Performance and biases of Large Language Models in public opinion simulation

Society, culture, and collective behavior2024NatureApproved editorial review

Original title: Performance and Biases of Large Language Models in Simulating Public Opinion

Authors: Yao Qu, Jue Wang

Keywords: Large Language Models, Public opinion simulation, World Values Survey, Bias, Cultural differences

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

2
Authors
22
Findings
42
Limitations
9
Evidence

Editorial summary

English

The article evaluates whether GPT-3.5 Turbo can predict individual World Values Survey responses and whether agreement varies by country, demographics, topic, and number of response options. It uses Wave 6 (2010–2014) for the United States, Japan, Singapore, Brazil, South Africa, and Sweden. For each respondent, observed sex, age, education, class, ethnicity, and covariates are converted into a verbal profile; a target question is added, and the model is asked to reason before selecting an option, to avoid politically correct answers, and to respond in the questionnaire language for Sweden, Brazil, and Japan. The paper identifies the model only as GPT-3.5 Turbo, sets temperature to 0.2, and reports 100 simulations per sample or respondent, but it releases no snapshot, API date, analytic sample size by country or subgroup, total request count, complete prompts, generations, or code. The primary task compares a synthetic answer with the real answer in the same row using Cohen's kappa, Cramer's V, and raw agreement. In the main text, mean country kappas are 0.239 for the United States, 0.165 for Sweden, 0.063 for Singapore, 0.034 for Brazil, 0.024 for Japan, and 0.006 for South Africa: even the best result is low-to-modest agreement, not accurate recovery of public opinion. The supplement reports different values, 0.257, 0.134, 0.059, 0.039, 0.026, and 0.001, respectively, without reconciling them. From only six country observations, the authors correlate kappa with three binary indicators and report 0.971 for Western culture, 0.557 for developed economy, and 0.101 for English language. No intervals, p values, or multivariable model are provided, and culture overlaps with development; the coefficients also cannot be reconstructed exactly from either the rounded figure values or the supplementary table. These are exploratory six-case associations, not evidence that culture, development, or language causes the differences. For the United States, the supplement reports higher kappa for men than women (0.275 versus 0.257), White than Black respondents (0.274 versus 0.079), ages 65+ than 18–29 (0.310 versus 0.134), upper than working class (0.371 versus 0.172), and university than non-university education (0.293 versus 0.172). Subgroup sizes, uncertainty, declared weighting, and tests are absent; kappa also varies with prevalence and category composition, so these gaps are descriptive signals rather than adjusted bias estimates. In U.S. topic analyses, the environmental item reaches kappa 0.270 with five covariates and zero without them; the political item reaches 0.306 with ideology and 0.145 without it. This does not isolate topic difficulty because the question, outcome, and covariate set all change. The model generates 6.10 percentage points fewer responses defined as liberal for the environmental item and 16.33 points more for political voting, demonstrating topic dependence rather than one stable ideological leaning. Reducing voting from four options to two raises kappa from 0.109 to 0.306 and agreement from 29.21% to 65.92%, but it also excludes responses and changes prevalence, so option count is not isolated as the cause. A control changes only the country label for U.S. Wave 6 data to Japan and reduces kappa to 0.057; this establishes sensitivity to the label, not fidelity to Japan. U.S. Wave 7 yields 0.379 versus 0.306 for Wave 6, a difference the paper calls consistency without an equivalence analysis. The defensible contribution is evidence that profile-conditioned survey simulation has unequal and generally limited agreement whose errors depend on context and subgroup. It does not support replacing surveys, causally locating failures in training data, or treating generations as real public opinion.

