Beyond the Feed: Why Social Media's Architecture Is Its Own Undoing

The Structural Flaws of Social Media

In recent years, the cracks in social media’s facade have become impossible to ignore. Echo chambers trap users in ideological bubbles, a small elite hoards the spotlight, and the most extreme voices drown out the moderate majority. These aren’t bugs; according to Petter Törnberg, a researcher at the University of Amsterdam, they are features—hardwired into the very blueprint of platforms like Twitter and Facebook. His work, which we first explored last fall, argues that the root causes are not algorithms, chronological feeds, or human appetite for negativity. Instead, the dynamics that breed toxicity are embedded in the architecture of social media itself.

Why Current Fixes Fail

Törnberg’s earlier research demonstrated that most proposed interventions—such as tweaking recommendation algorithms or promoting civil discourse—are doomed to fail. They treat symptoms, not the disease. The problem is that social media operates under fundamentally different structural conditions than physical-world interactions. In real life, conversations are bounded by time, space, and social cues. Online, these constraints vanish, allowing extreme viewpoints to spread unchecked and attention to concentrate among a few. Törnberg concluded that without a complete architectural overhaul, we are trapped in a loop of escalating polarization.

New Research into Echo Chambers

Since that interview, Törnberg has been prolific, producing two new papers and a preprint that deepen this structural critique. The first, published in PLoS ONE, zeroes in on the echo chamber effect. To study it, he employed a novel hybrid method: combining standard agent-based modeling with large language models (LLMs). He essentially created AI personas—digital stand-ins for real users—and set them loose in a simulated social media environment.

Simulating Online Behavior with AI Personas

These artificial users were programmed with basic preferences and biases, then allowed to interact, share content, and form connections. The LLMs gave them the ability to generate and respond to posts in a human-like manner. What emerged was a stark replica of the real world: the AI personas naturally gravitated toward like-minded peers, reinforcing their own views and ignoring dissent. The simulation confirmed that echo chambers are not accidental; they are an emergent property of the platform’s structure. Even when external moderation was introduced, the chambers persisted.

The Road Ahead

Törnberg’s findings suggest that minor adjustments won’t suffice. The architecture itself must be rethought—perhaps by introducing friction into interactions, or by redesigning how attention is distributed. But he remains skeptical that platforms, driven by profit motives, will voluntarily embrace such changes. As users, we may need to prepare for a messy transition, where the old social media model fades and something—unknown and unproven—takes its place. The research offers a sobering map, but the destination is uncertain.