The Architecture of Authority in Autonomous Systems

This paper examines a structural mismatch at the core of contemporary AI governance: autonomous systems are increasingly governed using frameworks designed for human-operated software. These frameworks assume that authority resides in actors and is exercised through identity, roles, and permissions. We argue that this assumption no longer holds for systems capable of selecting and executing actions without direct human intervention. In such systems, behavior is not determined by permission chains, but by the architecture, constraints, and admissible state space within which the system operates. Drawing on principles from cybernetics, robotics, and safety-critical control systems, the paper proposes a shift from permission-based governance to structural control paradigms. In this view, governance is not the regulation of actions after they are proposed, but the architectural definition of which states and transitions are possible in the first place. The paper introduces a conceptual distinction between three layers often conflated in governance discussions: the construction of state space, the evaluation of admissibility, and the execution of actions. It argues that governance mechanisms operating solely at execution time are necessarily reactive and insufficient, as they act on outcomes rather than on the structure that makes those outcomes possible. The central claim is that governance failure does not occur when an undesirable action is executed, but when such an action is allowed to be constructible within the system. From this perspective, governance becomes a property of system architecture rather than a function of external policy or permission structures. This work does not propose a complete framework, but outlines a shift in perspective and identifies a set of architectural questions that must be addressed to develop structurally grounded approaches to AI governance.