Anthropic's Korean book leak: Who knew what and when
While Western outlets only recently uncovered that Anthropic scraped Korean literature for AI training, a Korean translation institute had already been notifying affected authors since February. The delay raises questions about corporate transparency and who benefits from slow accountability.
The leak that arrived late — and then arrived differently
Anthropic downloaded roughly 4 million books from pirate websites between 2020 and 2021 to fuel its AI models. Among them were English translations of works by some of Korea’s most celebrated authors. A US court called it illegal. Anthropic agreed to pay $3,000 per title for nearly half a million of those downloads last November.
The story hit global headlines recently. But here is what most English-language coverage has missed: the Korean Literary Translation Institute (KLTI) had already begun notifying affected Korean authors about their right to compensation in February — months before the story became international news.
The timing matters. It reveals a pattern in which non-English creators are treated as afterthoughts in the conversation about AI and intellectual property, even when their government agencies are tracking the same violations as Western regulators.
What Anthropic took
According to the KLTI, 20 Korean authors and 21 works were identified as having been illegally downloaded. The list includes Nobel Prize winner Han Kang, alongside Kim Ae-ran, Kim Young-ha, Shin Kyung-sook, Jung Yoo-jung, and Hwang Sok-yong. Each of these writers confirmed to the institute that they would pursue compensation — either directly or through their literary agents.
But the KLTI’s list is incomplete by design. The institute told the JoongAng Ilbo that its outreach focused on authors whose works it had actively supported in translation. That means the real number of affected Korean writers is likely higher.
A separate investigation by the JoongAng Ilbo into Anthropic’s compensation portal found additional names not included in the KLTI’s initial notification: Kim Eun-soo, author of “The Designers”; economist Ha-Joon Chang; philosopher Byung-Chul Han; and deceased writers Cho Se-heui and Hwang Sun-won.
And those are only the Korean-language or Korean-authored works. The full 4 million books Anthropic scraped include titles from dozens of languages and countries. The Korean cases are visible because a government-linked institute tracked them. In other markets, no such alert existed.
The compensation portal itself operates under US law, with forms written in English and deadlines set by American courts. Authors outside that system must navigate a process that was never designed with them in mind. Several Korean writers told the KLTI they had no idea where to begin filing a claim — not because the right existed, but because the infrastructure to exercise it was invisible to them.
Why February matters
The KLTI received its information from South Korea’s Ministry of Culture, Sports and Tourism, which had been briefed on Anthropic’s legal settlements. The institute sent a formal notice to its roster of authors in February, explaining how to file compensation claims and noting a March 30 deadline.
Yet the broader public story — the one that shaped global understanding of the case — broke later. Major outlets reported on Anthropic’s Korean book downloads as if they were newly uncovered. The narrative framed the incident as a recent scandal, not a violation that had been known to Korean institutions for months.
This gap is not unique to Anthropic. It is structural. When a tech company faces scrutiny in English-language media, the companies’ legal and communications teams respond primarily to Western jurisdiction. Settlements are negotiated in US courts. Press statements are drafted in California.
Non-English rights holders often learn about violations through their own governments — or not at all. The KLTI’s February notice was an act of institutional diligence. Its silence in the global press was not negligence on the institute’s part; it was a reflection of where the conversation was deemed to matter.
There is also a secondary effect that rarely makes headlines: authors who heard about the leak through English-language coverage felt a delayed sense of betrayal. Several told the KLTI they had been reading about their own work in the news before they had received any formal communication from Anthropic or its legal representatives. The injury compounded — first the unauthorized use, then the erasure from the narrative about that use.
Who wins, who loses
Anthropic pays $3,000 per title. For 482,460 works, that totals roughly $1.45 billion — a large sum, but a fraction of what the company has raised. The payments are a legal resolution, not a precedent-setting admission of systemic harm. The models continue to run on data that was obtained illegally.
Korean authors who received the KLTI’s notice have a path to compensation. Those outside the institute’s referral network — translators, smaller publishers, authors whose works were scraped without their knowledge — may still be unaware of their rights. The KLTI itself acknowledged this, urging authors to spread the word widely.
The bigger loss is informational. When the global narrative around AI copyright begins in English, it privileges English-speaking perspectives on what constitutes harm, what deserves punishment, and what counts as recovery. Korean authors exist in this framework as casualties, not as stakeholders in the policy debate that follows.
The financial stakes for individual authors are modest. $3,000 per title does not approach the earnings a published work might generate over a decade. But the symbolic weight is heavier: it establishes that Korean literary property can be taken, digitized, and monetized by foreign systems with only a per-unit fine as consequence.
Legal experts in Seoul have noted that the settlement structure — a flat payment per title rather than a royalty-based model — sets a troubling precedent. It treats copyrighted works as items with a fixed breach cost rather than as ongoing revenue streams whose value is eroded by unauthorized reproduction. That framing benefits aggregators and penalizes creators whose works generate value across years and markets.
What happens next
The March 30 deadline for filing compensation claims has passed. Authors who missed it may still have legal recourse, but the window for the simplest path is closed.
For the Korean literary community, the immediate question is whether the KLTI and the culture ministry will press for a broader disclosure — a complete list of all Korean works found in Anthropic’s dataset, not just the ones the institute happened to track. The JoongAng Ilbo’s investigation suggests such a list would be significantly longer than what has been publicly acknowledged.
For the wider industry, the episode reinforces a pattern: AI companies scrape copyrighted material at scale, face modest fines, and continue operating. The authors whose work powered their models rarely shape the rules that govern that use.
Lawyers affiliated with Korean writers’ associations are now exploring whether class-action mechanisms could offer a more efficient route to compensation than individual filings. The KLTI has indicated it may publish a multilingual guide to the claims process, though no timeline has been set.
The most consequential outcome may not be financial but institutional. The fact that a Korean government body tracked these violations independently of Western media coverage suggests that future incidents may be met with faster domestic response — even if the global narrative continues to lag behind.
The KLTI’s early notice was a reminder that accountability is possible when institutions pay attention. The problem is that attention rarely travels beyond the borders where it is first registered.