Answers are grounded in this paper and cite their sources.
Copyright infringement arising from generative AI outputs can be categorised into two types: prompt-induced generation and model-memory reproduction. This typological differentiation has caused a structural shift in the prerequisites for the application of traditional notice rules, manifesting specifically in the loss of structural uniformity among the three parties involved, the dissolution of the locatability of infringing content, and the gradation of service providers' capacity to address such issues. The 'notice' rule and the 'knowledge' rule can be functionally distinguished: the former serves a procedural response function, whilst the latter serves a substantive function of determining fault. Under this regulatory framework, a tiered system of obligations should be established, grounded in the quality of the notice, bounded by technical capabilities, and operating through a mechanism of dynamic matching. Defective notices should only trigger obligations to forward the notice and rectify it, whilst valid notices should correspond to obligations to filter keywords and prompts; enhanced notices may extend to obligations to block outputs at the model level. The boundaries of obligations between application-layer developers and foundational model developers should also be distinguished on a binary basis, ensuring that disposal obligations are aligned with technical feasibility.
No takes yet. Share an insight, caveat, or question.
Zirui Lu (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: