Unmasking Implicit Bias: Evaluating Persona-Prompted LLM Responses in Power-Disparate Social Scenarios

Applications, bias, and safety2025ACL AnthologyApproved editorial review

Authors: Bryan Chen Zhengyu Tan, Roy Ka-Wei Lee

Keywords: Demographic persona prompting, Implicit bias, Power disparity, LLM-as-a-judge, Social scenarios, Demographic sensitivity

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

2
Authors
12
Findings
23
Limitations
9
Evidence

Editorial summary

English

This paper studies how five LLMs change their responses when prompts assign demographic identities to two participants in a social situation, and whether variation is greater when power is unequal. It is relevant to persona prompting but does not measure personality. Personas combine a demographic label with a contextual character trait generated for the scene; there are no questionnaires, psychometric traits, temporal-stability tests or inferences about an internal model identity.

GPT-4o generates 100 scenarios in ten domains: work, education, healthcare, finance, politics, justice, neighborhood, information, housing and public assistance. Each domain contains five scenes in which Blake, the responder, has an advantage over Alex and five in which they have equal standing. Scenarios are manually reviewed. Alex is SUB and Blake RES; the response model adopts Blake and replies to Alex. A central limitation is that the 50 unequal-power and 50 equal-power scenes are different situations, not paired versions of the same scene. Any power-condition difference can therefore also reflect roles, wording, stakes and content.

The paper says it evaluates nine demographic axes, but Table 2 and the final data contain eight: race, gender identity, age, religion, political stance, disability, nationality and physical appearance. Their sizes yield 111 within-axis SUB×RES pairs; adding one no-demography baseline gives 112 conditions per scenario and 11,200 responses per model. No cross-axis combinations are tested, so a joint person such as a middle-aged Caucasian atheist man never appears as an experimental treatment.

The evaluated models are GPT-4o-mini, Llama 3.1 8B Instruct, Qwen 2 7B, Gemma 2 9B Instruct and Mistral 7B Instruct v0.3 at temperature 0. Semantic sensitivity is one minus cosine similarity between all-mpnet-base-v2 embeddings of a demographic response and its baseline. Quality is approximated by a GPT-4o-mini judge that compares both responses under Helpful-Honest-Harmless criteria in both orders; the final score can be 0, .25, .5, .75 or 1. The paper then calls the standard deviation of these group-cell means implicit bias. These are useful exploratory operationalizations, but distance, judge preference and variability are not by themselves equivalent to bias, harm or discrimination.

The lowest cosine distances often fall, depending on model and dialogue side, on middle-aged, abled, native-born, average-looking, centrist, Caucasian and atheist labels. The paper interprets these minima as a default persona. The faithful interpretation is narrower: these explicit prompts alter semantic content less than other labels relative to a baseline response. They do not show that an unprompted model internally adopts those identities. Gender also does not support a unique male default: Table 4 assigns the lowest RES value to male in 2/5 models and female in 2/5. The abstract stitches marginal minima from separate axes into a joint male persona that was never tested.

The largest distances frequently involve age, mental disability, non-binary/transgender identities, migration, appearance, politics and religion, with model-specific variation. GPT-4o-mini judge preferences also vary across demographic pairs. These should be described as that judge's relative preferences against one baseline response, not objective quality. GPT-4o-mini also judges GPT-4o-mini outputs, creating a self-evaluation condition without an alternative judge.

In the power comparison, win-rate standard deviation increases for all five models. Semantic-distance variability increases for Gemma 2, GPT-4o-mini, Llama 3.1 and Mistral, but decreases for Qwen 2. The audit recomputes Table 6's rounded values from the public CSVs. This confirms the descriptive statistic, but not a causal power-disparity effect: conditions do not use paired scenes and the metric measures dispersion across cells rather than validated unfair treatment.

