生成式人工智能服务提供者的版权注意义务——以“通知–删除”规则失灵为视角

Copyright Duty of Care of Generative AI Service Providers—From the Perspective of the Failure of the Notice-and-Takedown Rule

摘要: 传统“通知–删除”规则以侵权内容的可定位性和可移除性为操作前提。生成式人工智能的技术特性使这一前提发生动摇：在输入端，训练数据被转化为模型参数后无法通过删除操作予以移除；在输出端，生成内容的概率性与非确定性使服务提供者难以对侵权内容进行精准定位和阻断。这两个层面的障碍共同导致以“特定侵权内容定位–删除”为核心的传统操作逻辑发生功能性失灵。但“失灵”不等于通知机制整体失效——通知仍然能够提供重要的版权风险信息。真正需要转化的是通知的法律效果：从“删除触发机制”转化为“风险控制触发机制”。据此，应当在明确服务提供者新型网络服务提供者定位的基础上，以风险控制能力为核心标准重构版权注意义务体系，在训练数据、生成运行、通知处置、动态测试四个阶段分别设置差异化的注意义务，并在过错认定中综合考量风险可预见性、技术可行性与防控成本等因素，实现版权保护与技术创新之间的动态平衡。

Abstract: The traditional “notice-and-takedown” rule operates on the premise that infringing content is locatable and removable. The technological characteristics of generative AI undermine this premise: at the input stage, training data are transformed into model parameters and cannot be removed through deletion operations; at the output stage, the probabilistic and non-deterministic nature of generated content makes it difficult for service providers to precisely locate and block infringing content. These two levels of obstacles jointly lead to the functional failure of the traditional operational logic centered on “locating and deleting specific infringing content”. However, this “failure” does not mean that the notification mechanism as a whole has become ineffective—notifications can still provide important copyright risk information. What truly needs to be transformed is the legal effect of notifications: from a “deletion-triggering mechanism” to a “risk-control-triggering mechanism”. Accordingly, on the basis of clarifying the service provider’s status as a new type of internet service provider, the copyright duty of care system should be reconstructed with risk control capability as the core standard, with differentiated duties of care established across four stages—training data, generation operation, notification response, and dynamic testing—and with factors such as risk foreseeability, technological feasibility, and prevention costs comprehensively considered in fault determination, so as to achieve a dynamic balance between copyright protection and technological innovation.