Abstract
Social scientists have long debated whether political polarization is driven by "top-down" elite cues or "bottom-up" ideological constraints. This study introduces "In Silico Sociology," using a version of GPT-3 frozen in 2019 to simulate partisan reactions to the then-future COVID-19 pandemic. The model accurately predicted the direction of partisan polarization in 84% of cases, suggesting that the political divide was already prefigured in the ideological discourse of 2019.
Introduction: The Roots of Polarization
The challenge in studying real-time crises like COVID-19 is that public opinion is often immediately influenced by political messaging. To isolate the role of pre-existing ideological values, researchers turned to Large Language Models (LLMs) as a "time capsule." By using a model with zero knowledge of the actual pandemic, they could test if the subsequent polarization was an inevitable result of the 2019 ideological landscape.
Methodology: Simulating Partisan Personas
The study utilized the original GPT-3, which completed training in October 2019. The researchers employed a multi-step process:
- Partisan Priming: The model was prompted with specific personas (e.g., "I am a strong conservative Republican...") before being asked to complete sentences regarding 49 COVID-related issues, such as mask mandates and vaccine efficacy.
- Token Generation: The model predicts subsequent words based on the probability distributions learned during training.
- Semantic Scoring and Validation: Responses were projected onto mathematical axes using embedding models to quantify valence (e.g., support vs. opposition). These simulated results were then compared against real-world survey data from 2020 (e.g., Gallup Panel) to verify accuracy.
Key Insights: Ideology as a "Track"
The findings suggest that history often moves along pre-existing linguistic "rails." Even for an unprecedented event like COVID-19, the reactions were highly predictable based on the values present in 2019 discourse.
The model revealed that the divide was rooted in institutional trust and a fundamental tension between collective safety and individual liberty. These themes were not new; they were already deeply embedded in the "latent probability distributions" of American political language.
Results and Discussion
GPT-3 achieved an 84% accuracy rate in predicting the direction of the partisan gap. The strongest predictions occurred in areas with clear ideological stakes, such as vaccine mandates and school closures.
However, the model failed in predicting emotional responses; it expected conservatives to be more fearful of the biological threat, whereas in reality, they were less so. This discrepancy highlights where political leadership (e.g., Trump’s messaging) successfully overrode prior ideological tendencies, shifting the "emotional framing" of the crisis.
Conclusion and Future Directions
This research demonstrates that LLMs can serve as a "Turing Test" for sociology. If a model can successfully impersonate a social group's reaction to a novel event, it suggests that the group's culture is a coherent, constrained system. Future "Causal Cultural Analysis" could involve training models on sequential checkpoints to pinpoint exactly when specific events or speeches shifted the cultural needle.
Summary updated