The Effects of Demographic Instructions on LLM Personas

Society, culture, and collective behavior2025ACMApproved editorial review

Authors: Angel Felipe Magnossão de Paula, J. Shane Culpepper, Alistair Moffat, Sachin Pathiyan Cherumanal, Falk Scholer, Johanne Trippas

Keywords: LLM personas, Demographic prompting, Sexism detection, Perspectivist annotation, Krippendorff alpha, EXIST 2023, Gender agreement, Age-group agreement, Prompt sensitivity, Bias evaluation

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

6
Authors
12
Findings
19
Limitations
6
Evidence

Editorial summary

English

This SIGIR 2025 short paper studies whether demographic instructions change five LLMs' agreement with human sexism judgments; it does not assess psychological personality or show that a model adopts an identity. It uses 7,958 bilingual EXIST 2023 tweets: 6,920 training plus 1,038 development items, not the separate 2,076-item test set. Each tweet has six binary labels, three female and three male and two in each 18-22, 23-45, and 46+ group, for 47,748 judgments. GPT-3.5, GPT-4, GPT-4o, Mistral Small 22B, and Qwen2.5-14B produce one label under a baseline prompt and prompts adding only sex or age. The outcome is ratio Krippendorff alpha between the model label and each group's per-tweet human proportion. In the baseline run, every model has higher alpha with the female than male aggregate: 0.415/0.371, 0.365/0.325, 0.228/0.191, 0.353/0.310, and 0.378/0.345. This is differential agreement, not by itself a causal bias, female personality, or normative gold standard. Age has no common direction. Demographic personas have mixed effects: female prompting raises target agreement for GPT-4, GPT-4o, and Mistral; male prompting only for GPT-4 and Qwen. GPT-4 and Qwen improve in all three target-age conditions, GPT-3.5 only for 46+, GPT-4o in none, and Mistral for 18-22 and 23-45 but not 46+. The prose incorrectly says Mistral improves consistently: Table 3 falls from 0.392 to 0.383 for 46+. The defensible conclusion is that a short demographic phrase can move agreement in a model-dependent way and is not reliable mitigation. Prompt selection also limits inference: three candidates were tried on twenty tweets and the one with greatest agreement among three LLMs was selected (75%, versus 70% and 55%); o1-preview then rewrote it. The criterion was cross-model consistency, not human validity, calibration, or bias reduction, with no held-out confirmation. Each final cell has one prediction per tweet; run-to-run, paraphrase, and API variation are unmeasured, Mistral/Qwen modifications are unspecified, and English and Spanish are pooled. Statistical auditing finds a material problem. Demographic rates are paired on the same tweets, yet the code uses an independent t-test and independent ANOVA/Tukey tests. The artifact reproduces p=0.237; respecting pairing, the female-minus-male difference is only 0.00716, but a paired t-test gives p=0.0463 and Wilcoxon p=0.000383. The nominal decision changes without making 0.72 percentage points a large effect. Age differences remain detectable with a paired Friedman check, although post-hoc tests need redesign. The intervals are also not reproducible. The paper claims 10,000 bootstraps and intervals narrower than 0.001, but the public function transposes the 2x7,958 matrix, flattens it, resamples cells, and reshapes it as 7,958x2; the library then sees another reliability structure and tweet pairing is broken. For female versus male, target alpha is 0.4768 and the defective callback 0.000254. In 500 audit resamples, a correct bootstrap gives about [0.4578, 0.4955], while the public method gives a narrow interval around zero. The pre-paper commit calls this function 10,000 times; final outputs are absent, so the exact manuscript run cannot be proven, but if this function was used its intervals do not bound the tables. Distance choice matters too: interval alpha is 0.647 rather than 0.477. The matching first-author repository compiles but is neither linked nor versioned in the paper and lacks the dataset, final predictions, Mistral/Qwen outputs, CSVs, intervals, logs, and tables; its intermediate JSON disagrees with several alphas. There are no tests, CI, tag, or release; Ruff reports 68 findings and the environment omits OpenAI. The API layer catches every exception and returns None without releasing final failure rates. No contamination test exists: the claim that labels were not public is not a guarantee because this experiment uses train and development data distributed by organizers. The tables support a practical warning about simple persona-prompt instability, not identity, empathy, personality, fairness, or generalization.

