The Korea Publishers Society’s new appeal turns an abstract copyright argument into three demands that authors and readers can follow: disclosure, permission and payment.
Korean publishers are asking a basic question of artificial-intelligence companies: which books did you use?
On October 16, the Korea Publishers Society opened signatures on 대한민국 정부와 국회, AI 기업에 보내는 호소문 (An Appeal to the Government, National Assembly and AI Companies). Its English slogan is direct: “No Books, No AI.” The association says generative systems depend on the sentences of writers, the choices of translators and the editorial work that makes a manuscript publishable, while the people behind those books often cannot determine whether or how their work entered a training set.
The appeal does not call for a general ban on AI. It asks for three conditions: companies should disclose the sources and legal basis of training data; use should begin with rightsholder permission rather than presumed consent; and creators and publishers should receive fair payment when their publications are used.
Those demands matter because they convert a technical dispute into information a working author can act on.
Disclosure comes before enforcement
A copyright holder cannot challenge, license or price a use that remains invisible. The appeal therefore starts with records: which work was used, where it came from and what legal basis the company claims.
That information would not settle every copyright question. Training can involve different kinds of copies, datasets, jurisdictions and model-development processes. A title appearing in a collection does not by itself establish how a particular system used it or what a court would decide.
It would, however, change the practical position of authors and publishers. Today, many must infer use from leaked datasets, litigation records or model outputs. A verifiable record would move the first question from “Can we prove our book was there?” toward “Was this use licensed or otherwise lawful?”
For translated Korean literature, the chain can be especially complex. A single edition may involve the Korean author, a domestic publisher, a translator, an overseas publisher and territory-specific rights. Disclosure is the prerequisite for identifying which agreement governs which text.
Silence is not a license
The second demand rejects an opt-out assumption. The society argues that a creator’s failure to announce a refusal should not be treated as permission.
That distinction shifts the burden. Under an opt-out model, every writer and small publisher must discover each system, understand its process and register an objection. Under a permission-based model, the company seeking the material must identify the rightsholder and secure a lawful basis before use.
The difference is not merely procedural. Large publishers may have rights departments and searchable catalogs. An individual illustrator, translator or small press may not have the staff to monitor changing technical policies across companies and countries.
A permission rule would strengthen that weaker party, but it would also require workable licensing infrastructure. AI development uses material at a scale that cannot be negotiated book by book through ordinary email. Collective licenses, rights registries or standardized machine-readable terms may become necessary if consent is to be meaningful rather than symbolic.
Payment requires a method, not only a principle
The appeal’s third demand is fair compensation. The difficult question is how to calculate it.
A flat payment per title is simple but may ignore edition, language, length and intensity of use. A revenue-share model connects payment to commercial success but can be opaque and leave creators waiting for uncertain income. Collective licensing can reduce transaction costs while raising questions about distribution: which authors receive money, and on what evidence?
The association’s statement does not announce a rate or completed licensing system. It establishes a negotiating position. That is useful, but the next stage will require details strong enough to prevent smaller rightsholders from being offered nominal terms simply because they cannot bargain alone.
Readers have a stake in that design. Publishing income finances editing, translation, advances, illustration and the next book. Compensation is not only a retrospective reward for material already created. It affects whether the people who make books can continue doing so.
The argument is no longer hypothetical
The appeal arrived as Korean publishing confronted evidence tied to a specific U.S. settlement. Korean reporting said the Literature Translation Institute of Korea had identified 21 Korean writers’ books on the Anthropic settlement compensation list and notified the authors. A settlement list is not the same thing as a public inventory of every model’s training data, but it makes the scale problem concrete.
The Korea Publishers Society also linked its campaign to reports that overseas AI companies had sought large quantities of Korean books through a domestic distributor. The association’s release did not identify the buyers or distributor, so the claim should not be expanded beyond what it states.
What can be said is narrower: Korean rightsholders are no longer debating only what a future training market might look like. They are trying to determine where existing books have already traveled and whether current legal processes give creators a usable remedy.
A signature campaign is leverage, not law
As of the morning of October 16, the society reported signatures from 158 organizations and 321 individuals, including writers, translators, illustrators and publishing workers. The form remained open, so that count was an early snapshot.
The campaign does not create a statute, force a company to release records or guarantee payment. Its immediate power is collective. It gives the government and National Assembly a clearer account of what publishing groups want, and it makes it harder to describe the dispute as opposition to technology itself.
The appeal’s strongest argument is structural. Books are not raw material without a production history. They contain layered labor and layered rights. Any training market that treats those layers as too inconvenient to identify will favor companies with computing scale over creators with legal ownership.
Disclosure, permission and payment will not answer every question about generative AI. They would make the questions auditable. For Korean authors and publishers, that is the first practical step from suspicion toward bargaining power.
