Source-linked AI summary
Large Language Models as Implicit Sociological Models: Reconstructing Voting Behaviour from Sociodemographic Profiles
Roman Neruda, Martin Bakoš, Josef Šlerka, Vít Tuček, Petra Vidnerová, Gabriela Kadlecová
TL;DR
LLM electoral simulations have been tested mainly in US two-party settings, leaving their ability to recover voting behaviour in challenging multilingual, multiparty contexts limited. This paper conditions LLMs on sociodemographic profiles and aggregates probabilistic outputs to reconstruct the 2021 Czech election, achieving low aggregate error while recovering political bloc structures and sociodemographic voting gradients.
Problem
Evidence remains limited on whether LLM electoral simulations generalise to multilingual, multiparty contexts such as the Czech Republic.
Method
The framework translates respondent sociodemographic attributes into demographic prompts, elicits probabilistic vote choices, and aggregates outputs through soft voting.
Results
The framework reproduced official Czech 2021 vote shares with low aggregate error, recovered known political bloc structures, and aligned with established sociodemographic voting gradients.
Takeaways & Limitations
The approach provides a systematic way to interrogate latent social knowledge in language models and assess its correspondence with observed social structures.
Takeaways & Limitations
Prompt wording may introduce unintended framing effects, and Czech-language encoding may be more fragile than English-language capabilities.
Abstract
from arXiv · showhide
Large language models (LLMs) trained on large-scale internet corpora encode extensive statistical regularities about social identities, attitudes, and political behaviour. This paper introduces and evaluates a methodological framework that leverages these latent representations to reconstruct aggregate voting behaviour from individual-level sociodemographic profiles. We operationalize LLMs as implicit sociological models by conditioning them on demographic descriptions, eliciting probabilistic turnout and party preferences, and aggregating individual outputs via a soft voting procedure. Using the 2021 Czech parliamentary election as a validation case, we demonstrate that contemporary LLMs reproduce official election outcomes with low mean absolute error, recover known political bloc structures, and align with independently established sociodemographic gradients. The contribution of this work is methodological rather than predictive: we show how LLMs can be systematically interrogated as compressed representations of social reality, offering a novel exploratory instrument for computational social science while clearly delineating its epistemic and ethical limits.
1 Introduction
The paper asks whether LLMs encode sociopolitical structures sufficiently to reconstruct aggregate voting behaviour from demographic information. It proposes and validates a methodological framework using demographic conditioning and probabilistic aggregation, treating LLMs as implicit sociological models rather than individual-level predictors.
- Motivation: LLMs trained on internet-scale corpora internalize structured statistical associations between social identities and political preferences.The paper frames these models as a novel form of cultural compression rather than systems encoding explicit rules or theories.
- Research question: The study tests whether LLMs can reconstruct aggregate voting behaviour from demographic information using the 2021 Czech parliamentary election.Multiple LLMs are conditioned on individual-level sociodemographic profiles from a representative survey and produce probabilistic voting choices for aggregation.
- Motivation: The Czech case is demanding because it combines a medium-resource language, documented Czech performance gaps, and a complex multiparty system that may be poorly captured by Anglophone training data.The introduction also notes context-dependent failures in non-US electoral simulations and systematic political-orientation shifts by language, including Czech.
- Contributions: The paper contributes a general framework for demographic conditioning, a soft voting procedure for converting probabilistic individual outputs into macro-level estimates, and multi-level validation.The framework is intended to handle individual-level noise while remaining diagnostic of systematic biases.
2 Related Work
Prior research documents strong links between sociodemographic position and party choice, while computational and LLM-based approaches seek to model these relationships despite interaction, language, and context challenges. This work extends silicon sampling to a non-US multiparty election by eliciting probability distributions, applying soft aggregation, and validating methodological reconstruction rather than survey replacement.