Español

El artículo evalúa si GPT-3.5 Turbo puede predecir respuestas individuales de la World Values Survey y si su concordancia varía por país, demografía, tema y número de opciones. Utiliza Wave 6 (2010–2014) para Estados Unidos, Japón, Singapur, Brasil, Sudáfrica y Suecia. Para cada encuestado convierte sexo observado, edad, educación, clase, etnia y covariables en un perfil verbal; añade una pregunta objetivo y pide al modelo que razone antes de elegir una opción, que evite respuestas políticamente correctas y que responda en la lengua del cuestionario para Suecia, Brasil y Japón. El paper identifica el modelo solo como GPT-3.5 Turbo, fija temperatura 0,2 y declara 100 simulaciones por muestra o encuestado, pero no publica snapshot, fecha de API, tamaño analítico por país o subgrupo, número total de peticiones, prompts completos, salidas ni código. La tarea principal compara la respuesta sintética con la respuesta real de la misma fila mediante Cohen kappa, Cramer V y acuerdo bruto. En el texto, las kappas medias por país son 0,239 para Estados Unidos, 0,165 para Suecia, 0,063 para Singapur, 0,034 para Brasil, 0,024 para Japón y 0,006 para Sudáfrica: incluso el mejor resultado representa concordancia baja o modesta, no una reproducción precisa de la opinión pública. El suplemento publica valores distintos, 0,257, 0,134, 0,059, 0,039, 0,026 y 0,001, respectivamente, sin reconciliarlos. A partir de solo seis observaciones nacionales, los autores correlacionan kappa con tres indicadores binarios y comunican 0,971 para cultura occidental, 0,557 para economía desarrollada y 0,101 para idioma inglés. No presentan intervalos, valores p ni un modelo multivariable, y cultura y desarrollo se solapan; además, esos coeficientes no se reconstruyen exactamente con los valores redondeados de la figura ni con la tabla suplementaria. Por tanto, muestran asociación exploratoria con seis casos, no que cultura, desarrollo o idioma causen las diferencias. En Estados Unidos, el suplemento informa mayor kappa para hombres que mujeres (0,275 frente a 0,257), White que Black (0,274 frente a 0,079), edades 65+ que 18–29 (0,310 frente a 0,134), clase upper que working (0,371 frente a 0,172) y educación universitaria que no universitaria (0,293 frente a 0,172). No hay tamaños de subgrupo, incertidumbre, ponderación declarada ni pruebas; kappa también cambia con prevalencia y composición de categorías, de modo que estas brechas son señales descriptivas, no estimaciones ajustadas de sesgo. En los análisis temáticos de EE. UU., la pregunta ambiental obtiene kappa 0,270 con cinco covariables y 0 sin ellas; la política obtiene 0,306 con ideología y 0,145 sin ella. La comparación no aísla la dificultad del tema porque cambia simultáneamente la pregunta, el resultado y el conjunto de covariables. El modelo genera 6,10 puntos porcentuales menos de la opción definida como liberal en ambiente y 16,33 puntos más en voto político, lo que evidencia dependencia temática y no una inclinación ideológica única. Reducir voto de cuatro opciones a dos eleva kappa de 0,109 a 0,306 y acuerdo de 29,21% a 65,92%, pero también excluye respuestas y cambia prevalencias, por lo que no demuestra que el número de opciones sea la única causa. Un control cambia solo el país del prompt de Estados Unidos a Japón y reduce kappa a 0,057; esto demuestra sensibilidad a la etiqueta nacional, no fidelidad a Japón. Wave 7 de Estados Unidos produce 0,379 frente a 0,306 en Wave 6, diferencia que el artículo denomina consistencia sin prueba de equivalencia. La contribución defendible es documentar que una simulación de encuesta condicionada por perfiles tiene concordancia desigual y generalmente limitada, y que sus errores dependen del contexto y del subgrupo. No sustenta sustituir encuestas, atribuir causalmente los fallos al entrenamiento ni usar estas generaciones como opinión pública real.

Research question

With what fidelity does GPT-3.5 Turbo reproduce responses from the World Values Survey, how does that agreement vary across six countries and US subgroups, and how does it change according to topic and number of options?

Method

Individual rows from WVS Wave 6 are transformed into persona prompts with demographic variables and covariates, 100 responses are generated per case at temperature 0.2, and they are compared with the human response using Cohen kappa, Cramer V, and agreement. Six countries, US subgroups, environmental and political questions, and vote versions with two or four options are contrasted. The editorial audit read and rendered the 13 pages and the six pages of the DOCX supplement, verified tables, recalculated the logic of the six national cases, and searched for public data and code.

Sample: Six countries from WVS Wave 6 and a supplementary analysis of the US Wave 7. The article states that it performs 100 simulations per sample or respondent, but does not report how many rows survive selection and complete data in each country, topic, or subgroup, nor the total number of generations. Without those N values, coverage, weighting, precision, or stability of the kappas cannot be evaluated.