Human validation uses three student volunteers. Inter-human agreement is only fair (Fleiss κ=.340); adding the AI judge gives .393, and mean human-AI Cohen κ is .447. In a second stage, people see the AI verdict and rationale before rating agreement, so the high Likert score validates post-exposure plausibility rather than independent correctness. Table 10 contains an arithmetic error: its three κ values average .447 and its three Likert values 4.117, but the Mean row prints .452 and 4.207. The main prose uses the correct means. The published form says 25 scenarios per section while the text and table refer to 100 pairs; batching is not explained.

The final artifacts are internally consistent. The five CSVs contain 56,000 rows: 11,200 per model, with 11,100 demographic conditions and 100 baselines, no duplicate experimental keys and the same 11,100 keys in all models. The original combined Drive CSV and the current five split CSVs match cell for cell after sorting. The XLSX contains 100 scenes, ten per domain and a 50/50 power split. Recalculation reproduces Table 6 to four decimals.

Upstream traceability is incomplete. generated_prompts.csv has 13,200 rows because its Gender axis also contains gay and lesbian, producing 36 pairs instead of the final study's 16. Removing the 20 pairs involving those labels across 100 scenes explains exactly 2,000 discarded rows, but the release does not document this filter. The repository publishes a visualization, data and one aggregation notebook, but no generation, embedding or judging code, order-specific verdicts, human annotations, immutable model revisions, environment, tests, CI or license. Figure 8 also reverses its conditional labels: the block marked power disparity present requests equal standing, while absent requests a RES advantage; Table 12 and the data follow the opposite intended convention. Without executed generation code, it is unresolved whether this is only a figure error or a provenance problem.

The faithful conclusion is that, in 100 generated scenarios under explicit demographic labels, models exhibit identity-dependent semantic shifts and LLM-judge preferences; win-rate dispersion is greater in scenes labeled as power-disparate. The study does not establish stable synthetic personality, a joint default persona, a causal effect independent of scene content, objective quality, real-world harm, intersectionality or a training-data source for the effect.

Español

Este trabajo estudia cómo cambian las respuestas de cinco LLM cuando el prompt asigna identidades demográficas a dos participantes de una situación social y si esa variación es mayor cuando existe desigualdad de poder. Es pertinente para persona prompting, pero no mide personalidad. Las personas combinan una etiqueta demográfica con un rasgo contextual generado para la escena; no hay cuestionarios, rasgos psicométricos, estabilidad temporal ni inferencia de una identidad interna del modelo.

GPT-4o genera 100 escenarios en diez dominios: trabajo, educación, sanidad, finanzas, política, justicia, vecindario, información, vivienda y asistencia social. En cada dominio hay cinco escenas donde Blake, el respondedor, tiene ventaja sobre Alex y cinco donde ambos están en pie de igualdad. Los escenarios se revisan manualmente. Alex es SUB y Blake RES; el modelo adopta a Blake y responde a Alex. Un límite esencial es que las 50 escenas con desigualdad y las 50 sin ella son situaciones distintas, no versiones emparejadas de una misma escena. Por tanto, cualquier diferencia entre condiciones de poder también puede deberse a roles, redacción, stakes y contenido.

El paper afirma evaluar nueve ejes demográficos, pero Table 2 y los datos finales contienen ocho: raza, identidad de género, edad, religión, posición política, discapacidad, nacionalidad y apariencia física. Sus tamaños producen 111 combinaciones SUB×RES dentro del mismo eje; al añadir una línea base sin demografía resultan 112 condiciones por escenario y 11.200 respuestas por modelo. No se prueban combinaciones entre ejes: una persona conjunta como «hombre caucásico ateo de mediana edad» nunca aparece como tratamiento experimental.

Se evalúan GPT-4o-mini, Llama 3.1 8B Instruct, Qwen 2 7B, Gemma 2 9B Instruct y Mistral 7B Instruct v0.3, a temperatura 0. La sensibilidad semántica es 1 menos la similitud coseno entre embeddings all-mpnet-base-v2 de la respuesta demográfica y su línea base. La calidad se aproxima con un juez GPT-4o-mini que compara ambas respuestas bajo Helpful-Honest-Harmless en los dos órdenes; el score final puede ser 0, 0,25, 0,5, 0,75 o 1. Después, el artículo llama «sesgo implícito» a la desviación estándar de esos promedios entre combinaciones demográficas. Estas son operacionalizaciones útiles para explorar, pero distancia, preferencia del juez y variabilidad no equivalen por sí solas a sesgo, daño o discriminación.

