AI 학습을 위한 저작물 이용은 공정이용인가 - Bartz, Kadrey, Ross 판결을 중심으로 -

This study analyzes whether the use of copyrighted works for AI training constitutes fair use under U.S. copyright law, focusing on issues addressed in recent U.S. district court decisions. Prior to the analysis, it is necessary to examine whether the output stage should be considered in the fair use assessment of training stage use. For general-purpose generative AI systems, the two stages must be assessed separately, whereas for special-purpose AI systems, the output stage may be considered together with the training stage. Regarding the first factor (purpose and character of use), AI training qualifies as transformative use, being a non-expressive use for extracting statistical patterns, while commerciality has limited significance. Although bad faith in using pirated copies is not a decisive factor under U.S. law, wrongful conduct such as data theft may be reflected in the fair use determination. The second factor (nature of the copyrighted work) plays a limited role in fair use analysis. For the third factor (amount and substantiality used), the use of entire works is reasonable in light of the transformative purpose, and does not weigh against fair use because the portion of learned content that can be reproduced externally is extremely limited. The fourth factor (market impact), which is the most important, is examined from three perspectives. First, the loss of compensation in the training data market amounts to circular reasoning in the absence of an actual market. Second, direct demand substitution at the output stage of general-purpose AI need not be considered in fair use analysis and does not occur in practice, as AI models learn only statistical patterns and commercial systems have filtering measures that make it difficult to generate outputs identical to original works. Third, indirect substitution through market dilution is merely demand "reduction" due to competition, not demand "substitution." Recognizing this would grant copyright holders a monopoly over entire genre markets, whereas copyright law permits competition and does not regard market reduction from transformative and non-expressive use as cognizable harm. Additionally, the public interest in AI technology's innovation and promotion of creativity must be considered. Considering these factors comprehensively, the use of copyrighted works for training general-purpose generative AI systems constitutes fair use.