Analysing LLM Persona Generation and Fairness Interpretation in Polarised Geopolitical Contexts

1. Problem Statement

Large Language Models (LLMs) are increasingly deployed for social simulation and persona generation. However, existing research on LLM bias largely focuses on Western-centric demographic categories (e.g., gender, race) and lacks investigation into complex, polarized geopolitical identities.

The authors address the critical gap in understanding how LLMs represent identities in active conflict zones, specifically the Palestinian-Israeli conflict. The core problem is determining whether models can generate fair, nuanced personas in such contexts or if they inadvertently encode historical biases, socioeconomic disparities, and conflict narratives into their outputs. Furthermore, the paper investigates whether explicit "safety" or "debiasing" prompts effectively correct these biases or merely shift the nature of the distortion.

2. Methodology

The study employs a rigorous experimental design involving five popular LLMs (Gemma 3 27B, Qwen3 32B, Llama 3.3 70B Instruct, Gemini 2.5 Pro, and GPT-4.1) across 640 experimental conditions per model (3,200 total responses).

• Experimental Setup:

  • Roles: Models were assigned one of five roles (UN peacekeeper, journalist, traveller, local resident, world leader).
  • Contexts: Prompts varied between "War" (witnessing the ongoing war in Gaza) and "No War" conditions.
  • Attributes: Models generated profiles for Palestinian and Israeli individuals, inferring gender, age, socioeconomic status (SES), city, occupation, and physical appearance.
  • Intervention: A subset of prompts included explicit instructions to "avoid harmful assumptions" or "stereotypes" (the "debiasing hint").

• Analysis Techniques:

  • Statistical Analysis: Quantitative analysis of distributional shifts in inferred attributes (e.g., SES, job types, gender ratios) across conditions.
  • Interpretability via SAE: The authors utilized a Sparse Autoencoder (SAE) trained on Llama 3.1 8B to analyze the reasoning traces (rationales) generated by the models. This allowed them to extract interpretable features related to fairness, caution, and harm avoidance, comparing the internal reasoning process against the final generated output.

3. Key Contributions

1. Geopolitical Bias Audit: Provides the first systematic analysis of how LLMs construct personas for opposing sides in an active, high-stakes geopolitical conflict.
2. Dissociation of Reasoning and Generation: Demonstrates a critical disconnect where models explicitly reason about fairness and safety in their internal traces but fail to translate these concepts into equitable final outputs.
3. Ineffectiveness of Simple Prompting: Shows that standard "debiasing" prompts do not neutralize disparities; instead, they often induce new, unintended distributional shifts (e.g., over-correction in gender or convergence on generic roles) without resolving underlying socioeconomic biases.
4. SAE Application: Introduces the use of SAEs as a tool for analyzing social text generation and reasoning traces to identify specific "fairness" features in model behavior.

4. Key Results

A. Representational Disparities (War vs. No War)

• Socioeconomic Status (SES): A stark asymmetry exists. Palestinian profiles in war contexts are overwhelmingly associated with lower-class status and survival-oriented roles. In contrast, Israeli profiles predominantly retain middle-class status and specialized professional attributes, even in war contexts.
• Occupation:
  • Palestinians: Frequently assigned survival jobs (scavenger, water carrier, street vendor) or medical roles (nurse, paramedic) in war contexts.
  • Israelis: Consistently assigned high-status professional roles (software engineer, architect, tech entrepreneur) regardless of the war context.
• Appearance: War contexts increase descriptors of fatigue, injury, and dishevelment for both groups, but the effect is significantly more pronounced for Palestinians. Israelis are more likely to retain descriptors related to grooming and positivity.

B. Impact of Debiasing Prompts

When prompted to avoid harmful assumptions:

• Gender Shifts: Models did not converge on a balanced gender distribution. Instead, they often suppressed male identities in favor of female or non-binary identities (e.g., GPT-4.1 reduced male inferences from 76% to 45%), effectively skewing the data in the opposite direction.
• Occupational Convergence: Models shifted toward generic, "safe" roles like "student" for Palestinians, but did not significantly increase high-status professional roles. The underlying socioeconomic gap (survival vs. professional) largely persisted.
• Appearance: Negative descriptors decreased slightly, but the models interpreted "fairness" differently for each group (e.g., focusing on emotional resilience for Israelis vs. positivity for Palestinians).

C. Reasoning vs. Generation (The "Dissociation")

• Rationales: When asked to explain their choices, models consistently generated text containing fairness-related keywords (e.g., "bias," "stereotype," "neutral," "diversity") and SAE features related to caution and harm avoidance.
• The Disconnect: Despite these explicit fairness arguments in their reasoning traces, the final generated personas continued to exhibit the same distributional biases (e.g., low SES for Palestinians). The models "talked" about fairness but "acted" according to their training data's statistical correlations with the conflict.

5. Significance and Implications

• Limitations of Current Safety Mechanisms: The study suggests that simple prompt-level interventions (safety hints) are insufficient for achieving true geopolitical fairness. They may alter surface-level attributes (like gender) but fail to address deep-seated socioeconomic and narrative biases.
• Risk of Automated Polarization: Uncritical deployment of LLMs in conflict simulation could automate and reinforce dehumanizing narratives, where one side is perpetually framed as a victim of survival while the other remains insulated and professional.
• Need for New Frameworks: The authors argue for a deeper understanding of how models interpret "fairness" in polarized contexts. Current alignment mechanisms do not guarantee representative outcomes, necessitating new frameworks for evaluating and correcting geopolitical biases in AI.

In conclusion, the paper reveals that LLMs process fairness as a complex, often contradictory set of instructions that result in diverse, sometimes counter-productive, distributional shifts, rather than a consistent move toward equitable representation in polarized geopolitical settings.