Abstract: Against the backdrop of the rapid development of generative artificial intelligence, news works, due to their timeliness and authenticity, have become an important corpus for AI training, thereby reshaping copyright governance into a new tripartite framework of “news copyright holders–AI trainers–the public.” However, the large-scale use of news works by AI may erode the copyright interests of news publishers, while also raising challenges regarding the legitimacy of AI training and the safeguarding of public interests. Traditional copyright limitations have shown regulatory failure in the new technological context, and current framework designs likewise struggle to achieve a balance of interests among the three stakeholders. By comparing the approaches of the United States, the European Union, and scholarly proposals in China, this study finds that none of them has fully realized an equitable balance in the context of AI and news publishing. It is therefore suggested to adopt a tiered model of copyright limitations inspired by the “educational needs” framework, and to construct a composite system combining “fair use + statutory licensing + licensing agreements, ” supplemented by collective management organizations, technical standardization, and transparency obligations. Such a framework aims to dynamically balance the protection of publishers’ economic interests, the fulfillment of AI training needs, and the safeguarding of the public interest.