As artificial intelligence spreads rapidly, lawsuits over the unauthorized use of copyrighted works to train AI models are piling up around the world. The sharper the race over AI performance becomes, the deeper the conflict grows between companies seeking high-quality data that is not available online and creators demanding fair compensation.
At least 15 overseas-published works by Korean writers including Han Kang, Hwang Sok-yong and Shin Kyung-sook are listed as eligible for compensation on Anthropic's copyright settlement website, according to reporting by The Seoul Economic Daily on the 15th.
The listings stem from Bartz v. Anthropic, widely seen as a landmark case in AI copyright law. In August 2024, three authors including Andrea Bartz sued Anthropic, alleging it used pirated digital books from illegal copy libraries such as Library Genesis to train its AI without permission.
Court proceedings laid bare the shortcuts Anthropic took to feed book data into its models. The company downloaded millions of books from illegal digital libraries around 2021 and 2022.
As the controversy grew, Anthropic shifted strategy and began buying large volumes of physical books from secondhand bookstores and book distributors worldwide in 2024. Internally it called the physical book purchasing effort the "Panama Project" and drew up plans to "destructively scan" every book in the world.
Documents Anthropic prepared in 2024 show it selected Woongjin Booksen, South Korea's largest book distributor, as its distribution partner in Korea and proceeded with a sales contract. One memo in the documents said books were being sourced mainly from the U.S. and the U.K. at that point, with orders from Taiwan, Japan and Korea planned soon.
Even so, South Korea has no rules governing the sale of books by distributors to AI companies. Under Article 20 of the Copyright Act, a copyright holder's right of distribution is exhausted once a book has been lawfully sold for the first time with the holder's consent.
Creators are pushing back. As AI advances, its output has reached a level of polish that can rival human work, and that poses a threat. They argue it is unfair for AI companies to endanger the creative ecosystem by training on copyrighted works without consent. They also say that using rare and antiquarian books for corporate-profit AI training and then discarding them damages the shared assets of humanity.
Courts around the world have reached conflicting conclusions. In July, the U.S. District Court for the Northern District of California, which heard Bartz v. Anthropic, ruled partly in the company's favor. The court drew a line between training on lawfully purchased books and downloading pirated digital copies. In the case of lawful purchases, it ruled the use was fair use and did not infringe copyright, because the content of the books is transformed through AI training. But it concluded that the pirated copies infringed copyright on the basis of the act of copying itself, regardless of their use in AI. Anthropic has since agreed to pay a total of $1.5 billion (about 2.2 trillion won) in settlement, covering only the roughly 500,000 illegally downloaded books. Han Kang and other Korean writers are on the list of those eligible to receive the money. Anthropic plans to begin paying rights holders of the covered books on Nov. 15.
By contrast, on Sept. 29 the U.S. Court of Appeals for the Third Circuit upheld the copyright claims of Thomson Reuters in its suit against Ross Intelligence, an AI-based legal startup, over the use of summaries of case law from Reuters' legal service to train AI.
Europe protects copyright more broadly. In July, the Munich Regional Court sided with GEMA, the German music copyright collecting society, in its suit against the music AI service Suno. The court found that the similarities in songs generated by the AI were "too extensive and too specific to be coincidental."
The European Union's AI Act took effect in August. By recognizing exemptions for training only for specific purposes set out in the law, such as research, criticism and news reporting, it protects copyright more strictly than the U.S. does. Works cannot be used for AI training if the rights holder, such as an author, has opted out. Germany applies the toughest standard on AI and copyright, strictly regulating the memorization of content inside AI models itself.
In the U.S., an AI powerhouse, copyright litigation is still unfolding. The New York Times is suing OpenAI and Microsoft, alleging they trained models on millions of its articles without permission. Judge Vince Chhabria of the U.S. District Court for the Northern District of California ruled in Meta's favor in a copyright suit against the company but noted that plaintiffs could prevail in other cases if they more fully demonstrated the harm AI inflicts on the market for existing works.