생성형 AI 학습에 대한 공정이용 첫 판결 -앤쓰로픽 및 메타 케이스의 비교 분석-

This paper compares and analyzes Bartz v. Anthropic PBC and Kadrey v. Meta Platforms, Inc., in which U.S. district courts held, for the first time, that training generative AI models constitutes fair use. Both cases focused on two key fair use factors:transformative nature of the use (first factor) and the effect on the market for the works (fourth factor). In the Anthropic case, the court analyzed training and data collection separately and held that, while AI training itself was transformative, data collection through piracy was also transformative. In contrast, the Meta case analyzed downloading and training together, holding that downloading from pirate sites could also fall within transformative use. This led to divergent judicial views regarding the use of infringing data. Both cases recognized that AI training is analogous to human learning and contributes to scientific progress. However, their market-impact analyses differed:the Anthropic court found that pirated copies substituted for book sales and harmed the market, whereas the Meta court acknowledged the potential for market dilution but rejected claims of market harm due to insufficient evidence from the plaintiffs. Both courts concluded that the “training data licensing market” constitutes a theoretical market not subject to copyright holders’ exclusive control, and thus could not be considered a potential market under the fourth factor. In conclusion, while both decisions affirmed the scope of fair use for AI training, they diverged in their legal interpretation of the use of infringing copies. These cases are likely to serve as important precedents in future AI copyright litigation, both in the United States and in other jurisdictions including Korea.