South Korea released three separate guideline documents on AI and copyright between June 2025 and February 2026. None of them are legally binding. None of them give you a clear answer on whether training an AI model on Korean content is lawful. And the only lawsuit testing these questions — filed by Korea’s three largest broadcasters against OpenAI — has no ruling yet.
If you are building or deploying AI systems that touch Korean copyrighted content, you are operating in a framework that is more ambiguous than Japan’s (which has a broad statutory exception) and less structured than the EU’s (which has an opt-out mechanism). Korea chose fair use — the most flexible tool, and the least predictable one.
Here is what the guidelines actually say, where the gaps are, and what you should do about it.
What Are the Three Documents?
Between June 2025 and February 2026, the Korean government issued three guideline documents through the AI-Copyright System Improvement Council, a public-private working group that has been active since 2023. All three are non-binding administrative guidance — important reference points for courts, but not law.
Document 1 (June 30, 2025): Copyright Registration Guide. Issued by the Ministry of Culture, Sports and Tourism (MCST) and the Korea Copyright Commission (KCC). Covers whether AI-generated content can be registered for copyright.
Document 2 (June 30, 2025): Dispute Prevention Guide. Same issuers, same date. Covers how to avoid copyright infringement when using AI outputs. An official English PDF is available on the KCC website. This guide explicitly excludes the training phase — it only addresses outputs.
Document 3 (February 26, 2026): Fair Use Guide for AI Training. Issued jointly by MCST, KCC, and the Ministry of Science and ICT (MSIT). This fills the gap Document 2 left open: whether using copyrighted works to train AI models qualifies as fair use under the Korean Copyright Act.
The three documents together cover the full AI copyright lifecycle — training, output generation, and registration — but they do it through guidance rather than legislation.
Can AI-Generated Content Be Copyrighted in Korea?
The Registration Guide (Document 1) draws a hard line between two categories.
Pure AI output cannot be copyrighted. If you type a prompt and the AI produces something with no human creative contribution, that output has no copyright protection in Korea. The KCC will not register it.
AI-assisted work can be copyrighted — but only the human-created portions. If a human makes a creative contribution using AI as a tool, the resulting work can be registered. The AI-generated portions within the same work remain unprotected.
The threshold question is what counts as “creative contribution.” The guidelines give examples in both directions.
Recognized as creative contribution:
- Inputting your own copyrighted material as a prompt, where that creativity is reflected in the output
- Modifying, augmenting, or eliminating aspects of AI output in a creatively significant way (inpainting, iterative refinement, combining AI outputs with original elements)
- Selecting, arranging, or reorganizing multiple AI outputs creatively (compilation authorship)
Not recognized:
- Simply typing a prompt — even a detailed one — unless it involves “significant creative decision-making and control over the output”
- Trivial edits: fixing typos, resizing, changing colors
For registration, applicants must clearly distinguish between AI-generated portions and human-created portions in the application. This connects to a separate obligation under the AI Basic Act — Article 31 requires labeling of generative AI outputs, which creates a paper trail that aligns with this registration requirement.
Can Copyrighted Works Be Used to Train AI Models?
This is the question the June 2025 guides deliberately deferred. The February 2026 Fair Use Guide (Document 3) addresses it — but provides no bright-line answer.
Korea does not have a Japan-style broad statutory exception for AI training. It does not have an EU-style TDM exception with an opt-out mechanism. Instead, it relies on Article 35-5 of the Korean Copyright Act — a general fair use provision modeled on the U.S. four-factor test.
The Fair Use Guide applies the four factors to AI training scenarios:
1. Purpose and character of the use. Is the training transformative? Is it commercial or non-commercial? Commercial use weighs against fair use but does not automatically disqualify it.
2. Nature of the copyrighted work. Factual and published works receive less protection than creative and unpublished works.
3. Amount and substantiality. How much of the original work was used relative to the whole?
4. Effect on the market. Does the AI’s use substitute for the original work?
The guide offers two worked examples. Academic papers — especially open-access ones — are likely to qualify as fair use for AI training. News articles used to train a system that then provides summary services without permission are likely not fair use. Beyond these examples, the guide states that courts will decide each case individually. Commercial web crawling is “not automatically excluded” from fair use.
The practical effect: every AI training use case involving Korean content must be assessed individually against these four factors. There is no safe harbor.
What Obligations Do AI Developers Have?
The guidelines create no binding legal obligations. But they signal what Korean courts and regulators will likely expect.
For AI businesses developing models on Korean content:
- Implement filtering and deduplication in training pipelines
- Secure licenses for training data where possible
- Clarify liability in terms of service
- Use legally cleared datasets where available
- Respond to rights holders who signal opposition through robots.txt or metadata
For AI users generating content:
- Avoid prompts that directly reference specific copyrighted works or characters
- Exercise caution with commercial use of AI outputs
- Accept responsibility for ensuring outputs do not infringe third-party rights
For rights holders concerned about AI training:
- Use technical measures — robots.txt, metadata embedding, explicit AI training refusals — to signal opposition
- Note that these are practical recommendations, not a statutory opt-out. Korea has no equivalent to the EU’s Article 4(3) CDSM opt-out mechanism.
The February 2026 announcement also included policy measures: MCST will build a rights information infrastructure to help developers identify rights holders. MSIT will link this to private data marketplaces and offer R&D tax credits for purchasing AI training data. The Korea Open Government License (KOGL) now includes a “Type 0” (unrestricted use) and an “AI Type” (permits AI training under specified conditions).
