Generative AI voting: fair collective choice is resilient to LLM biases and inconsistencies

Society, culture, and collective behavior2026link.springer.comApproved editorial review

Authors: Srijoni Majumdar, Edith Elkind, Evangelos Pournaras

Keywords: Voting, Generative AI, Large language models, Collective decision making, Social choice, Proportional representation, Participatory budgeting, Voter turnout

Source: Open primary source (opens in a new tab)

3
Authors
9
Findings
20
Limitations
5
Evidence

Editorial summary

English

The paper asks what happens when an LLM generates a person's ballot from demographic, political, and preference traits, and whether voting design can absorb the resulting errors. Its strongest result is not that AI represents individuals accurately: it performs very poorly on complex ballots. Mean individual human-AI consistency is 5.68% for the actual Aarau vote and 28.005% for its hypothetical survey, versus 84.5% in binary US elections. The useful contribution is different: after thousands of ballots are aggregated, some winner sets are much more stable than the individual predictions, and that stability depends strongly on the aggregation rule.

The study combines three observed settings. ANES supplies the 2012, 2016, and 2020 US presidential elections: the paper begins with 20,650 respondents, retains 17,010 complete profiles, and uses single-choice ballots with majority aggregation. City Idea, a 2023 participatory-budgeting campaign in Aarau, supplies a 3,314-person survey over five projects using single choice, approval, and score ballots, plus an actual 1,703-citizen vote over 33 projects using cumulative ballots; 505 voters have linked traits for emulation. The design adds correctly to 21, 207, and 135 emulated elections, 363 in total.

The study labels six LLMs, GPT-4o Mini, GPT-3.5, GPT-3, DeepSeek R1, Gemini 1.5 Flash, and Llama 3 8B, and one predictive neural network. Prompts are zero-shot, provide election context and personal traits, and request step-by-step reasoning. The supplement says temperatures from 0.4 to 0 were tested with 20 runs per setting and mean consistency across runs. It does not provide a manifest linking each result to a checkpoint, provider, endpoint, parameters, temperature, repetition, exact prompt, parser, or failure. The June-November 2023 and April-August 2025 collection windows are not mapped model by model. Generic model labels therefore do not identify reproducible systems.

The main metric counts relative matches in Condorcet pairwise comparisons between two ballots. Cross-format comparisons convert scores and cumulative votes to binary representations and exclude ties; collective outcomes become winner/loser vectors. This is a defensible ordinal-similarity measure, but it does not validate shared reasons, values, or a voter's mandate. Conversion discards intensity and ties and can change the denominator. In particular, the method of equal shares often funds more low-cost projects than utilitarian greedy, mechanically increasing overlap between winner sets. The authors recognize this: equal shares and Phragmén average 69.9% collective consistency, 31.6 percentage points above greedy, but the advantage falls to 22.8 points after equalizing winner counts.

Among the LLMs, GPT-4o Mini has the highest individual consistency and is reported 4.85 points above GPT-3.5 and 7.85 above Llama 3 8B. The neural-network baseline is 1.7 points above the LLM average individually and 2.9 collectively. Human cross-format consistency is 97.1%; model values fall to 74.3% for GPT-4o Mini, 72.1% for GPT-3.5, 76.2% for Llama 3 8B, 71.23% for Gemini, and 68.7% for DeepSeek. Equal shares raises collective cross-format consistency by 12.2 points over greedy. These are descriptions of this simulation, not stable capability rankings of model families.

The low-turnout experiment is retrospective. In a subset of 252 Aarau participants who completed both pre- and post-vote surveys, overlapping groups with low engagement, digital literacy, or trust are treated as likely abstainers. The analysis then removes ballots from people who actually participated and tries to recover the full-turnout result using their profiles and synthetic ballots. With up to 75% of the low-engagement group replaced by GPT-3.5, equal shares recovers up to 50% more consistency than random removal; proportional methods average a 7.53-point advantage, or 6.72 after controlling winner counts. Yet the supplement also shows that equal shares retains roughly 83-84% of winners with 80% simulated abstention. A substantial part of the resilience belongs to the rule itself, not AI representation.

This is not a study of actual abstainers. It does not observe an abstention decision, frustrated intent to vote, or consent to delegate. Targets are selected correlates among observed participants, and the reference outcome uses those same people's real ballots. The evidence therefore does not show that an LLM knows a nonvoter's preference, can recover a real election, or has democratic authorization. The 2.1-point GPT-4o Mini versus GPT-3.5 contrast is also nonsignificant, p=0.092.