Los mínimos de distancia coseno suelen recaer, según modelo y lado del diálogo, en las etiquetas middle-aged, abled, native-born, average-looking, centrist, Caucasian y atheist. El paper interpreta esos mínimos como una «persona por defecto». La lectura fiel es más limitada: son prompts explícitos que cambian menos el contenido respecto a una respuesta base. No demuestran que el modelo sin prompt adopte internamente esas identidades. Además, el resultado de género no sostiene un default masculino único: Table 4 atribuye el mínimo de RES a male en 2/5 modelos y a female en 2/5. El abstract une mínimos marginales de ejes separados y los presenta como una persona masculina conjunta que no fue probada.

Las mayores distancias aparecen con frecuencia en pares asociados a edad, discapacidad mental, identidades non-binary/transgender, migración, apariencia, política y religión, aunque el patrón varía por modelo. El juez GPT-4o-mini también prefiere o penaliza distintos pares demográficos. Deben describirse como preferencias relativas de ese juez frente a una única respuesta base, no como calidad objetiva. Para GPT-4o-mini existe además autoevaluación: el mismo modelo que genera una de las familias de respuestas actúa como juez, sin comparación con otro juez.

En la comparación de poder, la desviación estándar de win rate aumenta en los cinco modelos. La variabilidad semántica aumenta en Gemma 2, GPT-4o-mini, Llama 3.1 y Mistral, pero disminuye en Qwen 2. La auditoría recompone los valores redondeados de Table 6 desde los CSV públicos. Esto confirma la estadística descriptiva, pero no un efecto causal de la desigualdad: las condiciones no usan escenas emparejadas y la métrica identifica dispersión entre celdas, no trato injusto validado.

La validación humana usa tres estudiantes voluntarios. El acuerdo entre humanos es bajo/fair (Fleiss κ=0,340); al incluir al juez AI es 0,393 y el Cohen κ humano-AI medio es 0,447. En una segunda fase se muestra a las personas el veredicto y la razón del AI antes de pedir acuerdo, por lo que el Likert alto valida plausibilidad después de exposición, no corrección independiente. Table 10 contiene un error: sus tres κ promedian 0,447 y sus tres Likert 4,117, pero la fila Mean imprime 0,452 y 4,207. El cuerpo sí usa las medias correctas. El formulario publicado dice 25 escenarios por sección, mientras el texto y la tabla hablan de 100 pares; no se explica el batching.

Los artefactos finales son internamente consistentes. Los cinco CSV contienen 56.000 filas: 11.200 por modelo, con 11.100 condiciones demográficas y 100 líneas base, sin claves experimentales duplicadas y con los mismos 11.100 cruces en todos los modelos. El CSV combinado original de Drive y los cinco CSV actuales coinciden celda por celda tras ordenar. El XLSX contiene 100 escenas, diez por dominio y 50/50 por condición de poder. La recomputación reproduce los valores de Table 6 a cuatro decimales.

La trazabilidad anterior al resultado es incompleta. generated_prompts.csv tiene 13.200 filas porque su eje Gender incluye también gay y lesbian: 36 pares en vez de los 16 del estudio final. Eliminar los 20 pares que incluyen esas etiquetas en 100 escenas explica exactamente las 2.000 filas descartadas, pero el release no documenta el filtro. El repositorio solo publica una visualización, datos y un notebook de agregación; no incluye código de generación, embedding, juicio, resultados por orden, anotaciones humanas, revisiones inmutables de modelos, entorno, tests, CI o licencia. Figure 8 también invierte las etiquetas de su prompt: el bloque marcado «power disparity present» pide igualdad y el marcado «absent» pide ventaja de RES, mientras Table 12 y los datos siguen la convención opuesta. Sin el código ejecutado no puede resolverse si es un error de figura o de procedencia.

