AI Should Not Act Beyond Its Permission: A Governance Framework

The question: when should AI be allowed to act?

AI systems increasingly shape consequential decisions, but predictive accuracy alone does not settle whether a system should have authority to act. That distinction motivates Na Young Ahn’s conceptual study, Just AI Governance and the Forbidden-Boundary Principle.

The paper connects broad questions about justice and rights with a practical governance question: how should an AI agent be prevented from executing consequential actions beyond its authorized limits? Its approach combines an assessment framework with a proposed protocol for controlling execution.

A two-stage approach to governance

The proposed framework covers four layers: normative principles, data, models, and governance. It maps six justice requirements to observable evidence, decision rights, and remedies. The aim is to connect ethical requirements to practical questions: what evidence would matter, who gets to decide, and what recourse might be available?

The paper proposes a two-stage evaluation rule. First, exclude uses that fail independently justified minimum rights constraints. Then compare the remaining options across multiple performance and governance criteria. Under this approach, rights act as a threshold rather than one more score that strong performance might outweigh.

The supplied abstract does not give the names or details of the six requirements, or report an application of the framework to a real system. Readers therefore cannot assess those particulars from the available excerpt; this does not establish what the full paper contains on those points.

The “Forbidden-Boundary Principle”

For operational control, the paper proposes the Forbidden-Boundary Principle (FBP): an AI agent must not execute externally consequential actions outside pre-authorized boundaries. The author invokes Genesis chapter 2 as an explicitly limited normative analogy. The abstract does not present that analogy as an engineering proof.

The accompanying First-Principle Safe-Stop Protocol (FPSSP) is described as comprising an independently governed policy monitor, a mandatory execution gate, a restricted or alternative safe-state controller, audit trails oriented toward immutability, and accountable human authorization for recovery. In practical terms, the proposal is to monitor policy, check actions before execution, provide a constrained alternative state, preserve records, and require accountable human authorization when recovering.

The abstract distinguishes this proposal from existing Simplex and runtime-assurance architectures. It does not claim a novel safety switch or a formal proof of universal ethical compliance.

What the paper establishes—and what it does not

This is a conceptual study. The abstract says the author provides a policy schema, threat model, test matrix, outcome definitions, and a prospective comparison design intended to make the proposal falsifiable. It reports no FPSSP implementation, experimental dataset, numerical improvement, or verified latency bound. The abstract therefore offers no reported operational-safety evidence for the protocol.

These limits matter: a design and a plan for testing are not results from testing. As a general consideration for any such system, the effectiveness of a policy monitor and execution gate would depend on how boundaries are specified and enforced. The abstract does not report tests of these controls under real-world failures or adversarial conditions, so their behavior in those circumstances cannot be assessed from the supplied evidence.

The source available here is the arXiv abstract rather than the full paper, limiting what can be checked about the framework’s details and proposed evaluation. The source record identifies the work as an arXiv item; the supplied information does not establish peer review or journal publication. On this evidence, its contribution is a governance framework and a testable research agenda—not a demonstration that AI systems are safe or just under the proposed rules.