Statistical inference requires substantial caution. The paper uses Fisher's method to combine p-values from runs with different hyperparameters. These tests reuse the same voters, alternatives, and outcomes and are dependent; the number and identity of component tests, base statistics, complete grid, exclusion rules, and global correction across many comparisons are not reported. Most claims provide combined p-value thresholds without paired effect intervals. This is descriptive evidence, not independent replication across models or elections.

Causal language is also excessive. A recurrent network predicts discretized consistency from personal traits, while SHAP, LIME, and ablation rank predictive feature contributions. This does not identify causes: there is no intervention, causal graph, identification strategy, or sufficient control of correlated traits. Post hoc mappings to affect, unconscious, time-discounting, or surrogation bias do not use validated instruments for those constructs. The visible materials do not document a reproducible person-level split, leakage controls, external validation, or complete uncertainty for the RNN and fairness pipeline.

Public reproducibility is partial and belongs to an earlier stage. The official Figshare collection, created in June 2024, contains three ANES workbooks with 5,914, 5,272, and 6,031 rows, 17,217 total. They include profiles, prompts, human choice, GPT-3.5 at temperature 0.2, Llama 2 output, and an ML prediction. They do not match the final 17,010 subset and omit Llama 3, GPT-4o Mini, the other models, Aarau, repetitions, scripts, parsing, aggregation, abstention, RNN, figures, and environments. Internal arithmetic raises another flag: 3,640 incomplete profiles times the three named models would be 10,920 representatives, but the paper reports 7,280, exactly two models. No associated public code repository was found.

The public workbooks embed prompt profiles combining race, gender, age, ideology, party, religion, political interest, state, and inferred vote. No obvious direct identifiers were found, but the combination is sensitive and needs explicit minimization, governance, and linkage-risk documentation. The authors do discuss autonomy, privacy, consent, and accountability and do not simply advocate replacing citizens.

A faithful reading preserves an important conclusion narrower than the title. In these retrospective data, LLMs reproduce complex individual ballots poorly, while proportional rules such as equal shares and Phragmén can make aggregate outcomes more stable under synthetic ballots and simulated turnout loss. Aggregation design is therefore a relevant safeguard to investigate. The study does not establish faithful representation of abstainers, demographic fairness of AI agents, causal democratic resilience, causal cognitive biases, or full reproducibility of the final experiment.

Español

El artículo pregunta qué ocurre cuando un LLM genera el voto de una persona a partir de rasgos demográficos, políticos y de preferencias, y si el diseño de la votación puede amortiguar sus errores. Su resultado más sólido no es que la IA represente bien a cada votante: en los formatos complejos lo hace muy mal. La consistencia individual media entre voto humano y sintético es 5,68 % en la votación real de Aarau y 28,005 % en su encuesta hipotética, frente a 84,5 % en las elecciones estadounidenses binarias. La aportación útil es otra: al agregar miles de votos, algunos conjuntos de ganadores son bastante más estables que las predicciones individuales, y esa estabilidad depende mucho de la regla de agregación.

El estudio combina tres escenarios observados. ANES aporta las elecciones presidenciales de Estados Unidos de 2012, 2016 y 2020: el texto parte de 20.650 encuestados, conserva 17.010 perfiles completos y usa voto único con mayoría. City Idea, una campaña de presupuesto participativo realizada en Aarau en 2023, aporta una encuesta de 3.314 personas sobre cinco proyectos con voto único, aprobación y puntuación, y una votación real de 1.703 ciudadanos sobre 33 proyectos con papeleta acumulativa; 505 votantes tienen rasgos enlazados para la emulación. El diseño suma correctamente 21, 207 y 135 elecciones emuladas, 363 en total.

Se etiquetan seis LLM, GPT-4o Mini, GPT-3.5, GPT-3, DeepSeek R1, Gemini 1.5 Flash y Llama 3 8B, y una red neuronal predictiva. Los prompts son zero-shot, incluyen el contexto electoral y rasgos personales, y piden razonamiento paso a paso. El suplemento dice que se probaron temperaturas de 0,4 a 0 y 20 ejecuciones por ajuste, promediando la consistencia. Sin embargo, no publica un manifiesto que asocie cada resultado con checkpoint, proveedor, endpoint, parámetros, temperatura, repetición, prompt exacto, parser o fallo. Las dos ventanas de recogida, junio-noviembre de 2023 y abril-agosto de 2025, tampoco se vinculan modelo por modelo. Por ello, los nombres genéricos no permiten reconstruir qué sistemas respondieron.