- Sociodemographic foundations: Cleavage theory established structural divides linking sociodemographic position to European party choice, but Czech centrist populist parties may be better predicted by political cynicism than demographics.This suggests demographic conditioning may reconstruct some parties more faithfully than others.
- Computational electoral modeling: MRP and related hierarchical or machine-learning methods dominate computational electoral modeling, yet multiplying demographic dimensions makes relevant interaction effects combinatorially demanding.The common challenge is that analysts must specify or discover these interactions.
- LLM-based simulation: Silicon sampling conditions LLMs on demographic backstories and demonstrates algorithmic fidelity through demographically correlated responses reflecting attitudes and sociocultural context.Later work extends the paradigm to interactive agents and election simulations in US and European multiparty settings.
- Study contribution: The study targets a harder non-US multiparty election, elicits full party-choice probability distributions, uses soft voting aggregation, and validates vote shares alongside covariance, ideology, and polling-agency structure.These design choices distinguish the framework from discrete-response approaches and broaden validation beyond aggregate outcomes.
- Epistemic scope: The framework asks which sociopolitical regularities LLMs have absorbed through methodological reconstruction rather than claiming that LLMs can replace surveys.This explicitly limits the exercise’s intended interpretation.
- Context and bias: Beyond the US two-party system, evidence shows demographic prompting performs unevenly across European countries and languages, while political orientation varies by language and model scale.These findings motivate testing LLM electoral reconstruction in a Czech, multilingual, multiparty context.
3 Data
The study uses a representative Czech survey of 3,880 respondents with extensive sociodemographic, attitudinal, and political-behaviour data. Simulated elections are evaluated against both official results and self-reported voting, using ten profile features while excluding specific political stances.
- Dataset: The primary dataset is the representative “Society of Distrust” survey of N = 3,880 Czech adults, containing more than 130 questions.The survey covers sociodemographics, institutional trust, media consumption, conspiracy susceptibility, political attitudes, and self-reported voting in the 2021 parliamentary election.
- Evaluation targets: Evaluation uses two ground-truth levels: official Czech Statistical Office results and survey-based self-reported vote choice.This enables triangular comparison among simulated outcomes, claimed voter behaviour, and actual election results at regional and district levels.
- Input features: Simulated elections are conditioned on 10 variables selected for dataset availability and established theoretical relevance to voting behaviour.The feature set includes demographics, geography, socioeconomics, subjective living standard, and interest in politics.
- Input features: Subjective living standard and interest in politics are included as attitudinal predictors alongside structural sociodemographic indicators.Subjective economic evaluations are motivated by their established value for individual-level vote-choice prediction and populist-party support.
- Input features: Specific political stances, including attitudes toward NATO and the Russo-Ukrainian war, are deliberately excluded from the inputs.The design therefore focuses prediction on sociodemographic and general background profiles rather than explicit issue positions.
4 Methods
The method prompts LLMs with Czech-language sociodemographic respondent profiles and elicits turnout and conditional party-choice probabilities. Individual predictions are aggregated through soft voting, evaluated against election or survey ground truth, and compared with prior-based baselines and political inter-party dependencies.
- Respondent simulation: Each synthetic respondent received a natural-language prompt encoding selected sociodemographic variables and ending with a 2021 parliamentary-election vote-choice task.Simulations were conducted entirely in Czech, and each query was stateless and independent.
- Structured elicitation: Models returned structured JSON containing turnout probabilities and a conditional probability distribution over party choices.Outputs were validated against a Pydantic schema using the API’s Structured Outputs feature.
- Aggregation: Soft voting assigned each respondent’s party weight as P(p | i) · P(voted | i), preserving uncertainty when aggregating predicted vote shares.Aggregate party shares were computed by normalizing summed weights against the total weight of all parties.
- Evaluation: Mean Absolute Error compared simulated vote shares with official election results or self-reported survey data.The analysis also compared covariance and correlation matrices of simulated and official inter-party vote-share relationships.
