New AI method measures political partisanship across Bluesky and Truth Social
Analyzes 1.3M posts using transformer embeddings and news credibility signals
A new preprint from arXiv (2607.21842) introduces a cross-platform methodology for measuring political partisanship in social media content. Authors Fathima Ameen and Christopher Healey address the fragmentation of social media ecosystems, where existing tools fail to generalize across platforms like mainstream Bluesky and alt-tech Truth Social. Their approach uses a transformer-based sentence encoder to embed posts, then clusters them into topic groups labeled by aggregated media bias scores from AllSides, a news-credibility database. A partisanship axis is constructed from the difference between centroids of opposite-labeled clusters, and individual posts are scored by projection onto this axis.
The researchers tested their method on a corpus of approximately 1.3 million posts collected from Bluesky and Truth Social during the six months before the 2024 U.S. presidential election. This marks the first cross-platform comparison of partisanship distributions on these ideologically asymmetric platforms. The resulting partisanship scores correlated significantly with held-out AllScores both in-distribution and out-of-distribution on an independent Twitter corpus. The model also recovered within-platform partisan dynamics that platform identity alone could not explain, offering a scalable tool for studying polarization as the social media landscape continues to fragment.
- Method uses transformer-based sentence embeddings and AllSides media bias scores to create a platform-portable partisanship measure.
- Applied to 1.3 million posts from Bluesky and Truth Social ahead of the 2024 U.S. election.
- Partisanship scores correlate with external bias benchmarks and reveal intra-platform dynamics beyond simple platform labels.
Why It Matters
Provides a generalizable, scalable tool to track political polarization across emerging alt-tech and mainstream platforms.