Google is pursuing book purchase contracts with Korean publishers, following the controversy over AI company Anthropic's bulk acquisition of Korean books from a local book distributor. Industry officials question whether South Korea's small publishers can secure fair compensation from global giants. A particular concern is that Anthropic's $3,000-per-book settlement over pirated Korean literary works — which was damages for piracy, not payment for AI training — is being treated as a de facto price tag for AI training material, prompting calls for a fair pricing standard.
Google has been negotiating with major Korean publishers to buy books for AI training since the first half of this year, according to publishing industry sources on the 18th. Several industry officials said they understood that some contracts had already been completed. Which publishers signed and on what terms could not be confirmed, though academic publishers were reportedly among them. Google Korea did not respond to a request for comment.
After being caught training AI models on illegally downloaded books, global AI companies have shifted toward signing direct purchase contracts. The problem is that no legal basis exists to regulate distributors that sell content for AI training without the copyright holder's consent, raising concerns about infringement of the rights of authors and publishers. Google differs in approaching publishers rather than distributors, but the underlying issue is the same. If publishers have not obtained individual consent from authors, selling data for AI training carries a high risk of legal disputes.
Like Anthropic, Google has been drawn into copyright litigation over AI training abroad. In July, Hachette Book Group and others filed a class-action lawsuit against Google in the U.S. District Court for the Southern District of New York.
There is considerable concern over whether Korea's publishing industry, smaller than its overseas counterparts, can exercise any bargaining power against global big tech firms such as Google.
An official at a company working on AI copyright issues said many Korean publishers are relatively small and that the volume of titles held by any single publisher makes it difficult to negotiate on equal terms with large AI companies. The official added that there are also practical difficulties in meeting the file formats and quality standards AI companies demand, and a lack of benchmarks for judging whether proposed terms are reasonable.
The official expressed concern that there has been almost no discussion of appropriate compensation — that is, of price. Settlements that large overseas AI companies have agreed to pay in lawsuits should not be confused with copyright fees, the official stressed. "If each party negotiates without sufficient information or standards, a price accepted by a publisher in urgent need could harden into the market standard," the official said, advising that standards be established for how license prices should vary according to the scope of use. The official added that the industry also needs to prepare for a market that goes beyond one-off licensing for training, one in which AI systems reference and cite books in their answers and settle payments repeatedly based on usage. The division of royalties between publishers and authors could also become contentious. According to the U.S. pop culture magazine Vulture, copyright holders of the books covered by the settlement in Bartz v. Anthropic are to receive $3,000 per title, with most splitting the sum 50-50 between author and publisher. Some publishers, however, demanded 100% of the damages, and in some cases authors received nothing because copyrights were registered too late to meet Anthropic's eligibility requirements.
Concerns are also growing that the competitiveness of Korean AI models could fall further behind as global AI companies absorb the world's knowledge while bearing the litigation risk and enormous content acquisition costs. For Korean AI companies, which lag in scale and financial resources, the burden of buying content is heavy, leaving them unable to secure sufficient data and weakening the competitiveness of their models. An official at a Korean conglomerate developing AI models said the company is not training on Korean literary works because the data costs more to buy than general data. "For now, we are training our AI mainly on data obtained from the government," the official said.
An industry official who runs a publishing house said measures prepared at the industry level would be meaningless unless legislation passes. "Even if everything is taken from Korea by companies abroad, there is nothing we can do," the official said. "If this continues, it is obvious that the Korean content market will be damaged."
Multiple amendments to the Framework Act on Artificial Intelligence Development and Trust — which would require disclosure of the sources of generative AI training data and set out procedures for copyright holders to verify such data — have been introduced in the National Assembly but remain pending.