字号·· | 护眼
日经

Anthropic在日本大规模购书引发出版商强烈反对Anthropic's mass book buying in Japan stirs publisher backlash

点「原文对照」整页切到原文,或双击某段只看那段的原文。

Anthropic已购买实体书籍,将其数字化以用作人工智能训练数据。© 路透社Anthropic has purchased physical books to digitize for AI training data. © Reuters

东京——日本出版商对Anthropic从一家分销商处大量采购书籍的行为感到愤怒。尽管此举很可能合法,但再次引发了关于作者作品被用于训练人工智能时,其权利应如何保护的讨论。

TOKYO -- Japanese publishers have been angered by Anthropic's bulk purchases of books from a distributor, which, while likely legal, have reignited a debate on how authors' rights should be protected when their work is used to train artificial intelligence. From December 2024 to December 2025, Nippan Group Holdings sold large quantities of books from publishers in the Japan Book Publishers Association to the American AI developer.

从2024年12月到2025年12月,Nippan Group Holdings向这家美国人工智能开发公司出售了日本出版商协会成员出版社的大量书籍。

This came to light in early September, when it was learned that U.S. court documents from a copyright infringement lawsuit by authors against Anthropic mentioned a sales agreement with Nippan.

此事于9月初曝光,当时人们发现,在作者起诉Anthropic的版权侵权诉讼中,美国法院文件提及了与Nippan签订的销售协议。

The association sent a letter to Nippan in mid-September seeking an explanation of the transactions and terms.

该协会于9月中旬致信Nippan,要求解释相关交易及条款。该分销商周一回应称,承认存在销售行为,但表示此前并不知晓这些文本将被用于人工智能训练。Nippan和Anthropic均未就具体交易发表评论。

The distributor replied Monday, acknowledging the sales while saying it had been unaware the texts would be used for AI training. Both Nippan and Anthropic declined to comment on individual transactions. Anthropic said buying physical books on which to train large language models is widespread in the AI industry, and that none of the company's data acquisition programs involve buying and destroying rare books.

Anthropic表示,在人工智能行业中,购买实体书籍用于训练大型语言模型的做法十分普遍,且公司任何数据获取计划都不涉及购买和销毁稀有书籍。

It also denied that large amounts of texts are destroyed after training, as some critics have said.

该公司还否认了部分批评者所称的“训练后会销毁大量文本”的说法。Nippan则表示,其交易符合法律规定。东京的二手书店也收到了数百甚至数千册书籍的订单。

Nippan said its transactions complied with the law. Used book stores in Tokyo have received orders for hundreds or even thousands of books. A similar situation has emerged with used books in Japan.

日本二手书市场也出现了类似情况。拥有约500家会员企业的东京二手书协会表示,7月份收到了有关批量采购的咨询,据称来自美国的二手书卖家。从7月到9月,一些书店接下了数百甚至数千册书籍的订单。

The Tokyo Used Book Association, which has about 500 member businesses, said it received inquiries in July about bulk orders, purportedly from U.S. sellers of used books. Between July and September, some shops took orders for hundreds or thousands of volumes. The rapid influx left stores scrambling to send out the books and update their inventories.

突如其来的大量订单让书店不得不加紧发货并更新库存。一家接到2000件订单的书店甚至暂时歇业,并增派员工,用数天时间完成发货。

A shop that received a 2,000-item order closed temporarily and brought in more staff to ship out the books over several days. The purchases spanned many genres in the humanities and social sciences, but included virtually no manga nor anime-related texts -- and no duplicate titles at all.

此次采购涵盖了人文与社会科学领域的多个类别,但几乎没有任何漫画或动漫相关的书籍,也完全没有重复的标题。发货地址均与日本西部冈山县的商家有关。

The shipping addresses were associated with businesses in western Japan's Okayama prefecture. The used book association also received a request for a bulk purchasing partner.

该二手书协会还收到一份批量采购合作伙伴的请求。请求方表示采购量没有上限,可能高达数十万册,且对书籍类别和内容没有任何限制。

The requester said there would be no upper limit on volume, potentially running into the hundreds of thousands of books, and gave no restrictions on genre or content. "It felt quite different from the [typical] buying we've seen before," a representative said.

“这与我们以往接触过的[典型]采购方式大相径庭,”一位代表表示。该组织尚不清楚这些书籍是否将被用于人工智能训练数据。

The organization does not know whether the books were to be used for AI training data. Using physical books for AI training "is legal under Japanese copyright law," said Tatsuhiro Ueno, a law professor at Waseda University specializing in intellectual property.

专攻知识产权的早稻田大学法学教授上野达弘表示,使用实体书籍进行人工智能训练“在日本的版权法下是合法的”。

If the training is conducted by an American company in the U.S., then American law applies.

如果训练由美国公司在美国境内进行,则适用美国法律。不过上野指出:“只要不复制书籍本身的内容,根据近期的判例,在美国也极有可能是合法的。”

Still, "as long as the content of the books themselves isn't reproduced, then it's likely to be legal in the U.S. as well, based on recent precedent," Ueno said. In the U.S. class-action lawsuit, Anthropic was found to have downloaded roughly 7 million pirated books to use as training data.

在美国的一起集体诉讼中,Anthropic被认定下载了约700万本盗版书籍用作训练数据,并同意支付一笔数额不小的和解金。

Anthropic agreed to pay a sizable settlement.

即便Nippan的销售行为合法,出版商——作为该分销商的主要股东之一——仍对事件处理方式感到不满。

Even if Nippan's sales were legal, publishers -- which are among the distributor's top shareholders -- were unhappy with how the matter was handled.

一家大型出版社表示:“他们在销售前没有与长期合作过的出版商沟通,这种行为显然违背了诚信原则。”

"The fact that they didn't talk to publishers they've done business with for years before the sales was bad faith," a major publishing house said.

与此同时,大学研究人员也担忧,如果这些书籍从市场上消失,他们可能将无法购买到宝贵的学术资料。

Meanwhile, university researchers have voiced concern that they could become unable to buy valuable books if they disappear from the market.

尚未数字化的旧书被视为人工智能行业重要的训练资源,而许多业内人士认为,互联网上可用的数据已开始枯竭。

Old books that have yet to be digitized are seen as an important training resource in the AI industry, where many believe usable data on the internet is starting to be exhausted.

Anthropic认为,本地语言信息对于将其AI模型适应新市场至关重要。据信,该公司正在使用大量日语数据(包括二手书),以构建能够流利使用日语且对文化、法律等领域有深入了解的AI。

Anthropic views local-language information as crucial for adapting its AI models to new markets. It is believed to be using a wide range of Japanese-language data, including used books, to build an AI that is fluent in Japanese and well-versed in fields such as culture and law.

“希望这能推动合作取得进展,从而建立建设性的框架,包括为作者创造更多价值,”拥有AI法规经验的律师冈田敦志在谈到近期案件时表示。

"Hopefully this will lead to progress on cooperation to create constructive frameworks, including to return more value to authors," said Atsushi Okada, a lawyer with experience in AI rules, about the recent cases.