AI Study: Social Media's Dysfunctions Are Structurally Embedded

Much of the promise of social media—to foster healthy public discourse and connect individuals in a utopian digital square—has instead given way to profound dysfunction. Rather than bridging divides, these platforms frequently cultivate insular filter bubbles and echo chambers. A disproportionately small number of high-profile users command the lion’s share of attention and influence, while algorithms designed to maximize engagement often amplify outrage and conflict. This dynamic, researchers suggest, ensures the dominance of the loudest, most extreme voices, thereby exacerbating societal polarization.

Despite numerous proposed platform-level interventions to mitigate these issues, a recent preprint published in the physics arXiv suggests that most are unlikely to be effective. The study challenges the common belief that the blame lies primarily with much-maligned algorithms, non-chronological feeds, or even an inherent human inclination towards negativity. Instead, its authors contend, the negative outcomes are structurally embedded in the very architecture of social media. This implies that societies may be caught in endless toxic feedback loops unless a truly fundamental redesign emerges that can alter these deep-seated dynamics.

Petter Törnberg and Maik Larooij of the University of Amsterdam, the study’s co-authors, sought to understand the underlying mechanisms driving social media’s most problematic aspects: partisan echo chambers, the concentration of influence among a small elite (attention inequality), and the amplification of divisive, extreme voices. They employed a novel approach, combining standard agent-based modeling—which simulates how individual “agents” interact within a system—with large language models (LLMs). This allowed them to create AI personas, imbued with detailed characteristics from voter surveys, to simulate complex online social behavior. Surprisingly, Törnberg noted, these dysfunctional dynamics emerged organically from their baseline model, requiring no deliberate algorithmic inputs or model adjustments.

The researchers then rigorously tested six distinct intervention strategies commonly proposed by social scientists. These included switching to chronological or randomized feeds, inverting engagement-optimization algorithms to suppress sensational content, boosting viewpoint diversity to broaden users’ exposure to opposing political views, and implementing “bridging algorithms” designed to elevate content that fosters mutual understanding over emotional provocation. They also explored hiding social statistics like reposts and follower counts to reduce influence cues, and removing biographies to limit identity-based signals.

The results proved disheartening. While some interventions yielded modest improvements, none managed to fully disrupt the fundamental mechanisms responsible for the dysfunctional effects. In a few cases, interventions even exacerbated existing problems. For instance, chronological ordering, while effective at reducing attention inequality, simultaneously intensified the amplification of extreme content. Similarly, bridging algorithms weakened the link between partisanship and engagement and slightly improved viewpoint diversity, but at the cost of increasing attention inequality. Boosting viewpoint diversity, surprisingly, had no significant impact whatsoever.

Törnberg explains that these pervasive issues stem from the core dynamics of online social networks—the constant cycles of posting, reposting, and following. These actions, often driven by emotional and partisan reactions, not only spread toxic content but also actively shape the very network structures that emerge. This creates a self-reinforcing feedback loop: emotional actions lead to polarized network structures, which in turn dictate the type of content users encounter, leading to an increasingly toxic environment. Even platforms like Bluesky, which notably eschew traditional algorithms, appear to succumb to similar dynamics, lending credence to the study’s findings that the problem is structural, not merely algorithmic.

This inherent structure also distorts our perception of reality. Social media, Törnberg elaborates, acts as a “social media prism,” presenting a version of politics that appears far more toxic and polarized than it is. While actual polarization may be lower, the perceived polarization is significantly higher, largely due to the platform’s amplification of a tiny fraction of users—often those who are most outrageous or extreme—who then disproportionately influence the conversation. This “power law distribution” means that a mere one percent of users can dominate the entire discourse, creating an incentive structure where certain personalities thrive, reshaping not just how we see politics, but who becomes politically powerful.

The implications extend beyond individual platforms. The study suggests that even traditional media and broader culture are being reshaped by “social media logic.” Headlines, for instance, have become more “clickbaity” to align with what performs well online. This pervasive influence means that simply opting out of social media is not a solution, as its incentive structures continue to transform politics, empower specific individuals, and fundamentally alter the cultural landscape.

The researchers acknowledge that the current model of social media faces an existential crisis, particularly with the rise of increasingly powerful large language models. These LLMs, capable of mass-producing information optimized for attention—often false or highly polarized content—will likely overwhelm conventional social media structures. While this might lead to a retreat towards more curated, closed communities like private messaging groups, it’s unclear if this shift will ultimately lead to a healthier digital environment. The study ultimately suggests that if societies wish to cultivate genuinely constructive public discourse, they may need to move away from globally interconnected social network models towards more localized or group-based structures that avoid the pitfalls of highly influential, centralized nodes.