The deep involvement of generative AI (GenAI) in content production has severely impacted the existing internet copyright legal framework, making the allocation of infringement liability for GenAI service providers a core challenge that urgently needs to be addressed. As the cornerstone of balancing technological innovation and copyright protection in the internet era, the safe harbor rules are currently facing a severe test regarding their applicability to this new subject. This paper conducts an in-depth exploration of this issue from the perspective of the suitability of infringing subjects.From the history of institutional evolution, the application of the safe harbor rules highly depends on the suitability of subjects, having undergone the separation of internet service providers (ISPs) and internet content providers (ICPs), as well as the rise of content service providers. However, GenAI technology has achieved a leap from discovering information to reorganizing information and even creating information. GenAI service providers directly participate in the content production process, breaking the neutrality premise of traditional ISPs that do not participate in content creation. Meanwhile, because its content generation is subject to user instructions, and the cross-modal, large-scale production makes it difficult to trace the source of the content, it cannot be classified as an ICP within the meaning of copyright law. Therefore, the traditional dual-subject framework of internet service providers and content providers has fallen into a dilemma of applicability in the GenAI era.If GenAI service providers are forcibly regarded as traditional ISPs and the existing safe harbor rules are directly applied, it will plunge them into a dilemma of compliance. On the one hand, constrained by the technological essence of GenAI as a pen rather than paper, it cannot fulfill the duty of care to prevent the dissemination of infringing works. On the other hand, since AI-generated content is generated in real-time based on parameters rather than stored in a database in advance, service providers are technically unable to take necessary measures such as disconnecting links or deleting content under the notice and takedown rules. Therefore, the key to breaking the deadlock lies in abandoning the patching approach and legally recognizing GenAI service providers as a brand-new, independent liable subject. In determining their infringement liability, the focus should shift from the traditional safe harbor rules’ presumption of actual knowledge （informed or not) regarding dissemination to the standard of knowability (knowable or not) regarding content generation.Based on this, this paper proposes that a new AI safe harbor rule exclusively tailored for such subjects should be scientifically introduced and reconstructed within the framework of content producer liability under the Interim Measures for the Management of Generative Artificial Intelligence Services. Specifically, it includes three core mechanisms: First, establishing an AI training exemption mechanism, explicitly treating the technical processing behavior of using lawful works for model training as non-infringing use. Second, establishing a knowable exemption rule, requiring providers to ensure that the generated content and its sources are in a knowable state by adding visible and invisible watermarks and providing source retrieval links, and using this as a basis for liability apportionment and defense. Third, establishing a circumvention exemption rule, encouraging service providers to proactively deviate from training data through model optimization, rewriting, and security assessments to evade infringement risks. The construction of this set of brand-new rules aims to effectively maintain the institutional balance of the existing copyright order while safeguarding the continuous innovative development of GenAI technology.