Korea’s AI Copyright Fight Moves Into the Legislative Arena

On 5 May, 32 South Korean copyright and creator organisations issued a joint statement opposing the government’s “use first, settle later” approach under the Korea AI Action Plan. The proposal is framed as a way to accelerate AI development by widening room for fair use, but creator groups argue that it would weaken their ability to control how their works are used.

The dispute is not merely about compensation rates. It goes to the starting point of copyright bargaining: whether AI developers should obtain permission before using protected works, or whether mass use can be normalised first and priced afterwards. That question is likely to shape the next phase of South Korea’s copyright reform debate.

The proposal shifts leverage before a licence is negotiated

For AI companies, access to training data affects development speed, model quality and cost. A policy that lowers uncertainty around data use can therefore look attractive from an industrial strategy perspective. The difficulty is that “use first” changes the bargaining position before the parties even sit down.

Under a conventional licensing model, the rights holder retains a prior choice: whether to license, to whom, for which uses and under what conditions. A post-use settlement model keeps the language of compensation, but it moves the decisive moment to a later stage. Once works have entered training pipelines, creators may struggle to assess scale of use, number of model iterations, commercial relevance or downstream value. Without strong transparency obligations, settlement can easily become pricing led by the user rather than negotiation between equal parties.

Fair use cannot carry every industrial policy goal

The sensitive point in South Korea is the possible expansion of fair use. AI training is a new technical setting, and copyright law does need clearer answers for text and data mining, publicly available online materials and works whose ownership is hard to trace. But fair use is an exception. It is not designed to replace functioning licensing markets.

Creator groups are not rejecting every form of data analysis, nor are they denying the strategic importance of AI. Their sharper concern is that a broad exception for commercial model training could undercut licensing markets that already exist. Music, publishing, audiovisual content, journalism, webtoons and gaming assets are not just raw material for computation; many of them are traded through mature rights channels. If those markets are bypassed too easily, the price signal for original content is weakened.

Creators are defending control, not only payment

The reaction also reflects the economic weight of Korea’s content industries. K-pop, drama, film, webtoons, games and news content have become valuable cultural exports. The fact that high-quality Korean-language material is useful for AI training confirms its value. A system that turns this value into low-cost training input first and negotiates compensation later will face resistance.

The problem is sharper for individual creators and smaller rights holders. Large organisations may be able to join policy consultations, build licensing pools or operate collective management systems. A freelance writer, illustrator or composer may not even know whether a work has been scraped, let alone prove the scale of use. If the system leaves discovery, opt-out and recovery burdens mainly on creators, formal rights may offer limited practical protection.

Copyright reform may need a more granular design

If South Korea proceeds with legislative reform, a single rule is unlikely to satisfy either side. A more workable design may distinguish between content with clear ownership and mature licensing markets, materials with uncertain rights status, non-commercial research, public-interest uses and commercial model development. Different categories should not carry the same transparency and compensation requirements.

The details will matter. Lawmakers would need to decide which works can fall within any exception, who verifies rights status, whether training datasets must be documented or disclosed, how rights holders can refuse use, and whether compensation is set by market negotiation, regulation or collective management. A broad provision may look efficient at first, but it could simply move the dispute into courts.

Practical steps for AI and content businesses

Companies developing models for the Korean market, or using Korean-language content in training, should not treat the policy debate as a future safe harbour. Until the law is settled, safer practice includes preserving source records, separating licensed data from publicly collected data, assessing whether datasets include works from mature copyright markets, and building procedures for rights-holder queries and opt-outs.

Content businesses should prepare as well. Curated licensing datasets, collective licensing mechanisms, content fingerprinting, AI-training clauses and revenue-sharing models will all matter in the next round of negotiations. The durable solution is unlikely to be a simple enlargement of fair use. It will be a framework in which AI developers can obtain lawful, stable and auditable data, and creators can understand how their works are used and how they are paid.