How Does Korea Compare to Japan and the EU?
The three jurisdictions represent three distinct approaches to the same question — can copyrighted works be used to train AI?
| Dimension | South Korea | Japan | EU |
|---|---|---|---|
| Legal basis for AI training | General fair use (KCA Art. 35-5) — case-by-case | Broad statutory exception (Copyright Act Art. 30-4) — permitted by default | TDM exception (CDSM Directive Art. 3-4) — permitted unless opted out |
| Opt-out for rights holders | No statutory mechanism; robots.txt recommended | No opt-out regime | Yes — Art. 4(3) explicit reservation right |
| AI-generated content copyrightable? | Conditional — human creative contribution required | Conditional — human creative contribution required | No specific rule; CJEU requires “author’s own intellectual creation” |
| Developer transparency obligations | Non-binding guidelines | Soft-law Principle Code (draft, Dec 2025) | Mandatory — AI Act Art. 53 (in force Aug 2025). Training data summary required. |
| Key enforcement action | KBS/MBC/SBS v. OpenAI (filed Feb 2026, no ruling) | Yomiuri/Nikkei/Asahi v. Perplexity AI (filed Aug 2025, no ruling) | GEMA v. OpenAI — Munich court found ChatGPT liable for memorizing lyrics (Nov 2025) |
| Legal certainty | Low — no AI-specific copyright legislation | Moderate — clear general rule, exceptions untested | Low-moderate — framework clear, application contested |
| Policy orientation | Cautious balance — fair use without bright-line exception | Pro-innovation — Art. 30-4 deliberately promotes AI development | Dual regime — broad TDM permission balanced by opt-out and mandatory transparency |
Japan is the most permissive. Article 30-4 provides a broad statutory license for AI training as long as the purpose is “information analysis” rather than reproducing protected expression. No opt-out exists. Developers face legal risk only if they deliberately overfit a model to reproduce specific works or use known pirated sources.
The EU sits in the middle. Training is permitted by default, but rights holders can opt out under Article 4(3) of the Copyright Directive. The EU AI Act adds mandatory transparency — GPAI providers must publish training data summaries and implement policies to respect opt-outs. The GEMA v. OpenAI ruling (November 2025) established that memorization of specific works is not covered by the TDM exception.
Korea chose the hardest path. No bright-line exception. No structured opt-out. Just a four-factor fair use test applied case by case, with non-binding guidelines offering worked examples but no guarantees. For developers, this means more ambiguity than either alternative. For rights holders, it means fewer tools to proactively protect their work. For courts, it means every AI training dispute will require a full fact-specific analysis.
What About the KBS/MBC/SBS Lawsuit?
On February 23, 2026, Korea’s three major terrestrial broadcasters — KBS, MBC, and SBS — filed a copyright infringement lawsuit against OpenAI at the Seoul Central District Court. They allege OpenAI trained ChatGPT on “decades of news content” without permission. The Korea Broadcasters Association framed it as “infringement of Korea’s data sovereignty.”
This is the first Korean broadcaster lawsuit against a global AI company, and the first significant test of AI training copyright liability in Korean courts. No ruling has been issued as of April 2026.
The case matters because it will likely force Korean courts to apply the Fair Use Guide’s four-factor framework to a concrete commercial AI training dispute — particularly the fourth factor (market effect), given that the broadcasters’ core claim is that their content was used without licensing.
Parallel cases are playing out globally. In Japan, three major newspapers sued Perplexity AI in August 2025. In the EU, the GEMA ruling against OpenAI and the pending CJEU Grand Chamber case (Like Company v. Google Ireland, hearing held March 2026) are shaping the doctrine. Korea’s case will likely be influenced by — and will in turn influence — these proceedings.
What Should You Do This Week?
If you operate AI systems that interact with Korean copyrighted content, here are specific steps:
- Audit your training data for Korean sources. Identify any Korean news content, published works, or licensed databases in your training pipeline. The KBS lawsuit signals that Korean media companies are willing to litigate.
- Document your fair use analysis. For each category of Korean training data, apply the four-factor test from the February 2026 guide. Academic papers and open-access content carry lower risk. News articles and creative works carry higher risk — especially if your model outputs summaries or reproductions.
- Implement technical signals. Respect robots.txt and metadata-based training refusals from Korean sites. The Dispute Prevention Guide recommends this, and compliance strengthens your fair use argument.
- Label AI-generated outputs. The AI Basic Act Article 31 requires labeling of generative AI outputs. This is separately enforceable under the AI Basic Act and connects to the copyright registration framework.
- Watch the KBS case and the CJEU ruling. Both are expected to produce outcomes that will shape the interpretation of fair use for AI training globally. Korea’s courts will likely reference international developments.
- If you are a rights holder: Consider filing explicit refusals through robots.txt, terms of service, and direct communication with AI companies. Korea has no statutory opt-out, but documented refusals strengthen infringement claims.
For a broader view of how South Korea’s AI regulatory framework works beyond copyright, see our article-by-article guide to the AI Basic Act and our analysis of how the PIPC is enforcing data protection rules against AI companies.
Disclaimer: This content is for informational purposes only and does not constitute legal advice.
Last verified: April 10, 2026
Compare: EU vs South Korea
For the global keystone comparison across twelve dimensions — high-impact vs high-risk classification, mandatory vs voluntary conformity, KRW 30M vs €35M penalties, Korea’s innovation chapter, and a five-step dual-market compliance baseline — see EU vs South Korea AI Act: High-Impact vs High-Risk Compared (2026).