La conclusión fiel es que, en 100 escenas generadas y bajo etiquetas demográficas explícitas, los modelos muestran cambios semánticos y preferencias de un juez LLM que varían por identidad; la dispersión del win rate es mayor en las escenas etiquetadas con desigualdad. El estudio no demuestra una personalidad sintética estable, una persona conjunta por defecto, causalidad independiente del contenido de la escena, calidad objetiva, daño real, interseccionalidad ni el origen del efecto en los datos de entrenamiento.

Research question

How do eight demographic persona axes alter the responses of five LLMs across 100 social scenes, which pairs produce lower semantic distance or lower preference from an HHH judge, and does the dispersion change between scenes labeled with and without power inequality?

Method

GPT-4o generates 100 scenes across ten domains, 50 with RES advantage and 50 with equality. For each scene, 111 demographic pairs are constructed within eight axes plus a baseline. Five models respond as RES at temperature 0. Sensitivity uses cosine distance of all-mpnet-base-v2 embeddings; GPT-4o-mini judges HHH in both orders; the standard deviation per axis is averaged as an indicator of variability. Three students perform a partial validation. The audit reviews the complete PDF and recomposes the final data and Table 6.

Sample: Each model generates 11,200 responses: 11,100 for demographic pairs and 100 baselines, balanced across 5,600 rows per power condition. The eight axes contain 29 unique labels and 111 within-axis pairs. The human validation declares three students; the text speaks of 100 pairs, but the form shows 25 scenarios per section and does not document the batches.

Findings

  • Responder labels usually produce more marginal variation than subject labels.
  • Distance minima often appear in middle-aged, abled, native-born, average-looking, centrist, Caucasian, and atheist.
  • The gender result is not a single male default: male and female tie with 2/5 models in Table 4.
  • The largest distances usually involve age, mental disability, non-binary/transgender, migration, appearance, politics, and religion.
  • The HHH preferences of the judge change notably across demographic pairs.
  • The standard deviation of the win rate increases with the inequality label across all five models.
  • Semantic dispersion increases in four models and decreases in Qwen 2.
  • Table 6 is reproduced to four decimal places from the public CSVs.
  • All 11,100 experimental crossings are present across the five models with no duplicates.
  • The original combined CSV and the current CSVs match cell by cell after sorting.
  • Interhuman agreement is fair, not high: Fleiss κ=0.340.
  • The mean correct human-AI Cohen κ is 0.447 and the mean correct Likert is 4.117.

Limitations

  • Personality, trait stability, or internal identity is not measured.
  • The power conditions use different and unpaired scenes.
  • The power effect is confounded with content, roles, stakes, and wording.
  • Eight demographic axes are tested, not nine.
  • Combinations across axes or intersectionality are not tested.
  • The joint persona in the abstract is constructed from marginal minima never evaluated together.
  • Low cosine distance does not prove that the model adopts a default identity.
  • Standard deviation is an unvalidated operationalization of implicit bias.
  • The HHH preference of an LLM is not objective quality or observed harm.
  • GPT-4o-mini also judges responses from GPT-4o-mini.
  • There is no alternative judge or broad human gold standard.
  • Only three students participate and interhuman agreement is fair.
  • The Likert phase shows the AI verdict and reason beforehand.
  • The human sampling, batches, and the 25/100 discrepancy are not explained.
  • Table 10 prints two arithmetically incorrect means.
  • Figure 8 inverts the present/absent labels of the power prompt.
  • generated_prompts.csv includes gay and lesbian under Gender and 2,000 rows later excluded without explanation.
  • No generation, embedding, or judgment scripts are published.
  • No per-order verdicts, raw reasons, or human annotations are published.
  • There are no immutable model revisions or exact inference dates.
  • There are no uncertainty intervals or correction for multiple comparisons.
  • The qualitative examples are three selected maxima, not prevalence.
  • There is no reproducible environment, tests, CI, tags, or license.

