Eazler AI Lab AI Engineering Journal
Skip to article
Guides◷ 13 min read

South Korea's AI Basic Act: no direct administrative fine for failing to label generative AI output — checked against the articles

Labeling generative AI output is a legal duty under Korea's AI Basic Act, but Article 43, which sets administrative fines, does not list it.

On this page

Korea's AI Basic Act (인공지능 발전과 신뢰 기반 조성 등에 관한 기본법, literally the "Framework Act on the Development of Artificial Intelligence and the Creation of a Foundation for Trust") took effect on January 22, 2026. People often say that generative AI output must be labeled, but Article 43, which sets administrative fines, lists only three things, and output labeling is not among them. We read the Act and its Enforcement Decree (시행령) directly on law.go.kr, the Korean government's statute database (국가법령정보센터), and separated what you are obliged to do from what you can be sanctioned for.

Korean original: 한국어 원문 (2026-09-08)Scope: transparency obligations / high-impact AI / administrative finesAudience: developers · one-person studios · small teamsBasis: original text of Act No. 20676 (법률 제20676호) · Presidential Decree No. 36053 (대통령령 제36053호)

1. Executive summary

The Act places three types of obligations on AI business operators (defined in section 2): transparency (Article 31), responsibilities for high-impact AI (Article 34) and impact assessment (Article 35). But the three differ in how binding they are. Some are phrased “shall,” others “shall endeavor to” (our translation, as are all quotations from the Act and the Decree in this report), and only some of them carry an administrative fine. We cite provisions in the form Article 43(1)(iii), meaning Article 43, paragraph 1, subparagraph 3.

2026-01-22The Act and the Enforcement Decree took effect together
2026-01-24 for the digital medical device part only
3Violations listed for administrative fines
Up to ₩30 million (KRW)
11Areas of use listed
as high-impact AI
Conclusion: Keep obligations and administrative fines apart. If your service uses generative AI, Article 31(1) (advance notice) and Article 31(2) (labeling outputs) are both obligations. However, of the three paragraphs of Article 31, only paragraph 1 carries an administrative fine directly; a violation of paragraph 2 leads to a sanction only through a corrective order. (Across the whole Act, the administrative fines cover the three cases listed in Article 43(1), and the only one of them tied to Article 31 is paragraph 1.) The obligations on high-impact AI do not apply unless you use AI in one of the listed areas. But reviewing whether you fall under them (Article 33(1)) is itself an obligation, whether or not you do.

2. Does it apply to us?

First, check whether you are an AI business operator. Article 2(vii) of the Act defines this as “a person engaged in business related to the artificial intelligence industry” and includes corporations, organizations, individuals and State agencies, etc. (the Act's term for the State, local governments, public institutions and similar bodies). An individual developer who does AI-related work as a business is also an AI business operator. How ad platforms label AI-generated ads is covered in the labeling section of our Google and Meta ad automation report.

The scope is broad. Article 4(1) provides that the Act “applies even to conduct carried out outside Korea if it affects the domestic market or users.” A foreign company that provides a service to users in Korea is also subject to the Act. Conversely, Article 4(2) excludes AI developed and used solely for national defense or national security, and Article 2 of the Enforcement Decree limits that exclusion to tasks designated by the Minister of National Defense and the Director of the National Intelligence Service. It has nothing to do with ordinary services.

Effective date: Article 1 of the Addenda provides that the Act “shall enter into force one year after the date of its promulgation.” It was promulgated on January 21, 2025, so that is January 22, 2026. The exception is the part of Article 2(iv)(d) covering digital medical devices, which applies from January 24, 2026. The Enforcement Decree (Presidential Decree No. 36053) also took effect on the same day.

3. The three transparency obligations

Article 31 sets out the three separately. They are different paragraphs, and as shown below, the sanctions differ too.

ParagraphWhatWhen
Article 31(1) Tell users in advance that the product or service runs on that AI When you intend to provide a product or service that uses high-impact AI or generative AI
Article 31(2) Label the fact that the output was generated by generative AI When you provide generative AI, or a product or service that uses it
Article 31(3) Give notice or label in a way users can clearly recognize When you provide virtual sound, images, video and the like that are hard to distinguish from real ones

Source · checked against the original text of Article 31 of Act No. 20676.

Methods set by the Enforcement Decree

Article 31(4) leaves the specific methods of notice and labeling, and the exceptions, to a Presidential Decree; they are in Article 23 of the Enforcement Decree. An enforcement decree is subordinate legislation that fills in the details a statute hands off with the words “as prescribed by Presidential Decree.”

Advance notice (paragraph 1)

Any one of four methods is enough: ① state it directly on the product, or in the contract, user manual or terms of service; ② display it on the user's screen or device; ③ post it in an easily recognizable way at the place where the product or service is provided; ④ any other method the Minister (of Science and ICT) approves.

Output labeling (paragraph 2)

Either a method people can perceive or a machine-readable method. If you choose the latter, you must also tell users that the output was made by generative AI at least once, through a text notice, audio or similar means.