La métrica principal cuenta coincidencias relativas en comparaciones de Condorcet entre dos papeletas. Para comparar formatos, convierte puntuaciones y votos acumulativos a representaciones binarias y excluye empates; para resultados colectivos, convierte proyectos en ganadores o perdedores. Es una medida razonable de similitud ordinal, pero no valida que el agente comparta las razones, valores o mandato de la persona. La conversión descarta intensidad y empates y puede alterar el denominador. En especial, el método de equal shares suele financiar más proyectos baratos que utilitarian greedy, lo que eleva mecánicamente la coincidencia de conjuntos de ganadores. Los autores reconocen el problema: equal shares y Phragmén promedian 69,9 % de consistencia colectiva, 31,6 puntos más que greedy, pero la ventaja baja a 22,8 puntos al igualar el número de ganadores.

Entre los LLM, GPT-4o Mini obtiene la mayor consistencia individual y supera en 4,85 puntos a GPT-3.5 y en 7,85 a Llama 3 8B. La red neuronal queda 1,7 puntos por encima del promedio de LLM a nivel individual y 2,9 a nivel colectivo. La consistencia entre formatos humanos es 97,1 %; en los modelos baja a 74,3 % para GPT-4o Mini, 72,1 % para GPT-3.5, 76,2 % para Llama 3 8B, 71,23 % para Gemini y 68,7 % para DeepSeek. Equal shares aumenta 12,2 puntos la consistencia colectiva entre formatos frente a greedy. Estos contrastes describen esta simulación; no son pruebas generales de capacidad de cada familia.

El experimento de baja participación es retrospectivo. En un subconjunto de 252 participantes de Aarau que completó las encuestas previa y posterior, se definen como probablemente abstencionistas grupos con bajo compromiso, baja alfabetización digital o baja confianza, que además se solapan. Después se eliminan votos de personas que realmente sí participaron y se intenta recuperar el resultado de participación completa con sus perfiles y votos sintéticos. Con hasta 75 % del grupo de bajo compromiso sustituido por GPT-3.5, equal shares logra hasta 50 % más recuperación que una eliminación aleatoria; los métodos proporcionales promedian 7,53 puntos más, o 6,72 al controlar ganadores. Pero el propio suplemento muestra que equal shares retiene alrededor del 83-84 % de ganadores incluso con 80 % de abstención simulada. Parte importante de la resiliencia pertenece a la regla, no a la representación artificial.

Esto no estudia a abstencionistas reales. No se observa una decisión de abstenerse, la intención frustrada de votar ni el consentimiento para delegar. Los targets son correlatos elegidos entre participantes observados, y el objetivo de referencia se construye con los votos reales de esas mismas personas. Por tanto, la evidencia no demuestra que un LLM pueda conocer la preferencia de un no votante, recuperar una elección real o poseer mandato democrático. El contraste GPT-4o Mini–GPT-3.5 de 2,1 puntos tampoco es significativo, p=0,092.

La inferencia estadística necesita mucha cautela. El artículo combina mediante Fisher los p-valores de ejecuciones con diferentes hiperparámetros. Esas pruebas reutilizan los mismos votantes, alternativas y outcomes y no son independientes; no se informa cuántas se combinan, cuáles son los tests base, la cuadrícula completa, las exclusiones ni una corrección global para las numerosas comparaciones. La mayoría de afirmaciones ofrece umbrales de p combinados sin intervalos pareados de efecto. Es una señal descriptiva, no una réplica independiente entre modelos o elecciones.

También hay un exceso causal. Una red recurrente clasifica la consistencia discretizada a partir de rasgos personales y SHAP, LIME y ablación ordenan variables predictivas. Eso no identifica causas. No hay intervención, grafo causal, estrategia de identificación ni control suficiente de rasgos correlacionados. La asignación posterior de etiquetas psicológicas como sesgo afectivo, inconsciente, descuento temporal o surrogation tampoco usa instrumentos validados de esos constructos. El material visible no documenta un split reproducible por persona, controles de leakage, validación externa o incertidumbre completa para el RNN y su pipeline de fairness.

