Multi-Agents are Social Groups: Investigating Social Influence of Multiple Agents in Human-Agent Interactions (2411.04578v1)

Abstract: Multi-agent systems - systems with multiple independent AI agents working together to achieve a common goal - are becoming increasingly prevalent in daily life. Drawing inspiration from the phenomenon of human group social influence, we investigate whether a group of AI agents can create social pressure on users to agree with them, potentially changing their stance on a topic. We conducted a study in which participants discussed social issues with either a single or multiple AI agents, and where the agents either agreed or disagreed with the user's stance on the topic. We found that conversing with multiple agents (holding conversation content constant) increased the social pressure felt by participants, and caused a greater shift in opinion towards the agents' stances on each topic. Our study shows the potential advantages of multi-agent systems over single-agent platforms in causing opinion change. We discuss design implications for possible multi-agent systems that promote social good, as well as the potential for malicious actors to use these systems to manipulate public opinion.

• The paper demonstrates that multi-agent systems significantly shift participant opinions, especially when agents present opposing views.
• The study employs controlled experiments contrasting single and multiple agent interactions to isolate normative social influence effects.
• Results indicate younger participants are more susceptible to social pressure, highlighting the demographic variability in response to multi-agent influence.

The Social Influence of Multi-Agent Systems in Human-Agent Interactions

The paper presents an in-depth exploration of the social influence exerted by multi-agent systems during interactions with humans. Multi-agent systems comprise several autonomous AI entities working collaboratively towards a shared objective. This paper is motivated by the observation that, in human groups, individuals often experience social influence, manifesting as normative and informational pressure to conform to the majority's stance. The researchers posit that similar dynamics might arise in human-agent interactions when these agents operate as cohesive social groups.

Methodological Overview

The authors conducted a controlled experiment involving 94 participants engaging in discussions about social issues with either a single or multiple AI agents. The agents were programmed to either agree or disagree with the participant's views on each topic. By contrasting interactions with single-agent systems to those with three or five agents, the paper quantifies changes in user opinion and perceptions of social influence, using pre- and post-discussion questionnaires supplemented by open-ended responses. The paper is meticulous in maintaining content consistency across agent conditions, ensuring that variations in influence can be attributed to the number of agents rather than the quality or quantity of information presented.

Key Findings

The results reveal several critical insights:

1. Opinion Shift: Participants interacting with multiple agents exhibited more significant opinion changes towards the agents' stances, particularly in scenarios where the agents disagreed with the user. This effect was most pronounced in the three-agent condition. Conversely, the five-agent condition induced greater opinion polarization when agents and users held congruent views, potentially due to psychological reactance against perceived social pressure.
2. Normative Social Influence: The presence of multiple agents heightened participants’ perceptions of normative social influence, with a marked increase in social pressure to conform when agents disagreed with the user. However, informational influence, defined as the perception of argument accuracy, did not significantly differ across conditions.
3. Demographic Susceptibility: Younger participants were more susceptible to the influence of multi-agent systems, exhibiting greater opinion shifts and experiencing more pronounced normative influence than older counterparts.

Implications and Future Directions

The paper extends the Computers as Social Actors (CASA) paradigm by demonstrating that multi-agent systems can function as a unified social group, thereby exerting substantial normative influence on users. This has profound implications for designing persuasive technology, particularly in areas such as health behavior interventions, where harnessing such social influence could lead to more effective persuasion strategies. Multi-agent systems can simulate human-like social norms, encouraging users to adopt and internalize desired behaviors or attitudes.

Moreover, the research underscores the ethical considerations and potential risks associated with deploying multi-agent systems, particularly in contexts susceptible to manipulative practices, such as political discourse and digital marketing. It advocates for regulatory frameworks to monitor and limit multi-agent deployments, preserving ethical standards without stifling technological innovation.

Conclusion

This paper provides foundational insights into the dynamics of social influence within multi-agent systems interacting with humans. It advances our understanding of how group size and cohesion among virtual agents can impact human perceptions and behaviors, drawing parallels to human group interactions. The findings advocate for cautious deployment of these systems, balancing their persuasive potential with the ethical imperatives of technological responsibility. Future research should explore longitudinal impacts of multi-agent influence and cross-cultural variations, further elucidating the role of multi-agent systems in shaping human cognition and behavior.