- Baseline priors: Two prior strategies—Citizen Prior and Direct Prior—were evaluated nationally and regionally to separate persona-derived signal from underlying model knowledge.Each prior prompt was repeated N = 10 times, and the final prior averaged the returned probability distributions.
5 Results
The simulations closely reproduce aggregate Czech election outcomes while recovering turnout, sociodemographic, regional, and political-bloc structure. Smaller models more often produce absolute abstention and undefined party preferences, whereas larger models differentiate profiles more effectively.
- Turnout and party preferences: GPT-4.1-nano assigned exactly P(voted) = 0 to 61.37% of simulated respondents, unlike GPT-4o and GPT-4.1, which never assigned absolute zero turnout probability.The larger models maintained continuous uncertainty, while smaller architectures frequently expressed certainty that respondents would abstain.
- Turnout and party preferences: Smaller models’ undefined party preferences were fundamentally linked to absolute non-voter predictions, assigning zero probability to all parties for respondents predicted not to vote.The overlap analysis indicates that the turnout and party-preference outputs were functionally coupled.
- Aggregate and regional accuracy: Around 2.5 percentage points MAE was achieved by the best-performing models, with simulated vote shares closely matching official election results and GPT-4o models comparable to major Czech polling agencies.GPT-4o results fell within the range of late pre-election estimates reported by STEM, Median, and Kantar CZ.
- Aggregate and regional accuracy: Regional MAE remained relatively stable across Czech regions, with slightly higher errors where turnout was more volatile.The findings suggest demographic conditioning captures broad geographic structure but remains sensitive to turnout uncertainty.
- Prompt-information validation: All prompt input variables had statistically significant small-to-large MANOVA effects, while the unused “voted party” variable had large p-values in small models.The negative-control result indicates weaker, heterogeneous, or absent effects for that variable in the small models.
- Political structure: The models recovered two dominant blocs—SPOLU–PirSTAN and ANO–SPD with positive association to KSˇCM—mirroring correlations in real municipal election data without electoral statistics.GPT-4o showed SPOLU–PirSTAN correlation 0.76, ANO–SPD correlation 0.65, and ANO–KSˇCM correlation 0.53; GPT-4o-mini recovered the same qualitative structure with weaker magnitudes.
6 Discussion
Demographic conditioning adds interpretable structure to LLM outputs, enabling exploration of encoded demographic-to-political associations rather than improving national-level prediction consistently. The framework remains exploratory and limited by noncausal outputs, prompt and aggregation assumptions, model bias, generalisability, and ethical misuse risks.
- Encoded structure: Demographic conditioning produces systematically different probability distributions across profiles; education and age show the largest effects across all four models.The effects align with the documented education-based political cleavage, while predicted-vote correlations recover two political blocs.
- Prediction versus exploration: GPT-4.1 achieves MAE of 0.5 when directly asked for the 2021 Czech election result, so demographic simulation does not always improve national accuracy.All models possess substantial prior knowledge of the election, likely retrievable from training data.
- Prediction versus exploration: The simulation generates differentiated predictions across 3,880 individual profiles, supporting regional decomposition, inter-party correlation analysis, and diagnosis of influential sociodemographic variables.Its primary value is exploring sociological regularities encoded by the model and demographic-to-political mappings.
- Exploratory use: Controlled profile contrasts can manipulate one demographic variable while holding others constant, allowing researchers to examine shifts in predicted vote distributions.This enables systematic exploration of profiles absent from the survey data, such as differences in political interest or education.
- Epistemic status: Model outputs represent encoded associations and socially plausible simulations, not causal explanations, forecasts, or individual decision-making processes.The framework is intended for hypothesis generation and interrogation of regularities absorbed from training data, rather than replacing survey research.
- Limitations and ethics: Key limitations include prompt framing effects, potentially correlated aggregation errors, liberal high-income bias, restricted generalisability, temporal uncertainty, and potential political microtargeting misuse.The study covers one Czech election and one provider; responsible, transparent use is therefore advocated.
