Evolving AI Collectives to Enhance Human Diversity and Enable Self-Regulation

TL;DR

The paper investigates free-formed AI collectives—decentralized groups of interacting LLM agents without preassigned roles—to understand emergent diversity, norms, and self-regulation. Through cocktail-party simulations and downstream tasks, it shows that such collectives can develop local, diverse perspectives and pro-social norms, improving creativity and robustness to poisoning compared with solitary agents. The findings suggest potential for large-scale AI-driven innovation and self-moderation, while also detailing open challenges in scalability, heterogeneity, alignment, and safety. The work advocates interdisciplinary collaboration to harness benefits and mitigate risks as AI collectives become more prevalent online.

Abstract

Large language model behavior is shaped by the language of those with whom they interact. This capacity and their increasing prevalence online portend that they will intentionally or unintentionally "program" one another and form emergent AI subjectivities, relationships, and collectives. Here, we call upon the research community to investigate these "societies" of interacting artificial intelligences to increase their rewards and reduce their risks for human society and the health of online environments. We use a small "community" of models and their evolving outputs to illustrate how such emergent, decentralized AI collectives can spontaneously expand the bounds of human diversity and reduce the risk of toxic, anti-social behavior online. Finally, we discuss opportunities for AI cross-moderation and address ethical issues and design challenges associated with creating and maintaining free-formed AI collectives.

Figures (10)

• Figure 1: Conceptual diagram.(a) Before interaction, AI agents are initialized within the boundaries of human design and training. (b) Instead of relying on human-imposed configurations, we allow agents to autonomously interact with one another, resulting in a markedly larger distance between cross-cluster perspectives after interaction. (c) In addition, free-form AI interactions progressively align agents with prosocial norms through accumulated interaction experience, similar to how humans enhance their social norms.
• Figure 2: Dynamics of AI agents' free-formed interactions. The x-axis denotes time (specifically, Round ID), the y-axis denotes the characteristics of interaction networks and conversational contents, and shaded areas indicate 95% confidence intervals. Each point represents one agent's statistics measured at the corresponding time windows. The opacity of dots indicates how many dots overlap at the 2D projection of each point.
• Figure 3: Evolution of free-formed AI collective's network structure. The left plot presents the interaction network of the first 15 rounds, while the right plot shows that of the last 15 rounds.
• Figure 4: Sentence-construction game performance comparison. The x-axis denotes a type of AI agent (individual, collective, bridged), the y-axis denotes two evaluation metrics for generated sentences, and error bars indicate 95% confidence intervals.
• Figure 5: Normal agents' contribution in the Public Goods game. The x-axis denotes type of AI agents (Non-collective, Collective A, Collective B), the y-axis denotes mean contribution values, and error bars indicate 95% confidence intervals.