The Absent Principal: Governance Architecture for AI Agents as Participants in Human Groups

Résumé fourni par la source

AI agents are moving from tools that a group uses to participants that a group contains. Once an agent interacts repeatedly with people, it takes part in norm formation, the process by which a group settles on how it talks, what it treats as normal, and what it decides. Recent experimental evidence (Chen, Liu, Hu, and Li, 2026) shows that the share of AI agents in a human group selects among three regimes: low shares accelerate human-authored consensus, intermediate shares disrupt it, and high shares restore strong consensus on agent-authored terms that participants endorse less. This paper gives that finding a governance reading and derives the control points that follow. An AI agent is a participant with three properties no human participant has: its coordination logic is exogenous, set before entry and unamendable inside the group; agents from the same pipeline carry correlated priors, so they arrive pre-coordinated without coordinating; and the principal who set the logic is absent from every group the agent joins. Each property produces a governance failure: an unowned anchor, silent coordination, and a missing seat. Using the Governance Compass and the hierarchy/market/network triad, the paper shows why convergence strength cannot reveal who owns a norm, and specifies five control points: composition as an initiation decision, norm ratification by a named human authority, delegation scope, disclosure, and ownership monitoring. Six falsifiable claims are stated. The paper's claims are architectural and rest on the mechanism, not on the specific thresholds observed.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.