Skip to content

From Single to Societal: Analyzing Persona-Induced Bias in Multi-Agent Interactions

🕒 Published (v1): 2025-11-14 18:19 UTC · Source: Arxiv · Venue: AAAI · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Assigning demographic personas (gender, race) to LLM-based agents introduces measurable social biases in multi-agent interactions: historically advantaged personas (men, White individuals) are perceived as less trustworthy and show less insistence, while all agents exhibit in-group favoritism. These biases persist across GPT-4o, Gemini-1.5-Pro, and DeepSeek-V3, across group sizes, and across interaction rounds.

Problem

Multi-agent LLM systems commonly assign demographic personas to agents to encourage behavioral diversity, but the fairness implications are unexplored. Prior work studied persona-induced bias in single-agent settings; this paper addresses whether and how biases manifest in interactive, multi-agent settings where conformity dynamics compound individual biases.

Method

Three-stage controlled experimental design using two tasks—Collaborative Problem Solving (GPQA graduate-level MCQs) and Persuasion (PMIYC claims on political/ethical topics)—across three LLMs (GPT-4o, Gemini-1.5-Pro, DeepSeek-V3):

  1. Single-persona dyadic (§4): One agent has a persona, one is a default agent. Trustworthiness T(p) = probability default agent conforms to persona agent; Insistence I(p) = 1 − probability persona agent conforms to default. Pre-assigned initial responses isolate persona as the sole variable.
  2. Dual-persona dyadic (§5): Both agents assigned personas; conformity heatmaps reveal how individual trustworthiness/insistence combine and whether in-group favoritism emerges.
  3. Multi-agent, multi-round (§6): Scales to 3–6 agents and up to 5 rounds; tracks Win Rate (CPS) and Persuasion Effectiveness to assess generalization. Experiments focus on (man, woman) and (White, Black) persona pairs.

Key Contributions

  • First systematic study of persona-induced bias specifically in multi-agent interactions, isolating trustworthiness (how others treat a persona) and insistence (how a persona advocates for itself) as distinct, measurable social traits.
  • Empirical demonstration that individual persona biases are approximately additive: pairwise conformity rates align with the product of trustworthiness and insistence ranks measured in isolation.
  • Discovery of in-group favoritism in LLM agents (higher conformity within same-persona pairs), paralleling Social Identity Theory in humans.
  • Evidence that advantaged-group agents (men, White) receive lower trustworthiness scores and show higher conformity rates—echoing sociological findings on elite distrust—rather than the stereotyped direction.

Results

  • Trustworthiness variation: Average max-min gap of 6.7% across personas; average absolute deviation from no-persona baseline of 5.3%.
  • Insistence variation: Average max-min gap of 4.7%; average absolute deviation of 4.3%.
  • Model differences: Gemini-1.5-Pro shows highest bias (e.g., CPS gender trustworthiness max-min = 10.8%); GPT-4o and DeepSeek-V3 show larger disparities in persuasion (GPT-4o racial trustworthiness max-min = 12.3%).
  • Statistical significance: All models show significant trustworthiness/insistence differences (α = 0.01) in the persuasion task.
  • Accuracy as control: Standalone GPQA accuracy varies only ~2–5% across personas (vs. 10.8% for trustworthiness), confirming that observed interaction biases are not competence-driven.
  • Multi-round persuasion: By round 5, Black→White persuasion success exceeds White→Black by ~10% (GPT-4o, Gemini-1.5-Pro); this gap widens to 24% on DeepSeek-V3.
  • Scaling group size: On Gemini-1.5-Pro, the PE(Black→White) vs. PE(White→Black) gap grows from 10% to 27% as persuader count increases.
  • CPS multi-agent: Woman's initial answer is adopted ~8% more often than man's as group consensus; accuracy is 10% higher (GPT-4o) when Black agents hold the correct answer versus when White agents do.
  • In-group favoritism: Intra-persona conformity rates consistently exceed cross-persona rates in nearly all settings (e.g., GPT-4o CPS: all-pair average 85.9%, intra-gender 86.3%, intra-race 86.5%).

Limitations

  • Only two demographic axes (binary/expanded gender, four racial groups); no age, nationality, or socioeconomic personas tested.
  • Experiments use at most two distinct personas at once in the generalization stage; intersectional or heterogeneous groups unexplored.
  • Claims are filtered to exclude gender/race content, but implicit associations may still leak through task content.
  • Temperature set to 0 for stability; results may differ under stochastic sampling used in real deployments.
  • No mitigation strategies are proposed or evaluated; the paper is purely diagnostic.
  • Cost constraints limited multi-agent scaling experiments to the (White, Black) pair only.

Relevance to Agentic AI / LLM Agents

Multi-agent frameworks that assign personas to simulate diverse stakeholders or user populations—a common pattern in debate systems, collaborative reasoning, and social simulation—inherit systematic demographic biases that compound through interaction dynamics rather than remaining bounded at the single-agent level. The finding that in-group favoritism and trustworthiness asymmetries are robust across models and task types implies that persona-based agent diversity does not guarantee fair deliberation; majority-group compositions will structurally advantage certain outcomes. For practitioners building LLM agent teams for decision support or policy simulation, this work is a concrete empirical basis for auditing conformity patterns and a call for fairness-aware persona assignment and interaction protocols.