二手书销售近期出现异常激增，引发业界关注。独立书商观察到大量书籍被批量购买，随后运往远处仓库，其中一位经营者在从业三十年间"从未见过这样的情况"[1]。业界推测这一现象与AI公司训练大型语言模型的需求密切相关[1]。
Anthropic公司因大规模采购书籍而成为焦点。根据法庭文件，该公司内部通讯中称相关项目为"Project Panama"，目标是"破坏性扫描全世界所有的书籍"[1]。破坏性扫描是指将书籍运送到工业级数字化地点，包括移除书脊使所有页面可快速扫描，然后回收残留物[1]。2025年，美国法官William Alsup裁定Anthropic使用作者书籍是"极具转化性的"，因此不违反美国法律[1]。
对于其模型的训练数据来源，Anthropic发言人表示："Claude由公开网络数据、商业获取的数据集和我们自己生成的数据混合训练"[1]。值得注意的是，英国与美国版权法存在差异——英国的法律起点是所有复制行为都需要版权所有者的许可[1]。
Secondhand book sales have experienced an unprecedented spike, with independent booksellers reporting the bulk purchase and shipment of large quantities of books to distant warehouses [1]. In thirty years of operating in the secondhand book trade, one dealer stated he had "never seen anything like this" [1]. Industry observers have attributed the surge to AI companies acquiring books for training large language models.
The controversy intensified following a 2025 U.S. court ruling in which Judge William Alsup determined that Anthropic's use of authors' books for model training was "highly transformative" and therefore did not violate American copyright law [1]. Internal communications from Anthropic revealed the initiative was called "Project Panama," with the objective to conduct "destructive scanning of all books worldwide" [1]. Destructive scanning involves transporting books to industrial-grade digitization facilities, removing book spines to enable rapid page scanning, and then recycling the remnants [1]. An Anthropic spokesperson stated that "Claude is trained on a mixture of publicly available web data, commercially acquired datasets, and our own generated data" [1].
The practice has raised ethical questions about book destruction and copyright protection. British and American copyright law diverge significantly—in the United Kingdom, the starting point is that all copying activities require copyright holder permission, whereas U.S. law permits such use under certain circumstances [1].