2022年12月，生成式人工智能ChatGPT正式发布，它不仅能与用户进行高效对话，还能生成邮件、小说等各类文本。ChatGPT引发了生成式人工智能技术的潮流，相关产品相继面世并实现从单一的文本、图像生成到多模态应用的快速迭代。生成式人工智能的功能实现和技术发展建立在使用海量作品训练模型的基础上，其在极大便利创作、丰富文化生活的同时，也给著作权法带来了巨大挑战。生成式人工智能使用大量受著作权保护的作品进行模型训练引发的著作权问题成为各方争议的焦点，其是否构成著作权侵权以及如何在法律规制下合理进行亟待明确。著作权领域的作品使用路径中，授权许可模式下大规模作品使用需要的高昂交易成本和繁琐操作程序导致了“制度失灵”；合理使用和法定许可都是通过对著作权人权利的限制实现减少使用成本和简化交易程序的目的，能够满足生成式人工智能对海量作品的需求，两种制度在设计和适用上存在明显差异，前者侧重技术发展，后者强调利益均衡，究竟哪一制度才是当下生成式人工智能使用作品的合理路径，各方并未达成共识。本文在详细分析生成式人工智能使用作品的著作权侵权风险的基础上，从正当性和制度适用层面对合理使用和法定许可进行分析，参考国内外该领域立法司法经验，探索构建我国生成式人工智能使用作品的合理路径。

第一章对生成式人工智能的概念和主要类型进行介绍，在简单说明以ChatGPT和Stable Diffusion为代表的文本和图像生成式人工智能技术原理的基础上，探讨生成式人工智能在各阶段使用作品产生的不同的著作权侵权风险，并对我国已生效的生成式人工智能著作权侵权判决进行分析，得出初步结论：生成式人工智能在数据输入、模型训练阶段未经授权使用作品可能构成著作权侵权，但其定性存在较大争议；其在内容输出阶段的生成内容侵权问题则可沿用“接触+实质性相似”的规则进行判断。

第二章对授权许可、合理使用、法定许可这三种著作权法下使用作品的路径进行正当性和制度可行性分析，得出如下结论：授权许可模式意味着高昂的费用和繁琐的程序，会过分加重开发者的成本，因而无法满足生成式人工智能对海量作品的需求；合理使用制度虽然能便于人工智能开发者广泛获取作品进行模型训练，有利于技术进步，但其同时剥夺著作权人许可权和获酬权的方式会不合理地损害著作权人的合法权益，与著作权法的公平精神背道而驰，并不具备法理层面的正当性，且使用行为并不能满足“三步检验法”或构成“转换性使用”而落入合理使用的范畴；法定许可制度允许未经著作权人许可使用作品，但应当支付相应的费用，虽剥夺著作权人的许可权但确保其能从作品使用行为中获益，既能实现对著作权的保护又能促进生成式人工智能以使用海量作品为基础的快速发展，因此法定许可是当前解决生成式人工智能使用作品问题的最佳选择。但现行的法定许可制度不能很好地应对该问题，需要完善制度内容和配套措施以促进法定许可制度发挥价值。

第三章对欧盟、日本、美国的立法司法动态进行关注，从中认识到以欧盟文本与数据挖掘例外为代表的合理使用立法尝试并不能妥善解决当前生成式人工智能使用作品引发的著作权问题。日本的立法举措体现了促进人工智能发展的立场，但相关规定可能造成著作权人利益受到严重损害，因此引发了巨大争议。美国尚无人工智能使用作品著作权问题的针对性立法，但司法实践中多起生成式人工智能著作权纠纷的进程反映出美国法院对待该问题的谨慎立场。此外，各国相继开展人工智能监管也表明了全球范围内对生成式人工智能采取法律规制的趋势，其中包括对提供生成式人工智能服务的透明度要求在内的部分内容值得借鉴。

