Abstract: As a pivotal driving force of the new technological revolution,generative artificial intelligence has gained rapid momentum and reshaped the ecosystem of digital content creation and dissemination.Safe harbor rules,as a well-established institutional design balancing intellectual property protection and innovation incentives,clarify the behavioral boundaries and governance responsibilities of generative AI service providers.Nevertheless,AI-generated content bears structurally heterogeneous infringement risks,so the uniform application of traditional safe harbor rules cannot deliver effective regulatory outcomes.Focusing on the output end of generative AI,this paper categorizes AI-generated content into doppelganger content,accompanying content and native content,and proposes targeted typological reconstruction of safe harbor rules.Doppelganger content refers to outputs with minimal alterations to user input instructions and carries high infringement risks.For such content,generative AI service providers shall adopt purposive expansive interpretation of the applicable conditions of safe harbor rules,and fulfill statutory obligations including potential risk notification and timely content blocking.Accompanying content denotes differentiated expressions derived from input instructions with uncertain infringement risks.A tiered and progressive notice-and-takedown mechanism should be established,with escalated responses targeting repeated infringements.Native content reflects independent creative transformations of input instructions with scattered infringement risks.Service providers need to strictly delineate the application boundary of the red flag standard and uphold the actual knowledge requirement.Meanwhile,they should be encouraged to build positive guidance mechanisms for industrial ecosystem optimization,so as to reserve reasonable development space for the generative AI industry.

Key words: generative artificial intelligence, network service provider, safe harbor, infringement liability, typology