Limits of Safe AI Deployment: Differentiating Oversight and Control

Oversight and control are fundamentally different functions that shouldn't be conflated: The paper distinguishes control as an ex-ante or real-time operational mechanism aimed at preventing failures, while oversight is an ex-post policy and governance function focused on detection and remediation. This distinction matters because regulatory texts like the EU AI Act often use "human oversight" ambiguously, which could undermine meaningful implementation.

Current regulatory language risks being unenforceable without clearer definitions: The paper warns that vague terminology in regulations creates inconsistent interpretations of supervision requirements, potentially making it impossible to design or evaluate whether AI systems actually operate under meaningful human supervision as intended by policymakers.

Not all supervision mechanisms are equally feasible in practice: The framework proposed in the paper articulates specific conditions under which oversight and control are technically and organizationally possible, acknowledging that there are inherent limits to what each mechanism can achieve—a critical insight for regulators setting realistic expectations.

Supervision methods must be systematically documented and integrated into risk management: Rather than treating human oversight as a checkbox compliance item, the paper emphasizes that supervision approaches should be formally documented and connected to broader organizational risk management practices, drawing on maturity models to ensure meaningful implementation.