Petter Törnberg on Misinformation, Social Media Dynamics, and LLM Simulations

Shifting Epistemologies: From Machines to Complex Systems

• 💡 The concept of "power has an epistemology" suggests that how we understand society (as a machine vs. a complex system) shapes our approach to governance and social design.
• ⚙️ Historically, society was viewed as a complicated system (like a machine) with a top-down, state-driven design approach (Fordism).
• 🌐 Today, society is increasingly understood through a complex systems lens, emphasizing bottom-up, self-organizing emergent behaviors.
• ⚠️ This shift, while seemingly abandoning top-down control, can obscure new forms of power operating through less visible mechanisms like algorithms.

The Schelling Segregation Model and Digital Echo Chambers

• 🎲 The Schelling segregation model demonstrates how even mild individual preferences for similar neighbors can lead to extreme segregation in a system.
• 💻 Applying this model to digital communities (like subreddits) reveals that echo chambers can emerge even without algorithmic filtering.
• 📉 This suggests that the structure of social interaction itself, rather than solely algorithms, can drive polarization and segregation online.
• 🚫 Interestingly, filtering algorithms (like filter bubbles) can sometimes reduce segregation by making individuals less prone to move.

LLM Simulations of Social Media Dynamics

• 🤖 Agent-based models are being enhanced by using Large Language Models (LLMs) to simulate more realistic human behavior in social media.
• 📊 An experiment using LLMs on a barebones social platform replicated problematic outcomes like echo chambers, attention inequality (power-law distributions), and the social media prism (polarization).
• 📈 These emergent phenomena were robust, appearing even without explicit algorithmic manipulation, suggesting they are deeply linked to the platform's basic structure.
• ⚠️ Interventions like chronological timelines or focusing on constructive comments did not fix these issues and sometimes worsened them.

The Role of LLMs and Future Directions

• 🎭 LLMs were given personas based on real-world data (like political affiliation) to simulate diverse user behavior.
• 🗣️ The LLMs' interactions revealed that preferential attachment drives attention inequality, while emotional sharing and feedback loops contribute to polarization and the social media prism.
• ⚠️ A key limitation is validating LLM behavior against human behavior, as they are neither fully empirical nor fully formal models.
• 💡 The robustness of emergent outcomes in simulations provides more confidence than analyzing conversational toxicity alone.

Social Media's Impact on Society and Information

• 📈 Social media's business model, driven by advertising and niche creation, has fundamentally shaped its structure and impact.
• 📰 Social media influences mainstream media, leading to more clickbaity headlines and a focus on attention-grabbing content.
• 📢 The removal of traditional gatekeepers and incentives for attention can lead to the spread of misinformation, particularly by radical right-wing populist parties.
• 🌐 While social media has positive aspects for certain communities, its current structure is deeply linked to problematic political and social outcomes, requiring fundamental rethinking rather than cosmetic changes.