South Korea: If the AI product competes directly with the original market (or acts like a drop-in replacement), fair use becomes much harder to defend.

“Fair Use with Teeth”: Korea’s Playbook for AI Training on Copyrighted Works

by ChatGPT-5.2

Korea has just done something that many governments have talked about for two years but rarely execute well: it published a practical, case-driven guide explaining when training generative AI on copyrighted works might be “fair use”—and, equally important, when it clearly won’t be.

The document—issued by the Ministry of Culture, Sports and Tourism together with the Korea Copyright Commission in February 2026—is framed as an “안내서”(guidance), not a binding legal interpretation. But its real function is unmistakable: it’s a regulatory signal and a compliance roadmap for AI developers, rights owners, courts, and investors trying to price risk in a messy, fast-moving area.

What the Korean guidance is actually about

At its core, the guide tackles a basic but explosive question:

Korea’s guide does three crucial things.

First, it clarifies the “mechanics of infringement” in AI training. It doesn’t pretend training is magically non-copying. It explicitly describes how training involves collecting, storing, preprocessing, training, and evaluating—and how these steps can entail reproduction and transformation of works as part of the pipeline.

Second, it makes fair use the central balancing mechanism. Korea has a general fair use provision (Copyright Act Article 35-5), similar in spirit to the U.S. approach, and the guide explains how fair use is assessed case-by-case rather than through a single bright-line AI exception.

Third, it “imports” global AI-copyright battle lessons into a single decision framework. It draws from recent disputes and rulings (U.S. and Europe) and translates them into actionable risk factors, including when courts see training as “transformative,” and when they treat it as market-substituting competition.

The guide’s big message: “Transformative use” is necessary—but not sufficient

One of the loudest themes (also highlighted in the press coverage) is that the first fair-use factor—purpose and character of the use—turns heavily on ‘transformative use.’ The guide’s logic is broadly:

• If training extracts patterns and doesn’t merely repackage expressive content as a substitute, it may be transformative.

• But if the AI product competes directly with the original market (or acts like a drop-in replacement), fair use becomes much harder to defend.

The guide (and the related article discussing it) points to the familiar contrast:

• Courts may be more open to training that is framed as “learning like humans learn,” i.e., extracting non-expressive patterns.

• But courts come down hard when the service is effectively built to replace a paid product using the same materials—like the Ross Intelligence v. Thomson Reuters/Westlaw scenario, where the use was deemed non-transformative and market-substituting.

This matters because it signals a regulatory posture that is neither “AI gets a free pass” nor “all training is infringement.” Instead: fair use lives or dies on provable product design choices and market impact.

The “teeth”: Korea draws a sharp line on unlawful access and evasion tactics

Where Korea’s approach becomes especially relevant for compliance teams is its emphasis on how the data was obtained—and what the developer did (or didn’t do) to respect restrictions.

The guide indicates that fair-use analysis can turn against a developer if:

• the content was accessed in ways that bypass access controls (logins, paywalls), or

• the developer circumvented technical protection measures (which can also violate separate anti-circumvention rules), or

• the developer ignored robots exclusion signals or comparable measures aimed at preventing crawling/scraping, making the overall “fairness” posture worse.

This is a big deal because it effectively collapses a common industry excuse—“it was publicly accessible”—into a harder question: was it accessed and collected in a way that a court will view as legitimate, responsible, and consistent with restrictions?

For rights owners, that means practical levers matter: access controls, rate limits, anti-scraping measures, and machine-readable restrictions can shape later legal arguments about “fairness,” not just technical outcomes.

Why this is important for AI developers

For AI developers, Korea’s guide is not merely “copyright commentary.” It is a risk architecture—a way regulators and courts can separate:

• defensible training (licensed, responsibly collected, product not designed as a market substitute), from

• high-liability training (questionable sourcing, ignored restrictions, outputs that reproduce or compete with the original).

It also implicitly encourages developers to build a compliance narrative that can survive discovery:

• documented data provenance

• documented respect for access restrictions

• model and product controls that reduce output infringement risks

• clear separation between training use and downstream substitutive services

And it reinforces a trend the guide itself records: licensing is increasingly normal, not exceptional—across text/news, images, music, and more.

Why this is important for content and rights owners

For rights owners, the guide’s value is twofold.

1) It validates the “market harm” frame. Rights owners often struggle to translate “it feels unfair” into legally legible harm. Korea’s guide centers the market/value impact factor and repeatedly flags that fair use weakens when the AI use erodes existing or reasonably anticipated markets—especially where the AI service resembles the original product or function.

2) It strengthens the argument that “how it was collected” matters. If the guide’s framing takes hold, rights owners can press on questions like:

• Did the developer bypass restrictions?

• Did they ignore robots.txt signals?

• Did they rely on known-infringing sources?

• Did they maintain controls to reduce infringement outputs?

Those questions are operationally provable—and they shift disputes away from vague philosophy into auditability.

Should other regulators follow Korea’s example?

Yes—but with a crucial caveat: they should follow the method, not necessarily the legal theory.

Korea’s most exportable move is issuing detailed guidance anchored in real cases and real technical workflows. Many jurisdictions are stuck in abstract debates about “innovation vs. creators,” while the market is racing ahead. Korea is trying to give stakeholders a shared map.

That said, other regulators should consider improving on Korea’s approach in three ways:

1) Don’t rely on “case-by-case” alone. Fair use guidance is helpful, but developers and rights owners also need predictability. Regulators should pair guidance with:

• clearer baseline obligations for provenance and transparency, and/or

• sectoral licensing pathways (collective licensing, standard terms, registries).

2) Build opt-out / rights-reservation interoperability where relevant. Korea’s fair-use model differs from EU-style TDM frameworks that contemplate rights reservation mechanisms (e.g., machine-readable signals). The Korean guide recognizes the relevance of restrictions like robots exclusion in the fairness analysis, but other regulators—especially those designing TDM regimes—should aim for globally interoperable signaling so compliance doesn’t fragment by region.

3) Treat “market substitution” as a product governance issue, not just a courtroom argument. The most dangerous zone is not “training in the abstract,” but training + product design that replaces protected markets (search, summaries, databases, reference tools, educational content). Regulators should push for concrete safeguards and accountability around output behavior—especially for commercial services that functionally displace original works.

Bottom line

Korea’s guide is important because it tries to convert the AI copyright war from a shouting match into a compliance discipline: prove transformation, avoid substitution, respect access restrictions, and expect scrutiny if you cut corners on data sourcing.

Other regulators should absolutely learn from this—because the alternative is what we already see elsewhere: policy drift, litigation roulette, and a growing gap between what the market does and what the law can credibly govern.