The Deep Structure of Social Media's Toxicity: A Q&A with Researcher Petter Törnberg

Social media's worst problems—echo chambers, attention inequality, and the amplification of extreme voices—aren't mere glitches. According to University of Amsterdam researcher Petter Törnberg, they are baked into the very architecture of platforms. In a series of new papers, he explains why conventional fixes fail and what might actually work. Below, we explore his findings through seven key questions.

1. What is the core finding of Törnberg's research on social media negativity?

Törnberg's central insight is that toxic feedback loops—such as partisan divides, concentration of influence among a few elite users, and the dominance of extremist voices—are structurally embedded in how social media platforms are built. Unlike face-to-face interactions, online environments lack natural friction: content spreads instantaneously, signals amplify without attenuation, and network effects reward the most sensational material. This means that even well-intentioned platform tweaks—like promoting civil discourse or adjusting feeds—seldom address the root causes. Instead, the architecture itself creates a system where negativity is not a bug, but a feature of the dynamics.

2. Why are platform-level interventions unlikely to fix these problems?

Many proposed solutions—such as fact-checking, content moderation, or tweaking recommendation algorithms—target symptoms rather than the underlying structure. Törnberg's agent-based models show that these interventions often fail because they don't change the fundamental dynamics of attention inequality and network polarization. For instance, even if you remove algorithmic curation, users still gravitate toward like-minded peers, and extreme content still attracts disproportionate engagement. The architecture of social media inherently rewards divisiveness and consolidates influence among a small group, making piecemeal reforms largely ineffective without a complete redesign of the platform's core logic.

3. How does social media's architecture differ from the physical world?

In physical spaces, conversations have natural limits: you can only talk to a few people at once, voices fade with distance, and social norms constrain behavior. Social media removes those constraints. Information flows instantly to millions, attention is zero-sum, and the absence of non-verbal cues reduces empathy. Törnberg emphasizes that this isn't about algorithms or human nature alone—it's the structural design. For example, the “like” and share buttons create a direct feedback loop that amplifies the most engaging (often extreme) content, something that wouldn't happen in a town hall meeting. This fundamental difference means that online social dynamics are inherently prone to distortion.

4. What new methods did Törnberg use in his latest papers?

To study these dynamics, Törnberg combined agent-based modeling with large language models (LLMs). He created simulated “AI personas” that behave like real social media users—posting, reacting, and forming networks based on predefined rules and language patterns. This hybrid approach allowed him to test how different architectural changes would play out at scale, without the ethical or logistical challenges of experimenting on real platforms. By generating millions of interactions, he could observe emergent phenomena like echo chambers forming even when individual agents were programmed to be open-minded. The method offers a powerful tool for diagnosing why certain interventions fail before they are ever deployed.

5. What specific aspect did the PLoS ONE paper focus on?

The study published in PLoS ONE zeroed in on the echo chamber effect. Using his LLM-powered agent models, Törnberg showed that echo chambers can emerge spontaneously even without algorithmic curation or partisan bias in the agents' initial beliefs. The mere structure of a network where information flows preferentially among similar nodes creates self-reinforcing cycles. The paper also demonstrated that common countermeasures—such as exposing users to diverse viewpoints or debiasing content—often backfire, because they fail to account for how users selectively interpret and share information. The takeaway: echo chambers are not just a product of bad algorithms, but a natural outcome of how online networks are wired.

6. Is there any hope for a fundamental redesign?

Törnberg is cautious but not entirely pessimistic. He argues that meaningful change would require a radical rethinking of social media's underlying incentives—for example, shifting from engagement-based metrics to quality-of-interaction metrics, or redesigning platform architecture to mimic physical-world friction (such as limiting message reach or adding deliberation costs). However, he acknowledges that incumbents have little incentive to pursue such changes, given their reliance on advertising revenue tied to user attention. Without regulatory pressure or the rise of alternative business models, the cycle of toxicity is likely to persist. The research, while sobering, provides a clear roadmap for what a healthier system would need to look like.

7. How do algorithms and human nature factor into the problem?

Contrary to popular belief, Törnberg's work suggests that algorithms and human psychology are not the primary villains. Human biases toward negativity and homophily certainly exist, but they have been present in all societies. Similarly, algorithms amplify these tendencies but are not the root cause. The real issue is that the structural environment of social media magnifies these biases beyond what would occur in face-to-face contexts. For instance, while people may prefer agreeable discussions, in a physical setting they cannot easily avoid disagreement; online, they can curate a perfect echo chamber. Törnberg emphasizes that blaming algorithms or human nature misses the deeper point: the architecture itself enables and rewards these behaviors, and until that architecture changes, the mess will continue.

Explore more of Törnberg's research: Core findings | Methodology | Echo chamber paper