AI Authority Control Framework: A Sociotechnical Architecture for Bounding Consequential AI Authority

Advanced AI systems can generate substantial benefits in science, engineering, medicine, productivity, and public services while also creating misuse, malfunction, systemic, and loss-of-control risks. This paper proposes the AI Authority Control Framework (AACF), a sociotechnical framework for bounding consequential authority rather than treating model capability alone as the central safety variable. AACF distinguishes Granted, Obtainable, Deferred, and Systemic Authority; separates technical execution authority (AAL) from systemic influence (SIL); and specifies external control-plane, credential-mediation, authority-budget, safe-transition, human-oversight, governance, and portfolio controls. It introduces hard deployment gates for self-escalation, self-governance, uncontained persistence or replication, unilateral catastrophic authority, and other configurations that could make meaningful human control practically unrecoverable. The framework is complementary to alignment: alignment seeks to reduce the probability that dangerous behavior is attempted, while authority control seeks to limit the consequences of error, misuse, deception, or emergent behavior. To challenge the framework, the paper reports two rounds of adversarial theoretical review and retrospective validation against documented failures in software, cybersecurity, aviation, critical infrastructure, and agentic AI. It also defines a staged validation program culminating in controlled sandbox experiments that compare prompt-only safeguards with externally enforced controls. The central proposition is that advanced AI can become more useful without automatically receiving greater authority when access, permissions, resources, delegation, and consequential actions remain independently bounded and auditable across the surrounding sociotechnical system.