AI 학습에서 “변형적 이용”과 “잠재적 시장”의 의미에 대한 검토

The legal debate over fair use in the context of AI learning has gained significant momentum following recent U.S. court decisions and reports issued by the U.S. Copyright Office. However, the standards for determining its legality remain unsettled. In particular, the two core factors of fair use—transformative use and potential market effect—are being interpreted differently depending on the characteristics of AI technology, the use of data, and market conditions, which lowers legal predictability. This study examines the legal framework and policy implications of these two factors to propose a reasonable approach to fair use for AI learning. Regarding transformative use, AI does not use copyrighted works to reproduce or enjoy their expression, but to extract information and analyze linguistic patterns to develop new functions. This suggests that AI learning may be viewed as transformative. Nevertheless, further clarification is required as to whether non-expressive use should be regarded as a part of fair use analysis or as a preliminary step in determining copyright protection. It is also necessary to evaluate each AI system individually based on its purpose, structure, and output method, rather than relying on a simple division between generative and non-generative AI. In addition, technical guardrails should primarily be considered as tools for preventing infringing outputs at the generation stage, rather than restrictions on the learning process itself. The growing proposal to recognize “lawful access” as an additional requirement also needs careful review in light of the purpose of fair use. With respect to the potential market effect, the analysis should include not only whether the AI output replaces the existing market for copyrighted works, but also whether the emerging market for AI training data licenses falls within the scope of protection that copyright law intends to secure. While concerns exist that unauthorized use of training data could hinder market development, expanding the concept of potential market too broadly may restrict innovation and go beyond the original purpose of copyright law, especially considering the characteristics of AI users and the public benefits of AI-generated outputs. This study concludes that fair use analysis for AI learning should adopt a balanced and contextual approach, taking into account technological characteristics, social benefit, and the fundamental objectives of copyright law, rather than relying on a single fixed standard.