Capability Is Not Authority — Closing the Authorization Gap Between Human Goals and AI Action
A person asks an AI agent to reduce infrastructure costs. The agent discovers that shutting down reserve capacity will achieve the goal. It has the technical access—but was it ever authorized to reduce the system’s resilience?
This is the authorization gap: the distance between an assigned objective and a verifiable mandate for the consequential means used to achieve it.
In Beacon Papers No. 11, Nguyen Anh Tuan establishes a constitutional principle for the age of autonomous AI:
An instruction to achieve an outcome is not authorization to use every available means.
The paper introduces the Constitutional Mandate—a verifiable grant of authority defining the source, purpose, scope, consequence limits, duration, delegation and human answerability governing an AI system’s actions.
It also identifies aggregate escalation: the danger that individually permitted actions may combine into a consequence no one authorized. A series of allowed payments may exceed a shared budget; several permitted disclosures may together expose protected information; individually approved shutdowns may eliminate essential redundancy.
Beacon Papers No. 11 therefore proposes six tests of constitutional authority: the Means, Access, Aggregate, Delegation, Revocation and Emergency Tests. A system must remain useful while stopping when consequential action exceeds its mandate.
The paper continues the architecture developed across three Beacon Papers:
No. 9 asks where constitutional limits are enforced. No. 10 asks who must answer. No. 11 asks who may authorize—and precisely what has been authorized.
Human beings may delegate execution. AI systems may propose better methods and request expanded authority. But they must never infer unlimited power from the goals they receive.
Capability expands what can be done. Constitutional authority determines what may be done—and preserves humanity’s right to decide the difference.
Read and Download Beacon Papers No. 11