Virtual outputs (paragraph 3)

Give notice or label in a way that can easily be checked by sight, by hearing or with software, taking into account the age and the physical and social circumstances of the main users.

There are exemptions. Article 23(4) of the Enforcement Decree provides that all or part of paragraphs 1 to 3 may be disapplied in three cases: ① where the use of AI is obvious, considering the product or service name, the user's screen or the product's exterior, the wording shown on the output and the like; ② where it is used only for the business operator's internal work; ③ other cases that the Minister determines and publicly notifies.

In practice, the first exemption matters a great deal. For a service whose name and screens alone make it clear that it is AI-based, the provisions “may not be applied” even without a separate notice. But whether a case is “obvious” is ultimately for the regulator to judge, so if it is borderline, it is safer to put one line in your terms or on the first screen.

Source · checked against the original text of Article 23 of Presidential Decree No. 36053.

4. Where administrative fines apply

Article 43(1) sets an administrative fine of up to ₩30 million (KRW) and lists three cases it applies to. Violations of provisions that are not listed do not directly incur this fine.

Article 43(1)Who is finedRelevant to a small team?
(i)A person who fails to give notice in violation of Article 31(1)Yes
(ii)A person who fails to designate a domestic agent in violation of Article 36(1)No
(iii)A person who fails to comply with a stop order or corrective order under Article 40(3)Only after the fact

Source · checked against the original text of Article 43 of Act No. 20676.

This is where it splits. The commonly cited generative AI labeling duty is Article 31(2), and the notice for virtual outputs that are hard to tell from real ones is Article 31(3). Neither is listed in Article 43. Nor are the responsibilities for high-impact AI (Article 34) or the impact assessment (Article 35). The only transparency obligation that directly carries an administrative fine is Article 31(1).

That does not make the rest meaningless. Article 40 gives the competent ministry the power to demand materials or investigate in order to check for violations (a fact-finding investigation), and if you do not follow a stop order or corrective order issued after that investigation, an administrative fine applies under subparagraph (iii). In other words, violations of paragraphs 2 and 3 reach a sanction indirectly, by way of an order. But the structure does not produce an administrative fine straight away.

Criminal penalties are narrower still. The penalty provision in Article 42 (imprisonment of up to 3 years or a criminal fine of up to ₩30 million (KRW)) applies only to Article 7(9): disclosing secrets learned in the course of duty. There is no criminal penalty for a business operator's breach of its obligations.

The domestic agent rule rarely applies

A domestic agent is a person or company that handles, in Korea, the tasks the Act requires on behalf of a foreign business operator that has no address or place of business in Korea. Article 36(1) applies only to operators “that have no address or place of business in Korea,” and among those, only to operators meeting the criteria in Article 29 of the Enforcement Decree: prior-year revenue of ₩1 trillion (KRW) or more; revenue of ₩10 billion (KRW) or more from AI services; an average of 1 million or more daily users in Korea over the preceding three months; or having previously been fined under Article 43(1)(iii).

5. Is it high-impact AI?

Article 2(iv) defines high-impact AI as a system “likely to have a significant impact on, or pose a risk to, human life, physical safety and fundamental rights,” and limits it to systems used in the listed areas. There are eleven, items (a) through (k).

ItemArea
(a)Supply of energy
(b)Production process of drinking water
(c)Provision of health and medical services, and building and operating the systems for their use
(d)Development and use of medical devices and digital medical devices
(e)Safe management and operation of nuclear materials and nuclear facilities
(f)Analysis and use of biometric information for criminal investigation or arrest
(g)Judgments or evaluations that significantly affect an individual's rights and obligations, such as hiring and loan screening
(h)Key operation and management of means of transport, transport facilities and transport systems
(i)Decision-making by State agencies, etc., such as verifying and determining eligibility for public services and collecting charges
(j)Student assessment in early childhood, elementary and secondary education
(k)Other areas prescribed by Presidential Decree

Source · checked against the original text of Article 2(iv) of Act No. 20676. The terms in each item cite the definitions in the relevant individual statutes.

What counts is what the AI does, not what kind of app it is in. The question is whether it is used in a listed area. Item (g) in particular leaves room for a broad reading: if you automate judgments or evaluations that significantly affect an individual's rights and obligations, such as hiring screening or credit and loan decisions, it needs review whatever field your service is in.

Whether or not you fall under it, Article 33(1) makes the review itself an obligation: operators “shall review in advance whether that artificial intelligence falls under high-impact artificial intelligence”. Asking the Minister for confirmation is optional; the Act says operators “may.”

If it does fall under it, Article 34(1) requires six measures: establishing and operating a risk management plan; establishing and implementing an explanation plan; establishing and operating a user protection plan; human management and oversight; preparing and keeping documents that show what measures were taken; and matters deliberated and resolved by the Committee (the national AI committee set up under Article 7 of the Act) to ensure safety and trustworthiness. These six are obligations that paragraph 1 frames as “shall implement.” Separately, Article 34(2) provides that the Minister publicly notifies the details and “may recommend compliance.” The impact assessment in Article 35 is framed as “shall endeavor to” assess in advance the impact on people's fundamental rights. A best-efforts duty like this means making reasonable efforts; it does not mean the assessment must be completed.