第四章再次强调法定许可制度可以兼顾著作权人利益保护和人工智能技术发展，是生成式人工智能使用作品的合理路径。建议新增生成式人工智能以模型训练为目的获取、使用作品的法定许可，并围绕使用主体、使用方式等方面进行细化；同时建议完善配套措施以促进法定许可制度的实施，包括衔接集体管理制度、建立市场化的付酬机制以及构建著作权人权益保障机制。此外还建议规定生成式人工智能开发者和服务提供者提高作品使用行为透明度、优化模型设计和加强对用户使用行为的审查和管理等合理义务。

The generative artificial intelligence (for short, generative AI) ChatGPT was officially released in December 2022. It can make conversations with users and generate various texts such as emails and novels according to the prompt. ChatGPT triggered a trend of generative AI technology. All kinds of products have been launched one after another and achieved rapid iteration from single generation function to multi-modal applications. The function and technological development of generative AI are based on the exploitation of massive works to train models. While it greatly facilitates creation and enriches our cultural life, it also brings huge challenges to copyright law. The unauthorized exploitation of copyrighted works for training purpose has become the focus of controversy on all sides. Whether such behavior constitutes copyright infringement and how to exploit works reasonably under legal regulations need to be clarified urgently. Among the paths of exploiting works, obtaining licenses from all copyright holders means high transaction costs and cumbersome operating procedures, thus it’s not applicable for large-scale exploitation of works; both fair use and statutory license allow the unauthorized exploitation of copyrighted works which can reduce usage costs and simplify transaction procedures. Thus, they can meet the needs of works for generative AI. The difference between them is that fair use means exploiting works for free while statutory license is based on paying remuneration to copyright holders. The former focuses on technological development, while the latter emphasizes the balance of interests. Based on a detailed analysis of the copyright infringement risks of generative AI’s exploitation of works, this article did a comprehensive analysis about whether fair use or statutory license can be suitable for generative AI’s exploitation of works, and referred to domestic and foreign legislative and judicial experience in this field, then explored a reasonable path for generative AI’s exploitation of works.

The first chapter introduced the concept and main types of generative AI. A brief explanation of the principle of text and image generation was shown. Then the analysis was focus on possible copyright infringement at different stages, and a preliminary conclusion was drawn: during the data input and training stages, generative AI’s unauthorized exploitation of works may constitute copyright infringement; during the content output stage, whether output constitutes copyright infringement can be judged according to general rules for determining copyright infringement.

The second chapter analyzed three ways to exploit work according to copyright law: license, fair use, and statutory license, then drew the conclusion: licensed exploitation fails to meet the massive demand for works of generative AI. Although fair use can facilitate AI developers to obtain enough works for training thus play a major role in promoting technology advancement, it will seriously infringe the legitimate rights and financial interests of the copyright holders which is against the fair spirit of copyright law. Besides, neither “three-step test” nor “transformative use” principle can be satisfied. Statuary license allows the use of works without permission, but reasonable fees must be paid. It can not only protect the benefit for the copyright holders, but also promote rapid development of AI based on the exploitation of massive works. Therefore, statutory license can be a reasonable path for generative AI to exploit works.

The third chapter conducted research of the legislative and judicial developments in the European Union, Japan, and the United States. EU text and data mining legislation cannot solve the copyright issues well. Japan’s legislation reflects its stance on promoting the development of AI, but it may harm the copyright holders’ right. It has caused a huge controversy. The United States does not yet have targeted legislation on copyright issues in works using AI, but the process of multiple generative AI copyright disputes reflects the cautious stance of U.S. courts on this issue. The successive implementation of AI regulation in various countries shows the trend of legal regulation of generative AI on a global scale. Some perspectives including transparency requirements for the provision of generative AI services are worthy of study.

The fourth chapter emphasized that statutory license can take into account the protection of the interests of copyright holders and the development of AI technology. It is a reasonable path for generative AI to obtain works to train models. It was recommended to set a statutory license for exploitation of copyrighted works by generative AI. It was also recommended that improving supporting measures to promote the implementation of the statutory license, including using the collective management system, establishing a market-oriented payment mechanism and establishing a mechanism to protect the rights and interests of copyright holders. It was recommended to clarify the reasonable obligations of generative AI developers and service providers，including increasing transparency of exploiting works, optimizing model design and strengthening management of users’ behavior.