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

This paper investigates whether large language models (LLMs) implicitly encode enough sociopolitical structure to reconstruct aggregate voting behaviour using only individual-level demographic profiles. Rather than treating LLMs as raw predictors, the authors conceptualize them as compressed representations of social reality. They condition several OpenAI models on 10 sociodemographic and attitudinal variables—such as age, region, education, income, and subjective living standards—derived from a representative survey of the Czech population.

The empirical validation focuses on the 2021 Czech parliamentary election, a demanding testbed characterized by a multiparty system, post-communist legacies, and a medium-resource language. The researchers query the models to output probabilistic turnout and party preferences in Czech, using a zero-temperature setting and structured JSON outputs. To translate individual estimates into macro-level insights, they apply a soft voting aggregation procedure that multiplies party preference probabilities by turnout likelihoods, allowing individual-level uncertainties to cancel out at the population scale.

The study demonstrates that contemporary LLMs can reproduce official election outcomes with low mean absolute error when aggregated properly. The simulated vote shares align closely with independently established sociodemographic gradients and successfully recover known political bloc structures and inter-party correlations observed in the Czech political landscape. However, the models vary significantly by scale: while larger models maintain calibrated probabilistic uncertainty regarding turnout and voting choices, smaller models exhibit severe distribution compression and occasionally assign absolute zero probabilities to non-voters, highlighting vulnerabilities in handling minority or complex voter groups.

These results show that LLMs internalize durable social and political regularities from their training corpora, effectively serving as implicit sociological models. This provides computational social science with a novel exploratory instrument for simulating aggregate social phenomena and studying political polarization or demographic shifts. At the same time, the findings establish clear epistemic boundaries: successful macro-level reconstruction does not imply that individual-level synthetic responses are valid or unbiased human surrogates, urging caution in how generative simulations are interpreted.