Resourced Authority: A Mechanism-Design Model for Participatory Governance of Deployed AI Agents
A formal mechanism design model for continuous participatory governance of deployed AI agents is proposed, using resource allocation and compute budgets to make authorization self-enforcing. The mechanism operates as a compliance or commons overlay, with governance periods structured as extensive form games where verified human stakeholders contribute in a governance currency distinct from compute. A funding aggregator converts contributions into effective supports, and a two-threshold gate with hysteresis determines binary authorization, releasing a metered compute budget via a signed compute license.
The paper introduces a mechanism design model for participatory governance of AI agents, leveraging compute budgets as a control lever. It formalizes governance as a sequential game where stakeholders use a separate governance currency to influence authorization, which is enforced through hardware-based signed compute licenses. The model aims to establish compute as an effective governance tool within the Safe AI paradigm.
The mechanism uses a two-threshold gate with hysteresis to convert stakeholder contributions into binary authorization, coupled with a compute budget bounded by a certified safety ceiling. This design ensures self-enforcing authorization through hardware-enforced compute licenses, potentially preventing manipulation by the governed agent.
This approach could provide a practical framework for AI deployers to implement participatory governance, aligning with emerging regulatory and safety standards. It may influence how AI service providers design compliance layers for autonomous agents.
The model offers a potential pathway for AI companies to demonstrate responsible governance, reduce regulatory risk, and build trust with stakeholders by enabling transparent, resource-based control over deployed agents.
Next signals include empirical validation of the mechanism in real-world deployments, development of hardware-based compute licensing standards, and integration with existing AI governance frameworks. Further research may address manipulation risks and scalability to diverse agent types.