7 Conclusions
The paper presents a methodological framework that uses LLMs as implicit sociological models to reconstruct aggregate electoral behaviour from sociodemographic profiles. Its results support exploratory computational-social-science uses while distinguishing latent-pattern reconstruction from causal explanation, forecasting, or individual-level inference.
- Methodological contribution: The framework conditions LLMs on individual-level sociodemographic profiles and aggregates probabilistic outputs through soft voting to reconstruct aggregate political behaviour.It demonstrates reconstruction of key real-world electoral features without access to electoral statistics.
- Empirical validation: Using the 2021 Czech parliamentary election, the framework reproduced official vote shares with low aggregate error, recovered political blocs, and matched established sociodemographic voting gradients.These results arise from statistical regularities encoded in large-scale textual data rather than explicit political knowledge or causal reasoning.
- Scope and limits: The contribution is methodological rather than predictive: the approach does not explain voting causes, provide real-time forecasts, or support individual-level inference.It instead offers a systematic way to interrogate latent social knowledge in language models and assess correspondence with observed social structures.
- Applications and future work: Within epistemic and ethical boundaries, LLM-based demographic simulation may support hypothesis generation, survey robustness checks, and comparative analyses across models or contexts.Future work should replicate the approach across countries and electoral systems, test prompt sensitivity, and investigate the sources and limits of sociological knowledge encoded in LLMs.
Appendix A Prompt Template
Appendix A defines a neutral natural-language prompt template that encodes each respondent’s sociodemographic profile before eliciting their reported 2021 parliamentary vote.
- Prompt construction: Each respondent’s prompt encoded gender, age, education, region, district, municipality size, employment, household income, living standard, and political interest.These attributes were inserted into a first-person demographic profile.
- Voting elicitation: The template then asked which party the respondent voted for in the 2021 parliamentary elections.The voting question followed the sociodemographic profile.
- Neutral wording: All categorical values were rendered in plain, neutral language to minimize stylistic cues.The wording was intended to reduce stylistic influences on responses.
Appendix B Output Schema
The appendix specifies a structured-output schema separating turnout probabilities from party preferences conditional on voting. Models assign probabilities across nine party-choice categories, with probabilities summing to one within each component.
- Output components: Models returned structured outputs containing turnout probabilities and a conditional distribution over party choices.The turnout component comprised P(voted) and P(not voted), while party preferences were conditioned on voting.
- Normalization: Probabilities had to sum to one within each output component.This normalization applied separately to the turnout and conditional party-choice components.
- Party categories: The party-choice distribution covered ANO, SPOLU, PirSTAN, SPD, Pří̈saha, ČSSD, KSČM, Trikolóra, and Other.The distribution was represented as P(p | voted).
Appendix C Soft Voting Aggregation
The appendix defines soft voting aggregation by indexing respondents and political parties, assigning weighted votes, and computing aggregate party support. This preserves individual-level uncertainty while yielding stable aggregate estimates.
- Respondents and political parties are indexed as i and p, respectively, and each respondent contributes a weighted vote.
- Aggregate party support is computed from the respondents’ weighted contributions.
- The approach preserves individual-level uncertainty while producing stable aggregate estimates.
Appendix D Ethical and Methodological Scope
The methodology is an aggregate-level exploratory and descriptive instrument, not a tool for individual inference, prediction, or persuasion. Its outputs reflect training-data regularities, may reproduce historical biases, and should complement rather than replace survey research.
- Ethical and Methodological Scope: The methodology is not intended for individual-level inference, prediction, or persuasion.Analyses are conducted at the aggregate level and evaluated against publicly available benchmarks.
- Ethical and Methodological Scope: Simulated outputs reflect statistical regularities in training data and may reproduce historical biases.
- Ethical and Methodological Scope: The method is exploratory and descriptive, complementing—not replacing—traditional survey-based research.