대규모 언어모델 훈련 목적의 저작물 이용과 공정이용 문제 — 앤트로픽 사건과 메타 사건을 중심으로 —

초록

Does the unauthorized use of copyrighted works for the purpose of training large language models (LLMs) constitute copyright infringement, or should it be assessed as fair use? Whether in jurisdictions that have adopted specific exemptions for text and data mining (TDM), or in jurisdictions such as the United States and South Korea that must address this issue primarily through a general fair use provision, identifying a clear solution remains difficult. Recently, the U.S. District Court for the Northern District of California issued summary judgments in copyright (reproduction right) infringement actions brought by authors whose books were used to train Anthropic’s Claude and Meta’s Llama, respectively, holding that the use of the plaintiffs’ books for LLM training purposes constituted fair use. These two summary judgments are significant as the first decisions by U.S. courts to address the relationship between LLM training and fair use, and, in terms of their outcomes, they may be understood as aligning with prior U.S. precedents that have taken a relatively expansive view of fair use in TDM-related cases. Nevertheless, these decisions remain limited to the trial court level, and rather than articulating principles applicable to all LLMs, their reasoning focuses on specific factual circumstances and on particular models—Claude and Llama—through case-by-case analysis. Accordingly, their broader impact should not be overstated. More importantly, in assessing fair use, the courts did not recognize a potential licensing market for copyrighted works used for LLM training as a relevant market, a point that warrants further reconsideration.

키워드