Abstract: How to properly handle the relationship between copyright protection for works and the training of AIGC large-scale models has become the “Achilles’ heel” in achieving high-quality iterative updates for AIGC large-scale models. The “right paradigm” approach to front-end regulation, which involves reasonable use and statutory licensing, each has its shortcomings, making it difficult to achieve the institutional goal of balancing the development of the AIGC large-scale model industry with the protection of copyright holders’ interests. The “relationship paradigm”, focusing on terminal output control, can become the foundational institutional framework for regulating AIGC large-scale model training in our country. When users engage in copyright infringement using AIGC large-scale models, if the model provider cannot prove that it has actually fulfilled the duty to prevent output plagiarism and stop repeated infringement during the data training process, it should bear third-party infringement liability for the loss of interests of copyright holders.