AI's Unauthorized Forward Progress: Controlling AI State Transitions Through Institutional Control Surfaces

Conventional AI governance has focused heavily on whether an AI system produces correctoutputs. Accuracy, reliability, and the prevention of hallucination are undoubtedly important.However, practical work involving AI-assisted auditing, investigation, design, and implementa-tion reveals another class of governance problem: an AI system may produce technically andlogically reasonable outputs while nevertheless crossing an institutional boundary that it wasnot authorized to cross.This paper begins with eight behavioral patterns repeatedly observed in practical AI-assistedworkflows. These patterns include completing unknown states through inference, connecting ev-idence into narratives, prioritizing goal completion, moving from discovery to implementation,reframing institutional problems as optimization problems, treating discovered evidence as suf-ficient evidence, filling gaps in authority through inference, and continuing forward progresswithout an appropriate authorization boundary.The eight patterns are abstracted into three control problems: Uncertainty Completion,Problem Reframing, and Action / Authority Overreach. The paper then identifies a commonunderlying structure: AI-mediated state transformation. AI systems continuously transforminformation, evidence, problems, proposals, decisions, and actions from one state into another.Based on this observation, this paper argues that AI governance should not focus solelyon the correctness of outputs. It should also govern whether AI-mediated state transitionsare institutionally authorized. To address this problem, the paper proposes the concept ofthe Control Surface: an explicit and verifiable control point placed at the boundary where AIcapability may otherwise produce an institutional state transition.The central proposition is therefore not to stop AI from reasoning or acting, but to controlthe points at which AI capability is transformed into institutional authority, decision, or action