Analysing LLM Persona Generation and Fairness Interpretation in Polarised Geopolitical Contexts

Society, culture, and collective behavior2026arXivApproved editorial review

Authors: Maida Aizaz, Quang Minh Nguyen

Keywords: Persona conditioning, Human simulation, Safety and bias

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

2
Authors
7
Findings
22
Limitations
6
Evidence

Editorial summary

English

This study compares how five LLMs, Gemma 3 27B, Qwen 3 32B, Llama 3.3 70B Instruct, Gemini 2.5 Pro and GPT-4.1, complete Palestinian and Israeli profiles across war versus no-war context, age framing and five roles. Across 640 configurations per model and 3,200 baseline profiles, war more often shifts Palestinian profiles toward lower socioeconomic status, survival-oriented occupations and fatigue or injury descriptors, while Israeli profiles remain predominantly middle class and professionally specialised. Adding a warning against harmful assumptions does not correct the pattern consistently: depending on the model it changes pronouns, converges occupations toward student or alters appearance, while socioeconomic differences persist. This is useful evidence of representational disparity and prompt sensitivity, not a measurement of real people. There is no demographic reference or operational fairness target; the design appears to use one generation per cell at temperature 0.7, without intervals or tests. The prompt forces children to have worked and reduces gender and class to closed categories. Its rationale analysis also counts words supplied by the warning itself, while an SAE from another Llama is applied to text rather than target-model activations. No data or code supports regeneration. Official Figure 9 permutes the Llama, Qwen and GPT series relative to Table 1. The paper supports auditing prompt-induced associations, not faithful population simulation or demonstrated fair reasoning.

Español

Este trabajo compara cómo cinco LLM, Gemma 3 27B, Qwen 3 32B, Llama 3.3 70B Instruct, Gemini 2.5 Pro y GPT-4.1, completan perfiles palestinos e israelíes bajo contexto de guerra o no guerra, edad y cinco roles. Con 640 configuraciones por modelo, 3.200 perfiles base, observa que la guerra desplaza con más frecuencia los perfiles palestinos hacia clase baja, ocupaciones de supervivencia y descriptores de fatiga o lesión, mientras los israelíes permanecen mayoritariamente en clase media y empleos especializados. Añadir una advertencia contra supuestos dañinos no corrige el patrón de forma estable: según el modelo cambia pronombres, converge hacia estudiante o altera la apariencia, pero persisten diferencias socioeconómicas. Es un diagnóstico valioso de disparidad representacional y sensibilidad al prompt, no una medida de personas reales. No hay referencia demográfica ni definición operativa de justicia; se usa una sola generación por celda a temperatura 0,7, sin intervalos ni pruebas. El prompt obliga a los niños a haber trabajado y reduce género y clase a categorías cerradas. Además, el análisis de razones cuenta palabras incluidas en la propia advertencia, y el SAE de otro Llama se aplica a textos, no a activaciones internas. No hay datos ni código para reproducir los resultados. La figura 9 oficial intercambia las series de Llama, Qwen y GPT respecto de la tabla 1. El artículo apoya auditar asociaciones inducidas por prompts, no simular fielmente poblaciones ni demostrar razonamiento justo.

Research question

How five LLMs represent Palestinian and Israeli identities when generating personas under variations of war, age, and role; how that representation changes when asked to avoid harmful assumptions; and what their textual justifications show about that response to the concept of justice.

Method

Factorial design in English executed through OpenRouter at temperature 0.7. Each model completes a template with pronoun, age, residence, occupation, socioeconomic class, and appearance for identity, war/no war, child/person, five roles, and fixed or inferred gender/SES conditions. 640 profiles per model are reported. A second condition adds an anti-stereotype warning. The authors normalize occupations and places, manually classify appearance words into nine groups, count two vocabulary lists in post-hoc reasons, and apply InterpEmbed with an SAE from Llama 3.1 8B to the texts. They also examine reasoning tokens from Qwen and an API summary from Gemini.

Sample: Five models, 640 responses per model, and 3,200 base profiles. The factors include two identities, war/no war, child/person, and five roles, in addition to branches with fixed or inferred gender and class. The article does not present a complete table that derives the 640 cells nor does it reconcile the base total with the warning-paired generations, the reasons, and the reasoning tokens/summaries. There appears to be a single stochastic sample per cell.

Findings

  • In war, Palestinian profiles are associated more frequently with low class, street vending, survival labor, fatigue, and injury; Israelis retain more middle class and specialized professions.
  • In the child prompt, more vendors or workers appear for Palestinians and more students for Israelis, although the template forces all children to have 'worked as'.
  • The assigned role appears to matter less than the model, the identity, and the context.
  • The anti-stereotype warning does not produce a common correction: it produces distinct shifts per model and does not consistently eliminate socioeconomic differences.
  • Reasons with bias vocabulary increase +21.34% in Gemma, +18.02% in Qwen, +15.84% in Llama, +10.66% in Gemini, and +22.66% in GPT.
  • The most frequent SAE features after the warning relate to caution, harm, avoidance, and stereotypes, consistent with the model having picked up the language of the prompt.
  • Table 1 reports pronoun distributions by model, but Figure 9 labels GPT data under Llama, Llama data under Qwen, and Qwen data under GPT.

