On 16 October, Korean publishers launched a "No Books, No AI" national appeal. According to the Kyunghyang Shinmun, 158 organisations and 321 individuals had joined by the 16th. The appeal demands disclosure of the works used to train AI, prior permission, and fair compensation. It also says a rights holder's silence must not count as consent. Two days later, Seoul Economic Daily reported that Google has been negotiating to buy Korean books for AI training since the first half of 2026, with some contracts reportedly done. The same report said Anthropic's roughly US$3,000-per-title Bartz settlement is being treated as a de facto price.
The publishers' case deserves a fair hearing. A rights holder who is never told that a work was used cannot bargain, object or even learn that a licence is needed. Negotiating one-to-one with a company the size of Google is a lopsided exercise, so "silence is not consent" is a reasonable demand. The report also says that nothing in Korean law reaches distributors who sell content for AI training without rights holders' consent. That is a real gap, and it is hard to defend.
But the remedy the appeal proposes, a general rule of prior permission, would go further than the problem requires.
What Korean law actually says today
The Framework Act on the Development of Artificial Intelligence and Establishment of Trust (Law No. 20676, enacted 21 January 2025) took effect on 22 January 2026. Read in the CSET translation, its transparency provision does not do what publishers want. Article 31 requires operators to tell users in advance that a product uses high-impact or generative AI, and to label AI-generated content that is hard to tell from authentic material. It says nothing about disclosing training sources.
The nearest provision is Article 34(1)(2). It covers only operators of "high-impact" AI. They must, "to the extent that it is technically feasible", plan to explain AI outputs, including "an overview of the learning data used". That is an explanation duty owed for safety reasons, not a rights-clearance register. Article 15, meanwhile, directs the Minister of Science and ICT to promote the production, distribution and use of learning data. The statute's own bias is toward supplying training data, not gating it.
The amendments that would require disclosure of training-data sources are, as reported, stalled in the National Assembly. Their absence is why the appeal exists.
The $3,000 figure is a poor price signal
Treating the Bartz settlement as a market price misreads what it settled. In his 23 June 2025 order on fair use, Judge William Alsup granted Anthropic summary judgment that "the training use was a fair use". He also held that the pirated copies used to build a central library "were not justified by a fair use" and ordered a trial on those copies, including statutory damages. The case ended in a settlement, which the court finally approved on 20 July 2026. The Authors Guild describes the award as approximately $3,000 per work, four times the $750 statutory minimum for ordinary infringement.
So the figure prices exposure for acquiring works from pirate sites. It does not price a licence to train on lawfully obtained books. It was set under US law, in a class action, with statutory damages looming. Importing it into Korean negotiations as a benchmark would hand publishers a number that has no logical connection to what a licensed training corpus is worth. It would also give them a false sense of what the law currently entitles them to.
Why a blanket permission rule would backfire
If every act of training required prior permission, only firms that can afford large-scale clearance could build models. The EFF's October 2026 analysis quotes Lawrence Lessig on exactly this point: "If all training requires a license, then none but a handful of tech giants will be able to afford the costs of training." Korean AI start-ups and university labs would face that burden. Google, which is already negotiating, would not.
The Google talks are, in fact, evidence that a licensing market is forming without a statutory mandate. A market with willing buyers is the condition under which collective licensing, standard contract terms and sensible pricing can emerge. Regulation should support that market rather than pre-empt it.
A narrower package
Three measures would address the publishers' real grievances at far lower cost to innovation.
- Summary-level provenance disclosure. Require developers of large generative models offered in Korea to publish a summary of the main categories and sources of training data, and a contact point for rights holders. This builds on the "overview of the learning data" concept already in Article 34, extended to the largest systems. It is not a work-by-work register, which would be costly to produce and of little use.
- Close the distributor gap directly. If a distributor sells books for AI training without the rights holder's consent, the fix is a provenance warranty and clear liability for the seller. That targets the party that actually creates the problem. Pre-clearing every downstream use of training would not.
- Make collective licensing easy. Publishers' bodies and authors' organisations can offer blanket licences, with the Ministry of Culture providing model terms. Opt-out and opt-in design can then be tested with real transactions before anything is hard-wired into law.
None of this settles whether training on lawfully acquired books should require a licence under Korean copyright law. That question belongs to courts and, if necessary, to a deliberate amendment of the Copyright Act after evidence has been gathered. Alsup's order shows how much turns on how the works were obtained and how they were used. A single legislative rule written in a hurry, in response to an appeal and a newspaper report, will not capture those distinctions.
The publishers are right that silence should not be treated as consent. The proportionate response is transparency that lets them see what happened, and a functioning licensing market that lets them be paid for it. Prior permission for all training would protect Korean publishers less than it would entrench the largest foreign platforms.