
索尼音乐与环球音乐(UMG)再次对 AI 音乐公司 Suno 提起诉讼。两大唱片公司指控 Suno 新发布的 v6 模型仍然侵犯版权,理由是该模型此前的模型输出数据训练而成,而那些旧模型的训练数据来自从 YouTube 等渠道抓取的未授权音乐。
诉状中,索尼与环球音乐使用了"模型洗白"(model laundering)的说法:在一个侵权模型之上训练"新"模型并不能消除侵权,只是把侵权价值层层传递——从被复制的录音传入被污染的模型,再从模型传入输出,最后传入 v6。"v6 不是一个全新的开始,它只是同一棵毒树结出的果实。"
v6 发布时,Suno 方面曾表示新模型"从零开始、使用全新数据集训练",其中包含"用户数据"。Suno 发言人随后确认,训练数据包括合作伙伴授权内容、社区用户的创作与偏好信号,以及团队积累的经验,但未说明是否包括用户上传音频或基于上传音频生成的输出。
索尼方面还指控 Suno 使用知识蒸馏方式训练 v6,让新模型复刻基于侵权数据训练的旧"教师"模型的结果。诉状称,即便一个模型没有直接使用原告的录音训练,它也受益于 Suno 保留的未授权副本。两大唱片公司认为,除非 Suno 真正从零重建训练管线,其模型将始终建立在被盗数据之上。
本文来源:https://www.theverge.com/ai-artificial-intelligence/1000758/suno-sony-umg-lawsuit-ai-music
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Sony Music and Universal Music Group (UMG) have again filed a lawsuit against AI music company Suno. The two major record labels accuse Suno's newly released v6 model of continuing to infringe copyrights, arguing that the model was trained on the output data of earlier models, whose training data came from unauthorized music scraped from channels such as YouTube.
In the complaint, Sony and UMG used the term "model laundering": training a "new" model on top of an infringing model does not eliminate the infringement—it merely passes the infringing value down through layers, from the copied recordings into the contaminated model, from the model into its outputs, and finally into v6. "v6 is not a fresh start; it is simply the fruit of the same poisonous tree."
When v6 was released, Suno stated that the new model was "trained from scratch using a brand-new dataset," which included "user data." A Suno spokesperson subsequently confirmed that the training data included partner-licensed content, community users' creations and preference signals, and the team's accumulated experience, but did not clarify whether it included user-uploaded audio or outputs generated from uploaded audio.
Sony also accused Suno of using knowledge distillation to train v6, having the new model replicate the results of the old "teacher" model that was trained on infringing data. The complaint argues that even if a model was not directly trained on the plaintiffs' recordings, it still benefits from the unauthorized copies Suno has retained. The two major labels believe that unless Suno truly rebuilds its training pipeline from scratch, its models will always be built on stolen data.