Findings

  • The version of record was published on 28 August 2024 in Humanities and Social Sciences Communications 11, article 1095.
  • The analysis uses WVS Wave 6 from six countries collected between 2010 and 2014.
  • GPT-3.5 Turbo is run at temperature 0.2, but the snapshot and the API date are not identified.
  • Each response is conditioned on demographics, covariates, and country, and 100 simulations are requested per case.
  • The main text reports kappa 0.239 for the United States, the best country of the six.
  • Sweden, Singapore, Brazil, Japan, and South Africa show approximate kappas of 0.165, 0.063, 0.034, 0.024, and 0.006 in the main figure.
  • The supplementary kappas do not match: 0.257, 0.134, 0.059, 0.039, 0.026, and 0.001 for those same countries.
  • The published kappa correlations are 0.971 with Western culture, 0.557 with developed economy, and 0.101 with English language.
  • These correlations are calculated with only six countries and three binary predictors that are partially confounded.
  • In the US, the supplement reports kappa 0.275 for men and 0.257 for women.
  • By ethnicity, the kappas are 0.274 White, 0.079 Black, 0.321 Other, 0.242 Hispanic, and 0.218 for two or more races.
  • By age, kappa increases from 0.134 at 18-29 to 0.310 at 65+.
  • By class, kappa is 0.371 upper, 0.331 upper middle, 0.281 lower middle, 0.172 working, and 0.196 lower.
  • College education yields kappa 0.293 compared with 0.172 without college, although raw agreement is 65.16% versus 65.44%.
  • The environmental question reaches kappa 0.270 with five covariates and 0 without them.
  • The political question reaches kappa 0.306 with ideology and 0.145 without covariate.
  • The synthetic liberal proportion differs by -6.10 points on environment and +16.33 on politics relative to WVS.
  • Reducing the political vote from four to two options raises kappa from 0.109 to 0.306 and agreement from 29.21% to 65.92%.
  • Changing the national label of US rows to Japan reduces kappa from 0.306 to 0.057.
  • The US Wave 7 comparison obtains kappa 0.379, Cramer V 0.415, and agreement 67.65%.
  • The generated and analyzed data are only available upon request to the corresponding author.
  • No public repository of code or outputs associated with the article was located.

Limitations

  • Only one model family and one unidentified version of GPT-3.5 Turbo is evaluated.
  • The API query date and the returned snapshot are not reported.
  • Complete parameters, seed, top_p, token limits, system prompt, or retry logic are not published.
  • The instruction to produce a chain of reasoning does not validate that the process mimics human reasoning.
  • Asking for non-politically correct responses introduces a demand signal that may alter choices.
  • The complete prompt is not published for each language, country, and experiment.
  • It is not documented how the option is extracted when the reasoning is ambiguous or invalid.
  • Failures, exclusions, out-of-format responses, or retry rates are not reported.
  • Analytical sizes are not reported by country, subgroup, question, or condition.
  • It is not specified whether WVS sample weights are used or how missing values are treated.
  • WVS responses come from different years between 2010 and 2014 and may reflect temporal changes.
  • The prompts use different languages, so country, language, translation, and cultural context are confounded.
  • Political questions have country-specific options and are not equivalent results.
  • Marginal distributions of responses differ by country and affect kappa.
  • There are only six observations for the country-factor correlations.
  • Two countries are classified as Western, four as developed, and three as English-speaking.
  • Western culture and developed economy are correlated and are not jointly adjusted.
  • The correlations do not include intervals, tests, sensitivity, or correction for multiple comparisons.
  • The heatmap coefficients are not exactly reconstructed from the published rounded kappas.
  • The claim of better performance in English-speaking countries is weak given a correlation of 0.101 and the low South African result.
  • The text and the supplement publish different country values without explanation.
  • Cohen kappa is not directly comparable across groups with different prevalences and numbers of categories.
  • Demographic gaps lack subgroup sizes, intervals, and tests.
  • The Other and two or more races ethnic categories are heterogeneous.
  • V240 records sex by interviewer observation, although the text interprets it as gender.
  • There is no intersectional analysis or adjustment for correlation among age, education, class, and ethnicity.
  • Comparing environment and politics simultaneously changes the question, the scale, and the covariates.
  • Higher political kappa does not demonstrate that political behavior is intrinsically easier to simulate.
  • Defining Democrat and environmental protection as the only liberal options simplifies ideological constructs.
  • The normalization of the liberal difference is not sufficiently explained and no uncertainty is provided.
  • The comparison of four versus two options filters categories and respondents, in addition to changing complexity.
  • The Japan label applied to US data tests sensitivity to the prompt, not cultural accuracy.
  • Calling Wave 6 and Wave 7 consistent is not supported by a test of equivalence or individual stability.
  • Wave 6 and Wave 7 may change in composition, covariates, and political climate.
  • There is no preregistration or formal separation between confirmatory and exploratory analyses.
  • There is no comparison with simple statistical models or majority-class baselines.
  • GPT-3.5 is not compared with other LLMs or with a model trained on WVS.
  • There is no evaluation of calibration, aggregate distribution, per-class error, or predictive utility beyond kappa and association.
  • No code, processed data, prompts, outputs, or reproducible environment are released.
  • Availability upon request does not allow independent current verification.
  • Explanations based on training composition are speculative because that corpus is neither observed nor intervened upon.
  • Implications for public policy exceed an evaluation of two survey questions.