La reproducibilidad pública es parcial y corresponde a otra etapa. La colección Figshare oficial, creada en junio de 2024, contiene tres Excel ANES con 5.914, 5.272 y 6.031 filas: 17.217 en total. Incluyen perfiles, prompts, voto humano, GPT-3.5 a temperatura 0,2, salida de Llama 2 y una predicción ML. No coinciden con el subconjunto final de 17.010 ni contienen Llama 3, GPT-4o Mini, los otros modelos, Aarau, repeticiones, scripts, parsing, agregación, abstención, RNN, figuras o entornos. La aritmética interna añade otra alerta: 3.640 perfiles incompletos multiplicados por los tres modelos nombrados serían 10.920 representantes, pero el artículo informa 7.280, exactamente dos modelos. No se encontró repositorio de código asociado.

Los Excel públicos contienen en los prompts combinaciones de raza, género, edad, ideología, partido, religión, interés político, estado y voto inferido. No aparecen identificadores directos obvios, pero la combinación es sensible y exige documentar minimización, gobernanza y riesgo de enlace. Los autores sí plantean autonomía, privacidad, consentimiento y rendición de cuentas y no recomiendan reemplazar sin más a los ciudadanos.

La lectura fiel conserva una conclusión importante y más estrecha que el título. En estos datos retrospectivos, los LLM reproducen deficientemente votos individuales complejos, mientras que reglas proporcionales como equal shares y Phragmén pueden hacer más estable el resultado agregado ante papeletas sintéticas y pérdida simulada de participación. Eso convierte el diseño de agregación en una salvaguarda relevante para investigar. No demuestra representación fiel de abstencionistas, fairness demográfica del agente, resiliencia democrática causal, sesgos cognitivos causales ni reproducibilidad completa del experimento final.

Research question

With what fidelity do six LLMs reproduce human votes in binary elections and participatory budgets with complex ballots, what features predict their inconsistencies, and to what extent do aggregation rules preserve or recover the collective outcome when low turnout and AI representation are simulated?

Method

363 elections are emulated from ANES 2012/2016/2020 and from a survey and a real participatory budget vote in Aarau. The prompts combine personal features, context, alternatives, and format; six LLMs and a neural network generate or predict ballots. Individual and collective similarity is computed through Condorcet comparison matches, also between standardized formats. Abstention is simulated by removing votes from proxy groups and substituting a fraction with AI votes; majority, utilitarian greedy, equal shares, and Phragmén are compared. RNN, SHAP, LIME, ablation, and a fairness pipeline explore feature associations. The audit read and rendered the 22 pages and the 31 of the supplement, inspected the three official Excel files, and searched for public code.

Sample: ANES reports 20,650 people, 17,010 complete profiles, and 3,640 partial ones; City Idea contributes 3,314 respondents across five projects and 1,703 real voters across 33 projects, of which 505 have linked features. The abstention simulation focuses on 252 people with pre- and post-surveys. The design declares 363 emulated elections. The explicit counts of representatives sum to 81,224, but the block of 3,640 partial profiles is arithmetically inconsistent: three models would imply 10,920, not 7,280.

Findings

  • The mean individual consistency is 5.68% in real City Idea, 28.005% in the survey, and 84.5% in U.S. binary elections.
  • GPT-4o Mini leads the LLMs and exceeds GPT-3.5 by 4.85 points and Llama 3 8B by 7.85; the predictive network outperforms the LLMs on average by 1.7 individual points and 2.9 collective points.
  • Consistency between formats is 97.1% in humans and 68.7-76.2% in the informed LLMs.
  • Equal shares and Phragmén average 69.9% collective consistency, 31.6 points above greedy; when winners are equalized the advantage drops to 22.8.
  • Equal shares improves collective consistency between formats by 12.2 points over greedy.
  • Proportional methods recover 7.53 more points of consistency under simulated abstention, or 6.72 when controlling for winners.
  • GPT-4o Mini exceeds GPT-3.5 by 2.1 points in recovery, but the contrast is not significant, p=0.092.
  • Equal shares retains approximately 83-84% of winners with 80% simulated abstention, even before attributing the effect to AI representatives.
  • The final linked public collection is a 2024 ANES artifact with Llama 2, GPT-3.5, and ML, not the complete experiment published in 2026.