Español

Este artículo corto de SIGIR 2025 estudia si añadir instrucciones demográficas cambia la concordancia de cinco LLM con juicios humanos sobre sexismo; no evalúa personalidad psicológica ni demuestra que el modelo adopte una identidad. Usa 7.958 tuits bilingües de EXIST 2023: 6.920 de entrenamiento más 1.038 de desarrollo, no el test separado de 2.076. Cada tuit tiene seis etiquetas binarias, tres de mujeres y tres de hombres y dos por cada grupo 18-22, 23-45 y 46+, para 47.748 juicios. GPT-3.5, GPT-4, GPT-4o, Mistral Small 22B y Qwen2.5-14B producen una etiqueta con un prompt base y con instrucciones que añaden solo sexo o edad. El outcome es alfa de Krippendorff de razón entre la etiqueta del modelo y la proporción humana del grupo por tuit. En la ejecución base, los cinco modelos tienen mayor alfa con el agregado femenino que con el masculino: 0,415/0,371; 0,365/0,325; 0,228/0,191; 0,353/0,310; y 0,378/0,345. Esto describe concordancia diferencial, no identifica por sí solo sesgo causal, personalidad femenina ni qué grupo es normativamente correcto. Por edad no hay dirección común. Las personas demográficas producen efectos mixtos: el prompt femenino eleva el acuerdo objetivo de GPT-4, GPT-4o y Mistral; el masculino solo el de GPT-4 y Qwen. GPT-4 y Qwen mejoran en sus tres condiciones de edad, GPT-3.5 solo en 46+, GPT-4o en ninguna y Mistral en 18-22 y 23-45 pero no en 46+. El texto afirma erróneamente que Mistral mejora consistentemente: la tabla baja de 0,392 a 0,383 para 46+. La lectura defendible es que una frase demográfica puede mover el acuerdo de modo dependiente del modelo y no fiable como mitigación. La selección del prompt también limita la inferencia: tres candidatos se probaron sobre veinte tuits y se eligió el de mayor coincidencia entre tres LLM (75%, frente a 70% y 55%); o1-preview lo reescribió después. El criterio fue consistencia entre modelos, no validez humana, calibración o reducción de sesgo, y no hay confirmación reservada. Cada celda final contiene una sola predicción por tuit; no se estiman repetición, paráfrasis o deriva de API, los cambios para Mistral/Qwen no se especifican y se agregan inglés y español sin desglose. La auditoría estadística descubre un problema sustantivo. Las tasas demográficas son observaciones emparejadas en los mismos tuits, pero el código usa t-test independiente y ANOVA/Tukey independientes. El artefacto reproduce p=0,237; respetando el emparejamiento, la diferencia mujer-hombre es solo 0,00716, pero el t-test pareado da p=0,0463 y Wilcoxon p=0,000383. Cambia la decisión nominal sin convertir 0,72 puntos porcentuales en un efecto grande. Las diferencias de edad siguen detectándose con Friedman emparejado, aunque los post hoc deben rehacerse. Los intervalos tampoco son reproducibles. El paper afirma 10.000 bootstraps e intervalos menores de 0,001, pero la función pública transpone la matriz 2x7.958, la aplana, remuestrea celdas y la reconstruye 7.958x2; la librería interpreta entonces otra estructura y se rompe el emparejamiento. Para mujer-hombre, el alfa objetivo es 0,4768 y el callback defectuoso 0,000254. En 500 remuestreos de auditoría, un bootstrap correcto da aproximadamente [0,4578; 0,4955], mientras el público produce un intervalo estrecho alrededor de cero. El commit previo al paper llama esta función 10.000 veces; como no hay outputs finales, no puede probarse qué ejecución generó el manuscrito, pero si se usó esta función sus intervalos no acotan las tablas. La distancia también importa: alfa de intervalo da 0,647 en vez de 0,477. El repositorio coincidente del primer autor compila, pero no está enlazado ni versionado en el paper y carece de dataset, predicciones finales, outputs Mistral/Qwen, CSV, intervalos, logs y tablas; el JSON intermedio no coincide en varios alfas. No hay tests, CI, tag o release; Ruff encuentra 68 incidencias y el entorno omite OpenAI. La capa API captura cualquier excepción y devuelve None sin publicar tasas finales de fallo. Tampoco hay prueba de contaminación: la nota de que las etiquetas no eran públicas no sostiene una garantía, porque el experimento usa train y desarrollo distribuidos por los organizadores. Las tablas apoyan una advertencia práctica sobre inestabilidad del persona prompting simple, no identidad, empatía, personalidad, equidad ni generalización.

Research question

With which demographic groups of annotators do five LLMs agree most when classifying sexism in EXIST 2023, and does adding sex or age to the prompt reliably shift that agreement toward the indicated group?

Method

Observational and prompt intervention study. Aggregates six human labels per tweet in proportions by gender and age, obtains a binary label per model with a base prompt and with five single-dimension demographic instructions, and compares vectors using ratio Krippendorff's alpha. Selects the prompt by consistency across LLMs over twenty tweets and uses t-test, ANOVA/Tukey and bootstrap; the audit re-evaluates the matching, the bootstrap implementation and the public artifacts.

Sample: 7,958 tweets in English and Spanish: 6,920 train + 1,038 development. Each tweet has six labels, three from women and three from men, and two per each group 18-22, 23-45 and 46+: 47,748 judgments. The separate test of 2,076 tweets is not part of the analysis.