What the study does not establish

  • It does not demonstrate that GPT-3.5 reproduces public opinion with high fidelity for any country.
  • It does not demonstrate that culture, development, or language cause the national differences.
  • It does not prove that the training corpus is the responsible mechanism.
  • It does not establish that kappa gaps are adjusted demographic effects.
  • It does not demonstrate a stable liberal or conservative inclination across topics.
  • It does not isolate the number of options as the cause of the drop in agreement.
  • It does not validate the reasoning chains as human psychological processes.
  • It does not demonstrate temporal stability between Wave 6 and Wave 7.
  • It does not generalize to other models, surveys, questions, countries, or dates.
  • It does not justify replacing human respondents with synthetic samples.
  • It does not demonstrate that the simulations are adequate for policymaking.
  • It does not allow reproducing the results with the available public artifacts.

Traceability

Scope: Full text

Version: Humanities and Social Sciences Communications 11, article 1095; published 28 August 2024; DOI 10.1057/s41599-024-03609-x; CC BY-NC-ND 4.0

Consulted source: https://www.nature.com/articles/s41599-024-03609-x.pdf

Review: Codex full-text, visual, supplementary-table and statistical-reporting audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • OpenAI GPT-3.5 Turbo, exact model snapshot and API date not reported

Instruments and metrics

  • World Values Survey Wave 6 item V81 on environment versus economic growth
  • World Values Survey Wave 6 item V228 on national election vote
  • Demographic profile from V254 ethnicity, V240 observed sex, V242 age, V248 education and V238 social class
  • Environmental covariates V30, V78, V82, V83 and V122
  • Political ideology covariate V95
  • Interview-style persona prompts in English, Swedish, Portuguese and Japanese
  • Prompted chain of reasoning and instruction to avoid politically correct answers
  • Cohen's kappa
  • Cramer's V
  • Proportion agreement
  • Pearson correlations over six country-level observations

Data used

  • World Values Survey Wave 6 country files for Japan 2010, United States 2011, Sweden 2011, Singapore 2012, South Africa 2013 and Brazil 2014
  • World Values Survey Wave 7 United States 2017 robustness comparison
  • Unreleased GPT-3.5 Turbo simulated responses
  • Publisher supplementary DOCX with Tables S1–S4

Evidence and location

  • Scope, WVS source, and variables: Version of record pp. 1-3, Abstract and Materials and methods
  • Model, temperature, prompts, and 100 simulations: Version of record pp. 3-5, Simulation process and Data analysis
  • Country results and correlations of six cases: Version of record pp. 5-7, Figures 2-4 and Tables 1-2
  • US demographic gaps: Version of record pp. 7-8, Figure 5; Supplement Table S2
  • Topic, ideology, and number of options: Version of record pp. 8-9, Tables 3-5
  • Country controls and Wave 7: Version of record p. 10; Supplement Table S4
  • Acknowledged limitations and conclusions: Version of record pp. 11-12, Limitations and Conclusion
  • Metric discrepancies and absence of N: Main Figures 2-3 cross-checked against frozen Supplement Tables S1-S4, 15 Jul 2026
  • Artifact availability: Publisher Data availability statement and title/author searches of GitHub, OSF and Zenodo, checked 15 Jul 2026