Anthropic downloaded millions of books from digital piracy libraries in 2021 and 2022, and at least 15 works by Korean authors were among them, including Han Kang's "Human Acts" as well as books by Hwang Sok-yong and Shin Kyung-sook, Seoul Economic Daily reported. The disclosure adds Korean writers to a piracy dispute that has already reshaped how US courts treat AI training on books, and it lands weeks before Anthropic is due to start paying out a $1.5 billion settlement.
The downloads came to light during Bartz v. Anthropic, a lawsuit filed in August 2024 by author Andrea Bartz and others, according to Seoul Economic Daily. In a ruling in July, the U.S. District Court for the Northern District of California found that training AI on legally purchased books was a transformative use and therefore fair use. The same ruling held that the act of copying pirated editions itself constituted copyright infringement.
Anthropic agreed to pay $1.5 billion, about 2.2 trillion won, in a settlement covering roughly 500,000 books, Seoul Economic Daily reported. Korean authors are on the list of recipients, with payments set to begin on the 15th of next month.
After the controversy, Anthropic began buying and scanning physical books around the world in 2024 under what it calls Project Panama, according to the outlet. In Korea it signed a book purchase contract with Woongjin Booksen, which the company confirmed.
Under Article 20 of Korea's Copyright Act, books that have been legitimately sold may be resold. Critics cited by Seoul Economic Daily say there are no rules governing the use of such purchased books for AI training, and the outlet reported that Korean creators are pushing back against the practice.
A patchwork of rules
Court rulings on AI training have diverged by region, Seoul Economic Daily reported. In the United States, rulings have largely found no problem with training on legally purchased books. Europe takes a stricter approach, barring the use of works for AI training if the rights holder objects.
Korea ranks second in the world in paid ChatGPT subscribers, the outlet reported, a sign of how widely AI is used domestically even as its regulatory framework lags. An AI framework act took effect in January but contains no provisions protecting creative works, according to Seoul Economic Daily. The issue that has drawn the most attention in Korea so far has been the flood of AI-generated books, rather than the question of what material was used to train the underlying models.
The disclosure comes as Anthropic has been reporting rapid revenue growth while preparing an initial public offering. The company posted third-quarter revenue of $16.8 billion, up 45% from the previous quarter, after $4.7 billion in the first quarter and $11.6 billion in the second, according to Seoul Economic Daily's reporting on the company's financials. The Financial Times, in details relayed by Seoul Economic Daily, read the figures as a slowdown from earlier in the year, when revenue had been doubling.
For working authors, the settlement and the Woongjin Booksen contract point to two separate tracks running at once: compensation for past piracy, and an ongoing arrangement to supply AI developers with legitimately purchased print books for training. Seoul Economic Daily's reporting makes clear that Korean law has not caught up with the second track, leaving open how authors whose purchased books end up in AI training sets might be compensated going forward. Payments under the U.S. settlement are set to begin next month.
Sources
- Anthropic Pirated Han Kang's Books, Stoking AI Copyright Fears — Seoul Economic Daily,
Every quotation and figure in this article is checked against the source reporting linked above before publication. How we report