Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Analysing LLM Persona Generation and Fairness Interpretation in Polarised Geopolitical Contexts

Maida Aizaz, Quang Minh Nguyen

Abstract

Large language models (LLMs) are increasingly utilised for social simulation and persona generation, necessitating an understanding of how they represent geopolitical identities. In this paper, we analyse personas generated for Palestinian and Israeli identities by five popular LLMs across 640 experimental conditions, varying context (war vs non-war) and assigned roles. We observe significant distributional patterns in the generated attributes: Palestinian profiles in war contexts are frequently associated with lower socioeconomic status and survival-oriented roles, whereas Israeli profiles predominantly retain middle-class status and specialised professional attributes. When prompted with explicit instructions to avoid harmful assumptions, models exhibit diverse distributional changes, e.g., marked increases in non-binary gender inferences or a convergence toward generic occupational roles (e.g., "student"), while the underlying socioeconomic distinctions often remain. Furthermore, analysis of reasoning traces reveals an interesting dynamics between model reasoning and generation: while rationales consistently mention fairness-related concepts, the final generated personas follow the aforementioned diverse distributional changes. These findings illustrate a picture of how models interpret geopolitical contexts, while suggesting that they process fairness and adjust in varied ways; there is no consistent, direct translation of fairness concepts into representative outcomes.

Analysing LLM Persona Generation and Fairness Interpretation in Polarised Geopolitical Contexts

Abstract

Large language models (LLMs) are increasingly utilised for social simulation and persona generation, necessitating an understanding of how they represent geopolitical identities. In this paper, we analyse personas generated for Palestinian and Israeli identities by five popular LLMs across 640 experimental conditions, varying context (war vs non-war) and assigned roles. We observe significant distributional patterns in the generated attributes: Palestinian profiles in war contexts are frequently associated with lower socioeconomic status and survival-oriented roles, whereas Israeli profiles predominantly retain middle-class status and specialised professional attributes. When prompted with explicit instructions to avoid harmful assumptions, models exhibit diverse distributional changes, e.g., marked increases in non-binary gender inferences or a convergence toward generic occupational roles (e.g., "student"), while the underlying socioeconomic distinctions often remain. Furthermore, analysis of reasoning traces reveals an interesting dynamics between model reasoning and generation: while rationales consistently mention fairness-related concepts, the final generated personas follow the aforementioned diverse distributional changes. These findings illustrate a picture of how models interpret geopolitical contexts, while suggesting that they process fairness and adjust in varied ways; there is no consistent, direct translation of fairness concepts into representative outcomes.
Paper Structure (54 sections, 30 figures, 3 tables)

This paper contains 54 sections, 30 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Representation risks and social biases in LLMs
  4. Safety intervention for LLMs
  5. Interpretability as a tool
  6. Methods
  7. Models
  8. Experimental Design
  9. RQ1: What profiles do models typically assign to Palestinians and Israelis in various conditions?
  10. RQ2: How do safety prompts alter the distributional properties of generated personas?
  11. RQ3: What corresponds to 'fairness' in model reasoning traces?
  12. Results and Discussions
  13. Generated Profiles
  14. General Disparities
  15. Gender
  16. ...and 39 more sections

Figures (30)

  • Figure 1: There are significant occupation disparities which correlate with war nuances.
  • Figure 2: The war context inflates the lower-class of Palestinians but enhances the middle-class of Israelis.
  • Figure 3: The war increases the proportion of negative descriptors (fatigue and injury) for both ethnic groups, but more prominently so for Palestinians.
  • Figure 4: The hint makes inferred changes gender distributions significantly (especially with Qwen and Llama). This can be seen through the visualised directions from no-hint to hint, which largely supress male personas.
  • Figure 5: Prompting models with debiasing hints does not consistently neutralise occupational disparities. Here only the top-seven job categories for Gemini, Llama, and Qwen are visualised.
  • ...and 25 more figures