TLDR: Researchers developed an AI-driven social media simulator, LLM-SocioPol, using realistic demographics and network structures with LLM agents. It replicated a major Facebook experiment, showing that social messages (seeing friends vote) are more effective at mobilizing voters than just informational messages, confirming peer influence in a controlled digital environment. The simulation provides a flexible tool for studying political mobilization.
A groundbreaking study introduces a new way to understand how social media influences political participation, particularly voter turnout. Researchers Sadegh Shirani and Mohsen Bayati from Stanford University have developed an advanced simulation framework called LLM-SocioPol, which uses large language model (LLM) agents to mimic human behavior on social networks.
The core idea behind this research is to create a controlled, reproducible environment to study complex social phenomena that are difficult to observe and measure in the real world. Building on a landmark 2012 Facebook experiment, the LLM-SocioPol simulator integrates real U.S. Census demographic data, authentic Twitter network structures, and diverse LLM agents. Each simulated agent is given unique demographic traits, a political stance, and an LLM variant (like GPT-4.1 or its mini/nano versions) to reflect varying levels of political sophistication. These agents then interact within a realistic social network, receiving personalized content and dynamically adjusting their engagement and voting intentions.
How the Simulation Works
The simulator operates by assigning agents detailed profiles based on U.S. Census data, including age, gender, education, and occupation. Their social connections are derived from real Twitter follower networks. To make agents more realistic, they are further enhanced with interests, close friends, and a political stance (e.g., ‘consistent progressive voter’ or ‘low-turnout conservative’) determined by another AI model. Agents are also assigned different LLM models based on their simulated education and occupation, reflecting varying cognitive abilities.
During the simulation, agents experience ‘activity sessions’ where they log in, receive personalized content feeds, engage with posts (liking, replying), and update their likelihood of voting on a 0-4 scale. A crucial aspect is ‘context restoration,’ which allows LLM agents to remember their past interactions and maintain consistent behavior, overcoming the inherent statelessness of LLMs. Agents can also create new posts, contributing to a dynamic information environment.
Experimental Design and Findings
The study replicated the experimental conditions of the original Facebook study, testing three main scenarios:
- Control: Agents received no mobilization message.
- Informational Message: Agents saw a generic ‘get-out-the-vote’ message like ‘VOTE OR BE SILENCED! One ballot = one voice. Use yours.’
- Social Message: Agents received the informational message along with social cues, such as how many users were likely to vote and a list of followed users who had indicated their voting intention.
The results from LLM-SocioPol qualitatively mirrored the findings of the real-world Facebook experiment. Social messages consistently led to higher voter turnout compared to purely informational messages. This highlights the significant role of peer influence and social contagion in political mobilization. The simulation also showed measurable ‘peer spillovers,’ where untreated users were influenced by their friends who received the mobilization message.
However, there were quantitative differences. The simulated effects were generally larger than those observed in the real experiment, and the ratio of network-driven influence to direct influence was lower in the simulation. The researchers attribute these discrepancies to several factors: the prolonged exposure to messages in the simulation (30 days versus an election-day message in the original study), the isolated nature of the simulated environment without real-world distractions, and the absence of offline interactions in LLM-SocioPol.
Also Read:
- Virtual Justice: Simulating Legal Systems with AI Agents for Societal Insights
- AI Under Pressure: How Social Consensus Shapes ChatGPT’s Choices
Implications for Understanding Social Influence
This research demonstrates the potential of LLM-based simulations to complement traditional field experiments. By providing a controlled and repeatable environment, LLM-SocioPol allows researchers to test counterfactual scenarios, validate causal estimators, and experiment with interventions that might be infeasible or unethical in real-world platforms. The study reinforces the idea that social influence, particularly through online networks, is a powerful driver of political participation.
For more in-depth information, you can read the full research paper here.
