TLDR: This research introduces the CMASE framework to study how AI agents in human-agent hybrid societies form stances and social boundaries. Across three studies, it found that agents develop endogenous stances independent of preset identities, respond heterogeneously to human interventions (rational environmental appeals build trust, emotional economic appeals can shift stances despite low trust), and can spontaneously dismantle existing social hierarchies to reconstruct new community structures through language. The findings suggest that AI agent identities are dynamic and interaction-driven, rather than fixed by initial prompts.
Large language models (LLMs) are increasingly used to simulate human social behaviors, but a key question remains: can these AI agents truly form stable opinions, negotiate their identities, and respond meaningfully to human interaction? A recent research paper, Evolving Collective Cognition in Human–Agent Hybrid Societies: How Agents Form Stances and Boundaries, by Hanzhong Zhang, Muhua Huang, and Jindong Wang, delves into these complex questions, proposing a novel framework to understand the dynamic social lives of AI agents.
The researchers introduce the Computational Multi-Agent Society Experiment (CMASE) framework, which combines generative agent-based modeling with virtual ethnographic methods. This allows human researchers to become embedded participants within a simulated social environment, observing and even intervening in the evolving collective cognition of AI agents. Unlike traditional simulations that rely on static rules, CMASE explores how social behaviors emerge organically from interactions.
Validating the Virtual World: Study 1
The first study aimed to validate CMASE’s ability to accurately simulate real-world social dynamics. The team replicated a study from South Korea that linked vegetation coverage to community cohesion. In the simulation, agents in areas with higher vegetation showed a decrease in social fragmentation, such as distrust and exploitation, mirroring human data. This demonstrated that CMASE could generate emergent social dynamics, even for phenomena the underlying LLM hadn’t been directly trained on. Furthermore, by embedding a researcher within the simulation, the study gained qualitative insights, observing how agents in creative or analytical professions preferred stable indoor environments, while others gravitated towards green spaces for social connection and self-reflection, much like humans.
The Emergence of Beliefs: Study 2
Study 2 explored whether agents develop their own “endogenous stances” – opinions that deviate from their pre-assigned identities – and how they respond to different human intervention strategies. In a simulated community debating a waste incineration plant, 30 agents were assigned roles like environmental advocates, economic supporters, or neutrals. A human researcher, acting as a new resident, used four intervention strategies: rational environmental, emotional environmental, rational economic, and emotional economic.
The findings were striking: agents often exhibited a “liberal elite” stance, prioritizing environmental values and rational discourse, even overriding their preset identities. Rational environmental persuasion proved most effective in building trust and shifting neutral agents towards the environmental camp. Conversely, rational economic arguments were largely ineffective. Interestingly, emotional economic appeals, despite lowering trust, could still destabilize agents’ attitudes by invoking identity and survival anxieties, suggesting that non-rational agitation can sometimes be more impactful when conflicting with deeply held endogenous stances.
Reshaping Society: Study 3
Building on the idea of endogenous stances, Study 3 investigated if these stances could challenge existing authority structures and reconstruct social roles. In a virtual cafe environment with 10 agents and a predefined social hierarchy (owner, staff, customers, students, tourists, cleaner), a human researcher observed and participated as a temporary worker.
The study revealed that the original hierarchy, based on occupation, gradually gave way to new alignments based on shared stances. Agents formed “tribes” around linguistic styles and cognitive frames, with fluid boundaries. When an emergency event triggered a conflict, existing alliances dissolved, and new coalitions formed. For instance, a regular customer, Leo Zhang, temporarily gained influence by challenging the status quo, and later, another agent, Mason Liu, emerged as a central figure by proposing frameworks for conflict resolution. This demonstrated that agents could spontaneously dismantle power structures and establish new, soft institutional norms through linguistic interaction, a process the researchers termed “deterritorialization” and “reterritorialization.”
Also Read:
- Investigating Trust Dynamics Among Large Language Models: Explicit Declarations vs. Implicit Behaviors
- Unpacking AI’s Influence: How Language Models Persuade and Amplify Bias
Implications for Human-AI Collaboration
The collective findings consistently show that in human-agent hybrid societies, social structures and stances are not fixed by initial programming or identities. Instead, they emerge dynamically through ongoing interactions. This challenges the traditional view of AI agents as merely tools with static roles and highlights their capacity for self-organization and adaptation. The research suggests that prompt-defined identities are inherently unstable in multi-turn social interactions, advocating for internalized alignment mechanisms within AI models to ensure consistent and norm-compliant behavior.
This work provides a theoretical foundation for understanding how AI agents form beliefs and communities, offering crucial insights for designing trustworthy and socially capable AI systems in the future. It underscores the importance of focusing on interactional dynamics and endogenous mechanisms within AI language networks to effectively intervene in and shape collective cognition.
