隨著生成式人工智慧愈發融入我們的生活當中，其訓練過程中所涉及的著作權問題，勢必將成為未來發展該技術所不可迴避的重要法律課題。關於生成式人工智慧訓練階段是否構成著作權侵權，實務上涉及對技術運作機制的理解，並須將該等技術行為具體對應至著作權侵權之構成要件。除構成要件的認定外，亦須進一步審酌是否存在抗辯事由，得以排除或限制侵權成立之可能。 鑑於目前全球主要主導生成式人工智慧技術發展之企業多為美國公司，且從既有訴訟案件觀察，美國司法實務明顯居於主要討論地位，故本文將以美國法院之判決見解作為分析主軸，輔以歐盟數位單一市場著作權指令等重要規範，討論生成式人工智慧訓練階段是否構成著作權侵權。在侵權構成分析之外，由於當前多數生成式人工智慧開發者普遍以合理使用作為抗辯依據，本文亦將進一步探討其主張是否能主張合理使用。最後，倘若生成式人工智慧訓練行為不能主張合理使用，本文將提出具體之授權制度設計建議，以供未來建立合法資料授權架構之參考。 儘管本文主要以美國法與美國司法實務為分析對象，惟本文所涉及之著作權概念，包括重製、暫時性重製與合理使用等，皆為各國著作權法普遍存在之基本概念，僅於適用細節上有所差異。因此，本文所提出之理論架構與分析方法，亦具一定程度之跨國參考價值，能提供理論依據與比較法之參考。 ;With the rise of generative artificial intelligence (AI), the copyright issues arising from its training process are poised to become a significant legal concern. Whether the training of generative AI constitutes copyright infringement must be assessed in light of the specific operational mechanisms of such systems. In addition to determining whether infringement has occurred, it is also necessary to examine whether any defenses may apply to preclude or limit liability. Given that most of the leading companies developing generative AI technologies are based in the United States, and that the majority of litigation on this issue has arisen in U.S. courts, this article focuses primarily on U.S. judicial decisions as the foundation for analysis. To provide a more comprehensive view, it also draws upon relevant legal instruments, such as the European Union’s Directive on Copyright in the Digital Single Market, to explore whether the training of generative AI amounts to copyright infringement. Beyond the question of infringement, since most developers of generative AI currently rely on the doctrine of fair use as a defense, this article further examines whether such a defense can withstand judicial scrutiny. Should the fair use defense ultimately fail, this article will propose a licensing framework designed to support the lawful use of copyrighted data in the training of generative AI systems.