Findings

  • The five base models have higher alpha with the female aggregate than with the male aggregate under the chosen metric.
  • The base female/male alphas are 0.415/0.371; 0.365/0.325; 0.228/0.191; 0.353/0.310; and 0.378/0.345.
  • There is no common age direction across the five models.
  • The female prompt improves the objective agreement of GPT-4, GPT-4o and Mistral; the male only that of GPT-4 and Qwen.
  • GPT-4 and Qwen improve across the three age persons; GPT-3.5 only in 46+ and GPT-4o in none.
  • Mistral does not improve in 46+: the table drops from 0.392 to 0.383, contradicting the text.
  • The female human rate exceeds the male only by 0.00716 per tweet.
  • The independent test reproduces p=0.237, while the paired test gives p=0.0463.
  • Age differences remain detected with a paired check, but the published post hoc tests do not model repetition per tweet.
  • The public bootstrap estimates another statistic near zero and not an interval around alpha 0.477.
  • The choice of distance is sensitive: ratio alpha gives 0.477 and interval 0.647 for female-male.
  • The public artifacts do not reproduce the final tables or the intervals.

Limitations

  • The manipulation is a demographic phrase, without validation of personality, empathy or internal perspective.
  • Higher agreement does not causally identify bias nor establish which group is normatively correct.
  • The same tweets are treated as independent samples in the published tests.
  • The public bootstrap breaks the matching and changes the orientation of the matrix.
  • Ratio alpha and its interpretation over discrete proportions are not justified.
  • There are no valid contrasts of difference between alphas nor control for multiple comparisons.
  • Some improvements are minuscule, such as Qwen male 0.345 to 0.347.
  • There is only one run per model and condition.
  • The prompt is chosen by consistency across models on twenty tweets, not by human validity.
  • The final rewrite with o1-preview has no reserved confirmation.
  • The changes for Mistral and Qwen are not specified.
  • English and Spanish are aggregated without a breakdown by language.
  • Only two gender categories, three age bins and one dimension at a time are considered.
  • There is no analysis of tweet or label contamination.
  • 7,958 corresponds to train+development, not the full corpus with test.
  • The matching repository is not linked in the paper nor does it fix a release.
  • Data, final predictions, statistical outputs, logs and tables are missing.
  • The environment omits a necessary dependency and fixes macOS ARM builds.
  • Error handling can turn API failures into absences without final traceability.

What the study does not establish

  • That the LLMs have a female, male or age-specific personality.
  • That a demographic instruction produces faithful simulation of a human group.
  • That the model empathizes or adopts lived experiences.
  • That higher alpha with women alone demonstrates an inherent bias and its cause.
  • That the female aggregate is a gold standard of fairness or quality.
  • That persona prompting mitigates bias reliably.
  • That Mistral improves across the three age conditions.
  • That intervals smaller than 0.001 validly bound the published alphas.
  • That independent tests are adequate for paired observations.
  • That a difference of 0.72 percentage points is large.
  • That the pattern is stable between English and Spanish.
  • That the results withstand other templates, runs or versions.
  • That the 7,958 tweets include the EXIST 2023 test.
  • That the evaluated data could not have appeared in training.
  • That intersectional, non-binary, cultural or individual perspectives are represented.
  • That the public repository reproduces the final tables.
  • That the results generalize to other moderation systems.

Traceability

Scope: Full text

Version: SIGIR '25 proceedings article, pp. 3045-3049; arXiv:2505.11795v1; 5 pages, all rendered and visually inspected

Consulted source: https://doi.org/10.1145/3726302.3730255

Review: Codex 5-page full-text visual, DOI/arXiv, table arithmetic, paired-statistics, bootstrap implementation, EXIST split and author-repository audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • gpt-3.5-turbo-0125
  • gpt-4-turbo-2024-04-09
  • gpt-4o-2024-08-06
  • Mistral-Small-Instruct-2409 22B
  • Qwen2.5-14B
  • o1-preview for prompt refinement

Instruments and metrics

  • EXIST 2023 binary sexism-identification task
  • Baseline expert-linguist prompt
  • Single-attribute gender persona instructions
  • Single-attribute age-group persona instructions
  • Krippendorff alpha with ratio distance
  • Independent t-test and one-way ANOVA/Tukey HSD as published
  • Bootstrap confidence intervals as implemented in the related public code

Data used

  • EXIST 2023 training split
  • EXIST 2023 development split
  • Author-owned partial intermediate prediction JSON (not a final replication artifact)

Evidence and location

  • Publication, DOI, pages, authorship and license: ACM DOI metadata for 10.1145/3726302.3730255 and arXiv:2505.11795v1 checked 2026-07-16
  • Design, models, prompt, tables, results and conclusions: SIGIR '25 / arXiv:2505.11795v1 PDF, 5 pages; every page rendered and visually inspected
  • Partitions and availability of EXIST 2023 labels: AI-UPV at EXIST 2023 arXiv:2307.03385, 15 pages; every page rendered and visually inspected
  • Matching, alternative tests, alpha and bootstrap: AngelFelipeMP/Sexism-data-analysis at bf184d1f6855bb23f8af8d3cf3f3ce538f8e244c; reanalysis checked 2026-07-16
  • Missing artifacts, environment, API failures, compileall and Ruff: AngelFelipeMP/Sexism-data-analysis repository audit on 2026-07-16
  • Claim limits and Mistral 46+ error: reports/verification/article-273-sigir-demographic-persona-sexism-agreement-paired-statistics-bootstrap-code-and-claim-audit.json