人工智能在放大什么？——被放大的理性与取向敏感型治理

中文摘要

人工智能治理日益不仅关注安全与权利，也关注人的繁荣与积极对齐。然而，以结果为导向的语言留下了一个过程层面的问题尚未充分展开：人工智能系统持续把人的注意力、情感、判断、能动性和社会关系推向哪些方向？本文提出“取向敏感型治理”作为一项补充性分析层次。所谓“取向”，指人工智能持续把人的注意力、情感、判断、能动性和社会关系推向某一方向的力量；它是一种方向性压力，而非繁荣的终态。本文将人工智能理解为“被社会性地组织起来的理性放大机制”：分类、推理、预测、优化、生成和决策支持在制度层面被规模化，但不因此拥有人的理性。文章进一步提出，理性本身不内在道德，它可以服务于私善、私恶、公善与公恶四种方向。人工智能不创造新的道德方向，却会放大、自动化并制度化这些方向。为识别这些方向，本文发展出一个四层八维的具身-情感取向框架，涵盖具身脆弱性、调节完整性、社会取向和被放大的认知。该框架使治理能够在有害轨迹凝结为可测量的下游结果之前就识别它们——例如，当参与度优化逐步强化奖赏捕获、依赖、恐惧、敌意或判断外包时，即使系统仍然可靠且合规。文章最后提出取向敏感型治理的原则、工具箱、多层治理框架与研究议程，主张治理不仅要问人工智能是否安全、可控或繁荣，还要问它正在放大哪一种理性服务方向

Abstract

AI governance increasingly attends not only to safety and rights, but also to human flourishing and positive alignment. However, outcome-oriented language leaves underexplored a process-level question: in which directions do AI systems persistently push human attention, emotion, judgment, agency, and social relations? This paper proposes "orientation-sensitive governance" as a supplementary analytical level. "Orientation" refers to the force by which AI persistently pushes human attention, emotion, judgment, agency, and social relations in a certain direction; it is a directional pressure, not an end-state of flourishing. The paper understands AI as "socially organized amplification machinery of reason": classification, inference, prediction, optimization, generation, and decision support are scaled at the institutional level, but without thereby possessing human rationality. It further argues that reason is not intrinsically moral; it can serve four directions — private good, private evil, public good, and public evil. AI does not create new moral directions, but it amplifies, automates, and institutionalizes them. To identify these directions, the paper develops an embodied-affective orientation framework of four layers and eight dimensions, covering embodied vulnerability, regulatory integrity, social orientation, and amplified cognition. The framework enables governance to identify harmful trajectories before they crystallize into measurable downstream outcomes — for example, when engagement optimization progressively reinforces reward capture, dependence, fear, hostility, or judgment outsourcing, even while the system remains reliable and compliant. The paper concludes by proposing principles, a toolbox, a multi-layer governance framework, and a research agenda for orientation-sensitive governance, arguing that governance must ask not only whether AI is safe, controllable, or flourishing, but also which service direction of reason it is amplifying.

关键词：人工智能；人工智能治理；被放大的理性；取向敏感型治理；理性四象限；人的

Keywords: Artificial intelligence; AI governance; amplified reason; orientation-sensitive governance; four directions of reason; human flourishing