristics (political affiliation, bot probability, gender/age estimates, follower counts).

Results

Post-level association: Twitter posts containing links to low-quality news websites contained significantly more harmful language (β=-0.102, p<0.001; β=-0.095 with controls). False headlines contained more harmful language than true headlines (β=-0.190, p<0.001).

User-level association: Users who share links to low-quality news sources also use more harmful language overall (β=-0.135, p<0.001). This association held for both news posts (β=-0.094) and non-news posts (β=-0.134), indicating the relationship is not simply driven by hostile posts about low-quality content, but rather a broader user-level tendency.

Practical diagnosticity: Despite statistical significance, the associations were weak enough that most low-quality information contains no harmful language, and many high-quality posts do contain harmful language. Thus harmful language cannot serve as a practical screening tool for information quality.

Notes

The paper's key insight is that these two largely separate research literatures (misinformation detection and hate-speech moderation) may share underlying psychological mechanisms, suggesting value in integrating interventions. However, the weak diagnosticity finding is important for practitioners — one cannot use harmful language as a shortcut signal for misinformation, nor vice versa. The use of domain-quality proxies rather than individual headline accuracy is a limitation but reflects the constraints of large-scale Twitter analysis.