Why Social Media’s Toxic Dynamics Are Built Into Its Architecture

Research from the University of Amsterdam suggests that the most toxic elements of social media are not merely the result of specific algorithms, but are instead structurally embedded in the architecture of the platforms themselves.

Petter Törnberg, a researcher at the University of Amsterdam, has identified several recurring negative outcomes of social media design, including the creation of partisan echo chambers, the amplification of the most extreme divisive voices, and a phenomenon known as attention inequality.

Attention inequality is defined as the concentration of influence among a small group of elite users, which further skews the distribution of visibility and power on these platforms.

In research highlighted in August 2025, Törnberg found that numerous platform-level intervention strategies intended to combat these issues are unlikely to be effective.

The findings indicate that these negative dynamics are not caused by non-chronological feeds, specific much-hated algorithms, or a human tendency to seek out negativity.

Instead, the research concludes that these outcomes are a product of the fundamental architecture of social media, which differs significantly from the structure of the physical world.

Törnberg suggests that users may be doomed to endless toxic feedback loops unless the industry implements a brilliant fundamental redesign that manages to change those dynamics.

To reach these conclusions, Törnberg utilized a methodology that combined standard agent-based modeling with large language models (LLMs).

This approach involved creating AI personas to simulate online social media behavior, allowing researchers to observe how structural design influences interaction without the need to manually massage the model or rely solely on existing algorithms.

Building on these realizations, Törnberg has since produced a new preprint and two additional papers.

One of these papers, published in the journal PLoS ONE, specifically examined the echo chamber effect. This study employed the same simulation technique of using LLM-powered AI personas to model social media behavior.

The ongoing research emphasizes that because the architectural flaws are systemic, surface-level patches or moderation adjustments are insufficient to resolve the underlying issues of polarization and influence concentration.