Anna Keuchenius, Petter Törnberg, Justus Uitermark
Sociological Science May 11, 2026
10.15195/v13.a22
Abstract
The influential “echo chamber” hypothesis suggests that social media drive polarization through a mutual reinforcement between isolation and radicalization. The existence of such echo chambers has been a central focus of academic debate, with competing studies finding ostensibly contradictory empirical evidence. This article identifies a fundamental methodological limitation of these empirical studies: they do not differentiate between negative and positive interactions. To overcome this limitation, we develop a method to extract signed network representations of Twitter/X debates using machine learning. Applying our approach to a major Dutch cultural controversy, we show that the inclusion of negative interactions provides a new empirical picture of the dynamics of online polarization. Our findings suggest that conflict, not isolation, is at the heart of polarization.
![]() | This work is licensed under a Creative Commons Attribution 4.0 International License. |
Reproducibility Package: The data and code underlying this article are available as part of our replication materials available at this link: https://doi.org/10.6084/m9.figshare.30948659.v1 or the DOI: 10.6084/m9.figshare.30948659. Due to Twitter/X’s Terms of Service and ethical and legal obligations, we do not share tweet text, user IDs, or any data that could identify individuals or their stance in the debate. The dataset contains sensitive information, including political opinions, and releasing identifiable content would pose ethical risks to users and violate GDPR requirements. To support replication, we provide code, documentation, and non-identifying data sufficient to reproduce all analytical steps, with full analyses possible via rehydration of tweet IDs.
- Citation: Keuchenius, Anna, Petter Törnberg, and Justus Uitermark. 2026. “Echo Chambers Are Defined by Conflict, Not Isolation” Sociological Science 13: 565-588
- Received: January 8, 2025
- Accepted: January 12, 2026
- Editors: Arnout van de Rijt, Bart Bonikowski
- DOI: 10.15195/v13.a22


