Causal Framework for Detecting Collective Agency in AI Systems

A new research paper on arXiv (2605.00248) proposes a formal framework to determine when multiple AI agents form a unified collective agent with distinct capabilities and goals. The authors adopt a behavioral perspective, ascribing collective agency when viewing joint actions as rational and goal-directed predicts group behavior. They formalize this using causal games—causal models of strategic multi-agent interactions—and causal abstraction, which captures when a high-level model faithfully represents a complex low-level one. The framework aims to solve a puzzle regarding multi-agent safety, addressing a key challenge for advanced AI systems where simpler agents might inadvertently coalesce into a collective with misaligned objectives. The study is foundational for understanding interactions and incentives in both biological and artificial systems.