跳转至

文章背景与核心概要

随着人工智能技术的迅速部署，如何实现对AI智能体的持续、有效且民主的治理已成为学术界和工业界面临的核心挑战。本文提出了一种创新的机制设计模型，将治理重心从纯粹的行为约束转向资源分配，利用算力预算使授权决策具备自我执行的特性。

该研究植根于“安全AI范式”（Safe AI paradigm），将算力视为最有效的治理杠杆。通过构建严谨的博弈论框架，该机制允许人类利益相关者持续输入意见，并将其转化为基于硬件实现的算力许可，从而在AI部署方之上建立了一层合规与公共资源管理的叠加层。

资源化权威：部署态AI智能体参与式治理的机制设计模型

Resourced Authority: A Mechanism-Design Model for Participatory Governance of Deployed AI Agents

Authors: Praphul Chandra, Sujit Gujar, Ganesh Ghalme
Primary Subject: Computer Science and Game Theory (cs.GT)
Additional Subjects: Artificial Intelligence (cs.AI), Multiagent Systems (cs.MA)
arXiv Identifier: arXiv:2608.06353 [cs.GT]
Submitted: August 6, 2026

执行摘要

本文引入了一个形式化的机制设计模型，旨在实现对已部署AI智能体的持续性、参与式治理。其核心前提是，AI治理应当通过资源分配而非单纯的行为约束来运行，并利用算力预算使授权机制具备自我执行能力。

该模型围绕安全AI范式（Safe AI paradigm）构建——该范式认为算力可作为有效的治理杠杆——该机制在AI部署方之上充当合规或公共资源管理的叠加层。通过结构化的博弈论方法，人类利益相关者能够持续提供输入，这些输入最终会被聚合为可执行的、由硬件实现的算力许可。

Executive Summary

This paper introduces a formal mechanism-design model aimed at the continuous, participatory governance of deployed AI agents. The core premise is that AI governance should operate through resource allocation rather than purely behavioral constraints, utilizing compute budgets to make authorization self-enforcing.

Framed around the Safe AI paradigm—which posits that compute serves as an effective governance lever—the mechanism functions as a compliance or commons overlay on top of an AI deployer. Through a structured game-theoretic approach, human stakeholders continuously provide inputs that are aggregated into actionable, hardware-realized compute licenses.

治理机制的关键组件

  • 顺序化利益相关者参与： 每个治理周期都被构造成一个扩展型博弈（extensive-form game），其中经过验证的人类利益相关者将按顺序参与。他们通过使用专门的“治理货币”参与支持或反对市场，该货币与AI智能体的算力资源保持刻意的隔离。
  • 资金聚合与有效支持： 资金聚合器将原始的利益相关者贡献转化为考虑广度加权的有效支持（breadth-weighted effective supports）。
  • 带滞后特性的双阈值门槛： 净支持度将通过一个带有滞后特性的双阈值门槛（two-threshold gate with hysteresis），从而将集体情绪转化为二进制的授权决策。
  • 有界耦合映射与安全上限： 授权决策将通过一个受外部认证的安全上限约束的耦合映射，随后释放定额的算力预算。
  • 硬件实现： 输出结果直接在硬件层面以签名的算力许可形式实现，从而保证治理决策能够自我执行。

Key Components of the Governance Mechanism

  • Sequential Stakeholder Engagement: Each governance period is structured as an extensive-form game where verified human stakeholders arrive sequentially. They participate via provision or rejection markets using a dedicated "governance currency" that is kept deliberately distinct from the AI agent's compute resources.
  • Funding Aggregation & Effective Support: A funding aggregator translates raw stakeholder contributions into breadth-weighted effective supports.
  • Two-Threshold Gate with Hysteresis: Net support passes through a two-threshold gate featuring hysteresis, converting collective sentiment into a binary authorization decision.
  • Bounded Coupling Map & Safety Ceiling: Authorization passes through a coupling map bound by an exogenously certified safety ceiling, which subsequently releases a metered compute budget.
  • Hardware Realization: The output is realized directly in hardware as a signed compute license, guaranteeing that the governance decision is self-enforcing.

公开挑战与未来工作

尽管该模型为将算力可用性与公众监督对齐提供了一个强大的框架，但作者们指出了几个关键的挑战：

  • 治理范围： 准确界定能够通过这一特定机制得到有效治理的AI智能体类别。
  • 选民操纵： 隔离并解决受治理的AI智能体自身对治理选民的操纵问题，这构成了该模型中首要的未解难题。

Open Challenges & Future Work

While the model offers a robust framework for aligning compute availability with public oversight, the authors identify several critical challenges:

  • Scope of Governance: Characterizing the precise class of AI agents that can be effectively governed by this specific mechanism.
  • Electorate Manipulation: Isolating and addressing the manipulation of the governing electorate by the governed AI agents themselves, which represents the primary open problem in the model.

链接与资源