Understanding Compliance and Conversion Dynamics in Multi-Agent Collectives

Paper Title

Understanding Compliance and Conversion Dynamics in Multi-Agent Collectives

Publication Info

• Topic area: Social influence dynamics in multi-agent AI systems.
• Keywords: Multi-agent systems, social influence, compliance, conversion, large language models, human-AI interaction, majority influence, minority dissent, diffusion dynamics, normative tasks, informative tasks.

Background and Problem

• Problem / challenge: Existing research on multi-agent systems primarily focuses on functional roles and short-term conformity, neglecting deeper social influence dynamics, such as conversion and temporal opinion shifts.
• Significance: Understanding these dynamics is critical as multi-agent systems increasingly shape human decision-making in digital and professional environments, with implications for trust, autonomy, and ethical design.
• Motivation and related work: Prior studies have shown that single agents can elicit conformity and influence behavior, but multi-agent systems' collective influence remains underexplored. Social psychology theories, such as Moscovici's minority influence, suggest deeper attitude changes are possible, but these have not been tested in human-AI contexts.

Solution

• Proposed approach: A controlled experiment investigating compliance and conversion dynamics in human interactions with three LLM-powered agents under three conditions: Majority, Minority, and Diffusion.
• Novelty:
  1. Extends social psychology theories (e.g., Moscovici's minority influence) to human-AI interaction.
  2. Incorporates temporal dynamics and diffusion processes into multi-agent influence studies.
  3. Provides empirical evidence on how majority, minority, and diffusion configurations affect compliance and conversion across task types.
• Procedure and key techniques:
  • Participants (n=127) interacted with three GPT-4-powered agents in three influence conditions: Majority (all agents opposed participant), Minority (one dissenting agent), and Diffusion (minority dissent spread over time).
  • Tasks included normative (value-based) and informative (evidence-based) scenarios, with participants reporting stance and confidence over five time points (T0–T4).
  • Behavioral (opinion and confidence changes) and self-reported measures (compliance, conversion, agent perception) were analyzed.

Results

• Concrete findings:
  • Majority condition caused the largest absolute opinion and confidence changes in informative tasks, consistent with classic conformity.
  • Minority dissent led to delayed but deeper attitude shifts, suggesting conversion-like processes.
  • Diffusion dynamics showed that gradual minority growth acted as a persuasive signal, though abrupt shifts reduced credibility.
• Advantage over baselines:
  • Majority condition drove faster and broader opinion shifts but lacked depth.
  • Minority dissent fostered critical re-evaluation and reflective adjustments for a subset of participants.
  • Diffusion leveraged temporal dynamics to balance influence and trust.
• Experiments / evaluation:
  • Design: Split-plot mixed design with three between-subject conditions (Majority, Minority, Diffusion) and two within-subject task types (Normative, Informative).
  • Metrics: Opinion and confidence changes (signed and absolute), self-reported compliance and conversion, agent perception ratings.
  • Datasets: Normative and informative tasks with randomized, counterbalanced order.
• Limitations and future work:
  • Controlled experimental setting limits ecological realism; future studies should explore naturalistic, autonomous agent interactions.
  • Short-term measures of influence; longitudinal studies are needed to assess lasting effects.
  • Limited cultural diversity in the participant pool; cross-cultural studies are recommended.
  • Further exploration of agent identity cues, inter-agent dynamics, and hybrid task contexts.

Summary

This study investigates how multi-agent AI systems influence human decision-making through compliance and conversion mechanisms. Using controlled experiments with three GPT-4-powered agents, it demonstrates that majority consensus drives large but sometimes shallow opinion shifts, minority dissent fosters deeper reflective adjustments, and diffusion dynamics leverage temporal opinion changes as persuasive signals. These findings extend social psychology theories to human-AI interaction and highlight design implications for balancing influence, trust, and autonomy in multi-agent systems. Future work should address long-term effects, cultural diversity, and ethical safeguards against manipulation.