AI版权关键进展：美国连判两案，大模型“偷书”不算偷

Group 1 - The core issue revolves around whether using human works to train AI without authorization constitutes copyright infringement, with recent U.S. court rulings providing new references for this ongoing debate [1][2] - The rulings from the U.S. Northern District Court of California found that both Anthropic and Meta's use of copyrighted works for AI training fell under the "fair use" doctrine, emphasizing that the purpose of use was transformative and did not directly replace the original works [2][3] - The court highlighted that the determination of "fair use" is nuanced and depends on the legality of the data acquisition methods, with a distinction made between legal and illegal sources [4][5] Group 2 - In the Meta case, the court noted that the AI training was for a highly transformative purpose, as it was not intended for reading or disseminating the original works, but rather for generating tasks like writing code or emails [2][3] - The court also emphasized the importance of market impact, stating that if AI outputs could harm the market for original works, it might not qualify as fair use, although this was not proven in the Meta case [7][8] - The Anthropic case similarly recognized the transformative nature of AI training but differentiated between legal and illegal data sources, ruling that using data from illegal sources like "shadow libraries" constituted infringement [6][7] Group 3 - The rulings indicate a cautious approach by the courts, as they do not grant AI companies a blanket permission to use copyrighted works for training, stressing that each case must be evaluated on its own merits [3][6] - The distinction between the two cases lies in the treatment of data sources, with Meta's use of "shadow libraries" being viewed more leniently due to its failed attempts to obtain licenses, while Anthropic's establishment of a permanent internal library from illegally sourced materials was deemed infringing [5][7] - The ongoing legal disputes extend beyond literature, with similar copyright issues emerging in the film and visual arts sectors, indicating a broader industry concern regarding AI training practices [8]