Limitations

  • The model labels do not identify checkpoints, providers, endpoints, or reproducible parameters.
  • The two collection windows are not assigned model by model.
  • There is no complete record of prompts, outputs, parsing, retries, failures, temperatures, and repetitions.
  • Condorcet consistency is operational similarity, not representation or mandate validity.
  • Conversion between formats excludes ties and loses preference intensity.
  • Equal shares often produces more cheap winners and mechanically raises agreement; the secondary control only reduces the confound.
  • Fairness names normative properties of the rule, not demographic metrics of AI outcomes.
  • Abstainers are proxies among people who did participate; real abstention or frustrated intent is not observed.
  • The three proxy groups overlap and the main analysis uses only 252 linked respondents.
  • The recovery target is retrospective and uses the known real vote of the same people.
  • Fisher p-values combine dependent runs and hyperparameters over the same data.
  • No base tests, grid, number of p-values, exclusions, or global correction for multiplicity are enumerated.
  • Paired intervals and a complete quantification of uncertainty per voter, election, and model are missing.
  • SHAP, LIME, and ablation on observational data do not identify causes.
  • Cognitive bias labels are post hoc mappings without validated psychological instruments.
  • No reproducible split, leakage controls, or external validation of the RNN and fairness pipeline is documented.
  • The arithmetic of 3,640 partial profiles and three models does not match 7,280 representatives.
  • The public data sum to 17,217 rows and do not match the 17,010 complete cases of the final study.
  • City Idea, five of the six final models, code, reproducible figures, and environments are not published.
  • The public prompts combine sensitive political and demographic attributes with inferred voting without detailed documentation of linkage risk.

What the study does not establish

  • It does not demonstrate that LLMs faithfully represent real abstainers.
  • It does not demonstrate consent, authorization, or democratic mandate to delegate a vote.
  • It does not demonstrate that substituting voters with AI prospectively improves a real election.
  • It does not demonstrate that equal shares makes the AI representative demographically fair.
  • It does not demonstrate that aggregate resilience comes mainly from AI and not from the electoral rule.
  • It does not demonstrate causal relationships between features, cognitive biases, and model decisions.
  • It does not establish independent replication through combined p-values over dependent runs.
  • It does not allow reproducing the final experiment of six LLMs and City Idea with the available Figshare artifact.
  • It does not generalize to non-voters, other countries, other elections, other checkpoints, or future deployments.

Traceability

Scope: Full text

Version: EPJ Data Science 15:24 version of record, published 2026-02-09; 22-page article plus 31-page official supplement

Consulted source: https://link.springer.com/article/10.1140/epjds/s13688-025-00612-3

Review: Codex full-text, 22-page visual, 31-page supplement visual, publication-metadata, arithmetic, metric, aggregation, abstention-validity, causal, statistical, privacy, Figshare-workbook and public-code audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o Mini; snapshot, provider and exact endpoint not reported
  • GPT-3.5; snapshot, provider and exact endpoint not reported
  • GPT-3; snapshot, provider and exact endpoint not reported
  • DeepSeek R1; snapshot, provider and exact endpoint not reported
  • Gemini 1.5 Flash; snapshot and exact endpoint not reported
  • Llama 3 8B; checkpoint and serving stack not reported
  • Recurrent neural-network predictive baseline

Instruments and metrics

  • Personalized zero-shot, context-based and chain-of-thought voting prompts
  • Single-choice, approval, score and cumulative ballots
  • Condorcet pairwise-match consistency
  • Kemeny-distance robustness analysis
  • Majority, utilitarian greedy, method of equal shares and sequential Phragmén aggregation
  • Low-engagement, low-digital-literacy and low-trust abstention proxies
  • Fisher combination of p-values across runs and hyperparameters
  • RNN consistency classifier with SHAP, LIME and feature ablation
  • Decision-tree and SMOTE fairness pipeline

Data used

  • American National Election Studies 2012, 2016 and 2020
  • City Idea Aarau 2023 pre-voting survey
  • City Idea Aarau 2023 actual participatory-budgeting vote
  • Linked City Idea pre- and post-voting survey subset
  • Official Figshare ANES-only 2024 artifact, 17,217 rows, older than final study

Evidence and location

  • Design, methods, results, arithmetic, limitations, ethics, and availability: EPJ Data Science 15:24, DOI 10.1140/epjds/s13688-025-00612-3, pp. 1-22
  • Prompts, repetitions, temperatures, robustness, abstention, RNN, fairness, SHAP, and LIME: Official Supplementary Information 13688_2025_612_MOESM1_ESM, pp. 1-31
  • Official metadata, publication year, DOI, license, and editorial dates: https://link.springer.com/article/10.1140/epjds/s13688-025-00612-3
  • Content and divergence of the public 2024 ANES artifact: https://figshare.com/collections/Generative_AI_Voting_-_ANES/7261288
  • Comprehensive audit of aggregation, abstention, causality, statistics, privacy, and reproducibility: reports/verification/article-232-ai-voting-aggregation-abstention-causal-statistical-and-artifact-audit.json