欧洲独立书商发现了一笔离奇的大宗订单，涉及5000本来自不同领域的杂乱书籍，这引发了他们对AI公司是否在秘密采购书籍用于训练模型的怀疑。[1]美国已曝光AI企业Anthropic通过名为Project Panama的项目花费数千万美元购买并扫描数百万本书籍，最终与作者达成1.5亿美元的和解。[1]欧洲书商担忧这一做法可能违反当地版权法，并有可能加速传统出版业的衰退。[1]
德国出版商与书商协会将此类行为称为"具有破坏性且违法"。[1]该协会发言人Thomas Koch表示，扫描书籍将违反德国版权法。[1]柏林古籍书商Urban Zerfass指出，亚马逊在每笔交易中获利20%。[1]柏林拍卖行Bassenge的艺术历史买家Markus Brandis补充道，ISBN编号系统使机器人能够记录哪些书籍还未被扫描或缺失。[1]对此，Anthropic声称其Claude模型的训练数据来自公开网络数据、商业数据集和自生成数据。[1]
European independent booksellers have uncovered a suspicious pattern of large orders for thousands of obscure titles, fueling concerns that AI companies are systematically acquiring books through third-party vendors to scan for model training [1]. The discovery echoes a practice already exposed in the United States: Anthropic spent tens of millions of dollars purchasing and scanning millions of books through Project Panama, ultimately settling with authors for $1.5 billion [1].
Industry groups in Europe have raised alarm about the legal implications. The German Publishers and Booksellers Association contends that scanning books for AI training violates German copyright law and called the practice "destructive and illegal" [1]. Thomas Koch, a spokesperson for the association, stated that "scanning books will violate German copyright law" [1]. Booksellers in Berlin have also observed the mechanics of these operations, with one noting that Amazon profits 20 percent from each transaction [1]. Additionally, an art history buyer at a Berlin auction house suggested that ISBN systems enable automated tracking of books that remain unscanned or out of print [1].
Anthropic has maintained that Claude's training data derives from "public web data, commercial datasets, and self-generated data" [1], though this statement does not directly address the reported book acquisition activities. European booksellers worry that large-scale scanning operations could accelerate the decline of traditional book retailers and undermine copyright protections that have governed the industry since the ISBN system was introduced in 1966 [1].