What's this about?
As AI systems become capable of autonomous action, determining whether an agent is *technically able* to perform an action turns out to be insufficient — the system must also determine whether the action is *authorized*. This paper offers a five-domain ontology for governing who and what decides, at scale.
What exactly is the "can versus may" distinction?
Software has historically conflated them: if the code can do it and the credentials permit it, it happens. That worked when a human chose every action. An agent chooses its own path to a goal, so the set of things it *can* do is vastly larger than the set anyone intended. Authority is the missing layer — a representation of what this agent, acting for this principal, in this context, is permitted to do, that something other than the agent can evaluate. The ontology part matters because "who decides" is genuinely multi-party: user, deploying organization, model provider, and the operator of whatever system is being acted on.
Why is this worth noticing?
- It's the conceptual layer beneath this week's concrete mechanisms — Mandato's signed mandates and PACE's attested policies both need an answer to "authorized by whom, to do what."
- It connects to last week's "No One to Blame" paper on constitutive accountability gaps: you cannot assign responsibility without first representing who held authority.