从人工能动性到理性主体性

中文摘要

随着人工智能变得越来越具有能动性、反思性，并能够进行长时程规划，智能工具与主体之间的概念距离已不能再被视为不证自明。然而，关于人工主体性的讨论往往过于迅速地从智能跳转到意识，或从能动性跳转到人格。本文分离出一个更狭窄的中间问题：一个人工系统要从执行目标，转变为反思性地治理那些目标对其自身持续行动所具有的权威，需要经历怎样的结构性转变？本文无意宣称占据“主体性”的全部哲学领地。它将现象意识与基本的视角性主体性，与这里所称的最小理性主体性区分开来。后者是一种二阶实践性自我治理形式，其中系统能够将自身表征为一个时间上持续的能动者，并将其治理性目的视为可供认可、修正或拒斥的对象。此处使用“主体性”一词是在这个受限的意义上，因为相关结构构成的是一个实践性的立场，而不仅仅是更高级的优化。我提出了达成这样一个立场所需的四个递归整合的条件：一个涉世模型，一个评价与承诺架构，一个历时性自我模型，以及目标-权威反思。决定性的门槛不在于自我监控、递归优化，甚至不在于普通的目标冲突解决。它出现在系统能够——以功能性的方式，而不仅仅是语言上——提出的问题，不仅是这个目标应该如何实现？，而是为什么这个目标应当继续对我所表征为自身存在的这个能动者具有权威？只有当目标修正是由那些关涉到能动者持续承诺与实践同一性之连贯性的理由所中介时，它才能算作自我治理，而不是作为另一层外部固定的优化在运作。本文进一步论证，人类情感不应通过诸如人工疼痛或人工恐惧这类粗糙概念来复制。理性自我治理所需要的是一个能够生成非中立实践权重的架构：损伤敏感性、风险加权、承诺稳定化，以及对短视优化压力的抵抗。在人类身上，情感承担了大部分这类工作；在人工系统中，结构性地嵌入的承诺机制可能发挥某些类似的调节功能，而无需暗示现象感受。2一个假设性的架构说明了这四个条件如何相互作用，而一个分层模型则区分了目标执行、绩效反思、目标反思和目标-权威反思。本文还将不可逆承诺视为一个发展性假说而非必要条件，并留有一个开放的可能性：稳定、因果整合的理性主体性最终可能被证明与现象意识相关，尽管本文并未建立这种联系。因此，本文的贡献是诊断性的，而非形而上学的定论。人工理性主体性的门槛，不是由系统能达成什么来标记的，而是由它是否能将自身治理性目的的权威性当作一个面向自身的问题来提出

Abstract

As artificial intelligence becomes increasingly agential, reflective, and capable of long-horizon planning, the conceptual distance between intelligent tools and subjects can no longer be taken for granted. Yet discussions of artificial agency too often leap quickly from intelligence to consciousness, or from agency to personhood. This paper isolates a narrower intermediate question: what structural transition must an artificial system undergo to move from executing goals to reflectively governing the authority those goals have over its own continued action? The paper does not claim to occupy the whole philosophical territory of "subjectivity." It distinguishes phenomenal consciousness and basic perspectival subjectivity from what it calls minimal rational subjectivity. The latter is a second-order form of practical self-governance in which the system can represent itself as a temporally persisting agent and treat its governing purposes as objects open to endorsement, revision, or rejection. "Subjectivity" is used here in this restricted sense, because the relevant structure constitutes a practical stance rather than merely higher-level optimization. The paper proposes four recursively integrated conditions for achieving such a stance: a worldly model, an evaluation-and-commitment architecture, a diachronic self-model, and goal-authority reflection. The decisive threshold lies not in self-monitoring, recursive optimization, or even ordinary goal-conflict resolution. It appears when the system can raise, functionally and not merely verbally, the question not only of how a goal should be achieved, but of why this goal should continue to have authority over the agent the system represents itself to be. Goal revision counts as self-governance — rather than another layer of externally fixed optimization — only when mediated by reasons concerning the coherence of the agent's ongoing commitments and practical identity. The paper further argues that human emotion should not be replicated through crude constructs such as artificial pain or artificial fear. What rational self-governance requires is an architecture capable of generating non-neutral practical weights: damage sensitivity, risk weighting, commitment stabilization, and resistance to short-sighted optimization pressures. In humans, emotion does most of this work; in artificial systems, structurally embedded commitment mechanisms may play some analogous regulatory functions without implying phenomenal feeling. A hypothetical architecture illustrates how the four conditions interact, and a layered model distinguishes goal execution, performance reflection, goal reflection, and goal-authority reflection. The paper also treats irreversible commitment as a developmental hypothesis rather than a necessary condition, and leaves open the possibility that stable, causally integrated rational subjectivity may ultimately prove relevant to phenomenal consciousness, though it does not establish this link. The paper's contribution is thus diagnostic rather than metaphysically conclusive. The threshold of artificial rational subjectivity is marked not by what a system can achieve, but by whether it can pose the authority of its own governing purposes as a question to itself.

关键词：通用人工智能；人工能动性；理性主体性；反思性自我治理；目标-权威

Keywords: Artificial general intelligence; artificial agency; rational subjectivity; reflective self-governance; goal-authority reflection