摘要： 生成式人工智能的崛起标志着工具本质的范式转型。传统法律体系建立在确定性工具的技术逻辑之上，预设使用工具的人类行为与人类意图的线性映射关系，由此形成“设计者可控—使用者可责”的归责框架。然而，生成式人工智能作为或然性工具，其自回归生成机制与概率抽样特性颠覆了工具行为的可预见性原则：神经网络的参数空间阻断了设计意图的完整渗透，强化学习的动态调优消解了使用指令的确定性约束，技术应用呈现去中心化扩散效应。这种技术本体论层面的变革，使得传统法律中雇主责任、产品责任与平台责任的三重归责体系遭遇结构性危机。通过解构大语言模型的系统性幻觉机制，可以揭示算法黑箱与人类认知框架的深层张力，进而提出基于过程导向的动态合规路径——算法透明度强化、以人类为中心的审计机制与职业伦理再造，在技术创新与风险防控之间重构规范性平衡。这一探索最终指向法律科学从机械决定论向概率主义范式的认知跃迁，为人工智能时代的制度调适提供理论框架。
关键词:
生成式人工智能,
或然性工具,
算法幻觉,
法律主体性,
归责体系重构
Abstract:
he rise of generative artificial intelligence signifies a paradigm shift in the nature of tools. Traditional legal systems, based on the technical logic of deterministic tools, presuppose a linear mapping relationship between human behavior and human intent when using tools, leading to an attribution framework of “designer controllability user accountability”. However, generative AI, as a probabilistic tool, with its autoregressive generation mechanism and probabilistic sampling characteristics, subverts the principle of predictability of tool behavior. This ontological transformation at the technological level creates a structural crisis in the traditional triple attribution system of employer liability, product liability, and platform liability in law. By deconstructing the systematic illusion mechanism of large language models, this paper reveals the deep tension between algorithmic black boxes and human cognitive frameworks, and then proposes a dynamically compliant path based on process orientation. This exploration ultimately points to a cognitive leap in legal science from mechanistic determinism to a probabilistic paradigm, providing a theoretical framework for institutional adaptation in the age of artificial intelligence.
Key words:
generative artificial intelligence,
probabilistic tool,
algorithmic illusion,
legal subjectivity,
reconstruction of attribution systems
郑戈. 作为或然性工具的人工智能及其法律挑战[J]. 探索与争鸣, 2025(8): 32-46.
Zheng Ge. On Generative Artificial Intelligence as a Probabilistic Tool and Its Legal Challenges[J]. Exploration and Free Views, 2025(8): 32-46.