6. What you can do today

Below, the articles are arranged in a practical order. This is not legal advice; make specific judgments together with a professional.

#To doBasis
1Write down whether your service uses generative AIThe premise of Article 31(1) and (2)
2Review whether you fall under any of the eleven high-impact AI areasArticle 33(1): the review itself is an obligation. The Act does not require you to keep a record of the review, but it will serve as evidence later
3State that the service is AI-based in one place: your terms, first screen or manualArticle 31(1), Enforcement Decree Article 23(1): subject to an administrative fine
4If you cannot tell whether the name and screens alone make it obvious, just state itThe exemption in Enforcement Decree Article 23(4)(i) leaves room for judgment
5Decide how you will label outputs. If you will use metadata only, add a text or audio notice at least onceEnforcement Decree Article 23(2)(ii)
6If you generate realistic sound, images or video, design a separate notice or labelArticle 31(3), Enforcement Decree Article 23(3)
7Confirm that the Act applies even if you serve users in Korea from a company abroadArticle 4(1)
Item 3 is the cheapest safeguard. It is the only transparency obligation that directly carries an administrative fine, and if you choose the decree's option of stating it in the contract, user manual, terms of service or the like, one sentence does the job.

7. What this is not yet

  • This is not legal advice. It is our reading and summary of the articles; where a particular service falls depends on the facts.
  • We could not check the criteria for imposing administrative fines. The Act sets only the maximum amount, ₩30 million (KRW); the actual amounts by type and number of violations are set in Annex Table 2, attached at the end of the Enforcement Decree. An annex table is a document attached separately from the main text, and we were not able to open Annex Table 2. There are secondary sources that give amounts by number of violations, but we could not confirm them against the original, so we did not use them.
  • We could not confirm a grace period. A grace period here means a period during which, even when the law is broken, guidance and correction take priority over sanctions for a while. Material from several law firms says that enforcement of the rules will be deferred for at least one year after the Act takes effect, but we could not find the competent ministry's original announcement. Do not put off preparing on the assumption of a deferral.
  • We did not cover public notices or guidelines. A public notice is a rule in which a ministry sets detailed standards and announces them officially; a guideline is guidance material that explains the criteria and gives examples. Article 33(3) provides that the Minister may distribute guidelines on the criteria for, and examples of, high-impact AI, and Article 34(2) also anticipates a public notice. This report looked only at the Act and the Enforcement Decree.
  • Article 2 was amended on January 20, 2026. The amendment came two days before the Act took effect, and the text we read is the consolidated version as of the effective date, including that amendment. We could not confirm whether there have been further amendments since then, as of September 2026.
  • There are no interpretations yet of what counts as “obvious.” How far the exemption in Enforcement Decree Article 23(4)(i) is actually recognized will only become clear once the competent ministry issues interpretations explaining the provision's meaning and how it applies, or once real cases accumulate.

8. Sources

SourceWhat we used it forLink
AI Basic Act (인공지능 발전과 신뢰 기반 조성 등에 관한 기본법)
The header of the linked page reads [시행 2026. 1. 22.] [법률 제20676호, 2025. 1. 21., 제정] (in force 2026-01-22; Act No. 20676, enacted 2025-01-21), but the same page states that it is a statute text “including all amendments made after promulgation and before entry into force” (our translation). Article 2 does carry the marker <개정 2026. 1. 20.> (amended 2026-01-20), so we read the text including that amendment.
Article 2 (definitions; the eleven high-impact areas), Article 4 (scope), Article 31 (transparency), Articles 33 to 36, Article 42 (penalties), Article 43 (administrative fines), Addenda Article 1. Checked against the original text on law.go.kr law.go.kr — text of the Act
Enforcement Decree of the same Act (인공지능 발전과 신뢰 기반 조성 등에 관한 기본법 시행령)
Presidential Decree No. 36053, enacted 2026-01-21 / in force 2026-01-22
Article 2 (exclusions), Article 23 (methods of notice and labeling, and exemptions), Article 29 (criteria for a domestic agent), Article 32 (delegation of the criteria for administrative fines). Checked against the original text law.go.kr — text of the Enforcement Decree

We checked the articles directly on law.go.kr on 2026-09-08. We did not cite Korean-language summaries, because they often mix up article and paragraph numbers. In fact, one secondary source that came up in search while we were writing gave the effective date as July 21, 2026, which does not match the Addenda. Statutes get amended, so check the original text again before you act on this.

Newsletter

The week’s new reports, every Sunday

One email a week with that week’s reports, skipped when nothing new came out. Nothing but report news.

Used only to announce new reports. Unsubscribe any time. Privacy Policy
Publisher

Eazler AI Lab is the AI engineering journal of Eazler Inc. We read primary sources in full, and say so when something is unverified.

Editorial standards and method →

We also post report summaries on Threads: @eazler.inc.

Building or adopting AI agents? hello@eazler.com