生成式人工智能幻觉致损的侵权法回应

Tort Liability for Generative AI Hallucination

摘要: 生成式人工智能幻觉是指模型生成看似合理、实则与事实不符的内容的现象，是大语言模型基于概率预测机制运行的内生现象。这一技术现实对传统侵权规则提出了挑战。为能够在现行侵权法体系内回应生成式人工智能幻觉致损问题，本文首先阐明了人工智能幻觉的技术特质及其引发的侵权法困境，继而在梳理无过错责任、产品责任与过错责任等学说争议的基础上，论证了适用过错责任原则的正当性。在过错认定层面，本文提出了应以合理注意义务为核心，并结合应用场景、服务阶段与信息类型对服务提供者的注意义务进行动态调整；同时亦应当以动态的、多层次的标准综合考虑使用者的过错。上述方案旨在为生成式人工智能幻觉侵权这一新型侵权形态的司法裁判提供可操作的判断基准。

Abstract: Generative AI hallucination refers to the phenomenon in which models generate seemingly plausible but factually inaccurate content. It is an inherent byproduct of large language models operating on probabilistic prediction mechanisms. This technological reality poses challenges to traditional tort rules. In response to the tort law issues arising from hallucination-induced harm, this article first clarifies the technical characteristics of AI hallucination and the resulting doctrinal difficulties. It then examines competing theories including no-fault liability, product liability, and fault liability, ultimately demonstrating the propriety of the fault liability principle. Regarding the determination of fault, this article proposes that reasonable duty of care should serve as the core standard, with dynamic adjustments based on application scenarios, service stages, and information types, and that, on the basis of such a dynamic framework, the user’s contributory fault should also be taken into account comprehensively. The proposed framework aims to provide an operable analytical benchmark for judicial adjudication of this emerging type of harm caused by generative AI hallucination.