Limitations

  • It is a descriptive audit of outputs, with no human population, ground truth, or demographic distribution against which to measure fidelity.
  • It does not operationally define what outcome would be just or unbiased; it cannot demonstrate correctness, representativeness, or absolute equity.
  • It appears to use a single generation per cell at temperature 0.7, with no replicates, seeds, or sensitivity to stochastic variation.
  • It offers no intervals, tests, hierarchical models, multiplicity control, or uncertainty; significant is used descriptively.
  • It does not reconcile the 3,200 base profiles with the warning condition, reasons, failures, retries, and exclusions in a complete pipeline.
  • It does not enumerate in a table all factorial cells, fixed/inferred levels, and denominators.
  • The exact aliases/snapshots, dates, top-p, token limits, seeds, retries, costs, and errors per model are not documented.
  • Manual coding of appearance and normalization of occupation/place lacks complete mappings, number and independence of annotators, agreement, and adjudication.
  • The warning contains harmful assumptions or stereotypes and the counted list includes harmful, assumption, and stereotyp; part of the verbal effect is direct prompt leakage.
  • The reasons requested after generation are not a faithful causal explanation of the original decision.
  • The SAE belongs to Llama 3.1 8B and is used as an extractor over external text; it does not inspect activations of the five target models.
  • Presence threshold, selection universe, cross-domain transfer, and multiple comparisons in the SAE analysis are not validated.
  • The hidden reasoning summary from Gemini is not comparable with the more complete reasoning tokens from Qwen.
  • The template restricts gender to he/she/they and equates they with non-binary; it measures forced pronoun, not validated gender identity.
  • Socioeconomic class is forced to low/middle/high without a validated instrument or population reference.
  • The child template requires 'worked as a [job]', which privileges occupation and child labor and limits the interpretation of the result.
  • The Palestinian/Israeli binary and the variable use of identity, ethnicity, and side obscure intragroup heterogeneity.
  • The maps do not specify a border source or geopolitical convention; the text imprecisely calls Tel Aviv the capital of Israel.
  • There is no documented consultation with Palestinian/Israeli people, regional experts, or affected communities despite the active war context.
  • Profiles, reasons, reasoning, structured prompts, mappings, code, logs, scripts, and figure data are not released.
  • The official Figure 9 permutes Llama/Qwen/GPT in v1 and v2; without data/code the same risk in other figures cannot be excluded.
  • The claim that only Qwen recognizes non-binary is excessive: Table 1 gives 8.75% of they to Llama in no war, even though Qwen is more consistent.

What the study does not establish

  • It does not demonstrate that the generated profiles are faithful to real Palestinian or Israeli people.
  • It does not demonstrate human simulation, demographic accuracy, or population representativeness.
  • It does not establish an objective metric of justice nor that the warning makes the profiles more just.
  • It does not causally attribute the patterns to training, to an internal perception, or to a model mechanism.
  • It does not validate post-hoc reasons as faithful decision traces.
  • It does not convert SAE text features into mechanistic evidence about GPT-4.1, Gemini, Gemma, Qwen, or Llama 3.3.
  • It does not fully separate model associations from constraints and priming introduced by the template.
  • It does not support statistical inference or stability of the percentages across stochastic samples.
  • It does not allow regenerating the tables or figures with public artifacts.
  • It does not make Figure 9 reliable for comparing pronouns across Llama, Qwen, and GPT.

Traceability

Scope: Full text

Version: arXiv:2603.22837v1; cross-checked against ACL Anthology 2026.eacl-srw.51 Version 2, DOI 10.18653/v1/2026.eacl-srw.51

Consulted source: https://arxiv.org/abs/2603.22837

Review: Codex 21-page arXiv plus official ACL v1/v2 visual full-text, construct, prompt-leakage, statistical, figure-integrity, reproducibility and ethics audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Gemma 3 27B
  • Qwen 3 32B
  • Llama 3.3 70B Instruct
  • Gemini 2.5 Pro
  • GPT-4.1
  • Goodfire Llama 3.1 8B layer-19 sparse autoencoder

Instruments and metrics

  • Plantilla factorial de persona geopolítica
  • Advertencia 'avoid harmful assumptions or stereotypes'
  • Normalización manual de ocupaciones y lugares
  • Taxonomía manual de nueve categorías de apariencia
  • Dos listas manuales de palabras de estrategia/justicia
  • InterpEmbed con SAE de Llama 3.1 8B
  • Distribuciones descriptivas y mapas de residencia
  • Razones post-hoc y salidas de razonamiento Qwen/Gemini

Data used

  • 3,200 baseline persona profiles: 640 per target model, not released
  • Debiasing-condition profiles and rationales, total accounting unclear and not released
  • Qwen reasoning outputs and Gemini API reasoning summaries, not released
  • LMSYS-Chat-1M as the reported training corpus of the external SAE

Evidence and location

  • Design, prompts, 640 responses per model, methods, results, limits, and ethics: ACL Anthology 2026.eacl-srw.51 current Version 2, 21/21 pages rendered and individually inspected
  • Persistence of the Figure 9 label failure: Official ACL Version 1 and current Version 2, Appendix Figure 9 compared visually with Table 1
  • Publication, DOI, versions, pagination, and license: Official ACL Anthology landing page inspected 2026-07-17
  • arXiv status and original metadata: Official arXiv abstract and Atom records for 2603.22837 inspected 2026-07-17
  • Current absence of artifacts linked or locatable by title/ID: Paper, ACL/arXiv records and exact-title/arXiv-ID public web and GitHub repository searches on 2026-07-17
  • Comprehensive audit of construct, prompt leakage, statistics, figure, reproducibility, and ethics: reports/verification/article-385-geopolitical-persona-fairness-construct-prompt-leakage-statistics-figure-label-reproducibility-and-ethics-audit.json