Large Language Models as Implicit Sociological Models: Reconstructing Voting Behaviour from Sociodemographic Profiles

Research from arXiv suggests that Large Language Models (LLMs), by virtue of their training on vast internet data, implicitly encode extensive statistical regularities concerning social identities, attitudes, and political behavior. A new methodological framework proposes leveraging these LLMs as 'implicit sociological models' to reconstruct aggregate voting behavior from individual-level sociodemographic profiles. Validated against the 2021 Czech parliamentary election, current LLMs demonstrated the ability to reproduce official election outcomes with low mean absolute error, indicating their potential for modeling complex social phenomena.

Why it matters

This development indicates a novel application of AI for understanding and potentially predicting complex societal dynamics, such as electoral outcomes, based on demographic data. It highlights the embedded sociological knowledge within advanced AI systems, which could inform strategic planning across various sectors dependent on public sentiment or societal trends.

Key insights

• LLMs trained on internet corpora capture statistical regularities related to social identities, attitudes, and political behaviors.
• A new framework uses LLMs as implicit sociological models, conditioning them on demographic profiles to predict probabilistic turnout and party preferences.
• Individual LLM outputs are aggregated through a soft voting procedure to reconstruct overall voting behavior.
• The method was validated against the 2021 Czech parliamentary election, showing LLMs can reproduce official outcomes with low mean absolute error.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.15871