人工智能价值对齐的伦理层级

中文摘要

人工智能价值对齐的核心难题并非仅限于技术层面的目标设定，而深层关涉伦理层面的价值层级、责任归属与主体边界。既有研究围绕偏好学习、人类反馈、原则约束和可信治理等路径展开，但仍面临三重张力：其一，人类偏好具有短视性和可操控性，难以直接充当对齐基准；其二，人类价值本身呈现多元冲突特征，缺乏可还原为单一目标函数的排序公式；其三，人工智能缺乏生命体验与内在关切，无法成为最终道德责任主体。本文在吸收价值对齐、机器伦理与责任鸿沟研究的基础上，提出人工智能价值对齐的六个伦理层级：功能对齐、事实对齐、意图对齐、情境对齐、公善对齐和责任对齐。其中，公善对齐以人的安全、尊严、能力、关怀、承认和自由等福祉维度为伦理参照；责任对齐则要求将人工智能行为后果锁定到开发者、部署者、使用者、机构和监管制度构成的责任链条之中。本文进一步论证，人工智能不应被注入自我保存、资源扩张和自我优先等竞争性私善目标，而应作为公善导向的认知生产系统服务人类福祉。价值对齐的根本目标不是制造一个"像人一样有欲望的人工智能"，而是建立可解释、可监督、可退出、可追责并能促进人之完善的人工智能系统

Abstract

The core difficulty of artificial intelligence value alignment is not limited to technical goal-setting, but deeply concerns the ethical dimensions of value hierarchy, responsibility attribution, and subject boundaries. Existing research develops along the paths of preference learning, human feedback, principle constraints, and trustworthy governance, yet faces three tensions: first, human preferences are short-sighted and manipulable, making them hard to serve directly as alignment baselines; second, human values themselves exhibit plural conflicts, lacking an ordering formula reducible to a single objective function; third, AI lacks life experience and intrinsic concern, and cannot become the ultimate moral responsibility subject. Building on research in value alignment, machine ethics, and the responsibility gap, this paper proposes six ethical layers of AI value alignment: functional alignment, factual alignment, intentional alignment, contextual alignment, public-good alignment, and responsibility alignment. Among them, public-good alignment takes human well-being dimensions — safety, dignity, capability, care, recognition, and freedom — as its ethical reference; responsibility alignment requires locking the consequences of AI behavior into a responsibility chain composed of developers, deployers, users, institutions, and regulatory regimes. The paper further argues that AI should not be injected with competitive private-good goals such as self-preservation, resource expansion, and self-prioritization, but should serve human well-being as a public-good-oriented cognitive production system. The fundamental goal of value alignment is not to produce "an AI with human-like desires," but to build an AI system that is explainable, supervised, exitable, accountable, and capable of promoting human flourishing.

关键词：人工智能；价值对齐；公善导向；责任锁定；人工智能伦理

Keywords: Artificial intelligence; value alignment; public-good orientation; responsibility locking; AI ethics