跳转至

文章背景与核心概要

随着人工智能朝着旨在延伸人类注意力、记忆和协调能力的持久型助手方向发展，一个根本性的挑战随之出现：AI应当在何时以及如何独立采取行动？

本文引入了一个关于受控主动能动性（governed proactive agency）的概念与形式化框架。它解决了“激活问题”（activation problem）——即AI如何独立决定某种情况是否需要采取行动、以何种模式进行、以及是执行动作、询问、监控、推迟还是克制。通过将激活机制与常设授权、问责制以及有界个性化联系起来，该研究为共生式AI系统奠定了结构基础，能够在增强人类能力的同时，严格维护人类的权威与判断力。

当智能转化为能动性：共生式AI系统的受控主动能动性理论 (When Intelligence Becomes Agency: A Theory of Governed, Proactive Agency for Symbiotic AI Systems)

作者： João Dias Ferreira
提交时间： 2026年9月7日
主要主题： 人工智能 (cs.AI)
arXiv: 2609.07741
全文访问： 查看 PDF | HTML 版本

执行摘要 (Executive Summary)

随着人工智能朝着旨在延伸人类注意力、记忆和协调能力的持久型助手方向发展，一个根本性的挑战随之出现：AI应当在何时以及如何独立采取行动？

As artificial intelligence shifts toward persistent assistants designed to extend human attention, memory, and coordination, a fundamental challenge emerges: when and how should an AI act independently?

本文引入了一个关于受控主动能动性的概念与形式化框架。它解决了“激活问题”——即AI如何独立决定某种情况是否需要采取行动、以何种模式进行、以及是执行动作、询问、监控、推迟还是克制。通过将激活机制与常设授权、问责制以及有界个性化联系起来，该研究为共生式AI系统奠定了结构基础，能够在增强人类能力的同时，严格维护人类的权威与判断力。

This paper introduces a conceptual and formal framework for governed proactive agency. It addresses the "activation problem"—the challenge of an AI deciding independently whether a situation warrants behavior, in what mode, and whether to act, ask, monitor, defer, or refrain. By linking activation to standing authorizations, accountability, and bounded personalization, the research provides a structural foundation for symbiotic AI systems that augment human capability while rigorously preserving human authority and judgment.

核心概念 (Key Concepts)

  • 激活问题： 决定情况是否需要采取行动、何时介入、使用哪种操作模式，以及是执行动作、请求澄清、静默监控、推迟还是刻意克制的挑战。
    • The Activation Problem: The challenge of deciding if a situation warrants behavior, when to engage, which operational mode to use, and whether to execute an action, ask for clarification, monitor silently, defer, or deliberately refrain.
  • 共生能动性： 在常设、可撤销的授权下被建模为委托（delegation），其特征包括：
  • 与委托人状况的持续耦合。
  • 对委托人状态的校准推断。
  • 有界的个性化与自适应克制。
    • Symbiotic Agency: Modeled as delegation under a standing, revocable mandate, characterized by:
  • Continuing coupling to the principal's situation.
  • Calibrated inference of the principal's condition.
  • Bounded personalization and adaptive restraint.
  • 作为分析单元的行为片段： 通过感知、意图、情感-意动评估、约束和反馈循环，在时间维度上组织行为。
    • Behavioral Episodes as the Unit of Analysis: Organizing behavior across time through perception, intent, affective-conative appraisal, constraints, and feedback loops.

提出的框架贡献 (Proposed Framework Contributions)

  1. 受控主动能动性： 一个将激活决策与授权感知、行为选择、权威限制和可追溯克制联系起来的综合说明。
    1. Governed Proactive Agency: An integrated account linking activation decisions to authorized perception, behavior selection, authority containment, and traceable restraint.
  2. 能动性分类方法： 区分自主能动性与委托能动性层级的系统化方法。
    1. Agency Classification Method: A systematic approach to distinguishing between autonomous and delegated agency tiers.
  3. 评估框架与基准： 旨在评估AI协助是否及时、已授权且可答复的提议场景——超越了简单的任务完成度指标。
    1. Evaluation Framework & Benchmarks: Proposed scenarios designed to assess whether AI assistance is timely, authorized, and answerable—going beyond simple task completion metrics.
  4. 参考架构： 旨在指导始终在线的个人助手和具身支持系统开发的结构蓝图。
    1. Reference Architecture: A structural blueprint intended to guide the development of always-present personal assistants and embodied support systems.

引用与元数据 (Citation & Metadata)