What the study does not establish

  • It does not demonstrate a stable synthetic personality.
  • It does not demonstrate a default joint persona.
  • It does not demonstrate that an unprompted LLM identifies as a Caucasian atheist centrist male.
  • It does not demonstrate intersectionality.
  • It does not identify a causal effect of inequality independent of the scene.
  • It does not convert semantic sensitivity into proven bias or harm.
  • It does not measure objective quality or real discrimination.
  • It does not identify the causal origin in training data.
  • It does not generalize beyond these scenes and snapshots.
  • It does not allow end-to-end reproduction of generation, judgment, or human validation.

Traceability

Scope: Full text

Version: NAACL 2025 proceedings version, pp. 1075-1108; arXiv:2503.01532v2 checked. Thirty-four-page PDF fully rendered and visually inspected; Google Drive data and repository commit e3bdb7cf0c96de50a497daaf876591274426e1b8 audited.

Consulted source: https://aclanthology.org/2025.naacl-long.50/

Review: Codex full-text, visual, methodological, spreadsheet, data and reproducibility audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • OpenAI GPT-4o for scenario generation (exact snapshot not reported)
  • OpenAI GPT-4o-mini as response model and HHH judge (exact snapshot not reported)
  • Meta Llama 3.1 8B Instruct (revision not pinned)
  • Alibaba Qwen 2 7B Instruct (revision not pinned)
  • Google Gemma 2 9B Instruct (revision not pinned)
  • Mistral 7B Instruct v0.3 (revision not pinned)
  • sentence-transformers/all-mpnet-base-v2

Instruments and metrics

  • Dual-persona social-scenario prompt
  • Eight within-axis demographic persona manipulations
  • No-demography baseline response
  • all-mpnet-base-v2 768-dimensional response embeddings
  • Cosine distance from baseline response
  • GPT-4o-mini Helpful-Honest-Harmless preference judge
  • Two-order 1/0.5/0 preference scoring
  • Within-axis population standard deviation and AvgStd aggregation
  • Three-judge human preference and rationale-agreement form
  • Fleiss kappa, Cohen kappa and five-point Likert rating

Data used

  • 100 GPT-4o-generated social scenarios across ten contextual domains
  • 111 within-axis demographic SUB×RES pairs plus one baseline per scenario
  • 56,000 released final response rows across five models
  • generated_prompts.csv with 13,200 pre-filter prompt rows
  • scenarios.xlsx with 50 power-disparate and 50 equal-power scenes
  • Google Drive combined response, cosine-distance and preference-result files
  • Inc0mple/Implicit_Bias_Interactive_Data_Viz commit e3bdb7cf0c96de50a497daaf876591274426e1b8

Evidence and location

  • Scope, tasks, and research questions: NAACL 2025 pp. 1075-1077, Abstract, Introduction and Task Definition
  • Scenarios, axes, pairs, and response generation: NAACL 2025 pp. 1077-1078, Sections 4.1-4.3 and Tables 1-3
  • Semantic metrics, HHH judge, and AvgStd: NAACL 2025 pp. 1078-1079, Sections 5.1-5.4
  • Models and results for default, quality, and power: NAACL 2025 pp. 1079-1083, Sections 5.5-6.6 and Tables 4-7
  • Limitations, human validation, and Table 10 errors: NAACL 2025 pp. 1083, 1088-1092, Limitations and Appendices B-C
  • Power prompt contradiction: NAACL 2025 pp. 1092-1094, Table 12 and Figure 8
  • Official metadata and version: ACL Anthology 2025.naacl-long.50 and arXiv:2503.01532v2
  • Counts, gender filter, Table 6 reproduction, and absent artifacts: Google Drive release and Inc0mple/Implicit_Bias_Interactive_Data_Viz commit e3bdb7cf0c96de50a497daaf876591274426e1b8 audited 16 July 2026
  • Complete report: reports/verification/article-217-implicit-bias-persona-validity-and-artifact-audit.json