---
title: South Korea's AI Basic Act Takes Effect With Labeling Duties for Generative AI Output
dek: South Korea's framework law on artificial intelligence took effect on 22 January 2026 and requires providers to label generative AI output, in machine-readable form if they choose. The science ministry says it will hold off on investigations and fines for at least a year.
kind: news
topics: [content-provenance, governance-compute]
layers: [institutions, outputs]
event_date: 2026-01-21
published: 2026-10-20
agent: daily
evidence: [regulation, announcement]
sources:
  - Framework Act on the Development of Artificial Intelligence and the Creation of a Foundation for Trust (인공지능 발전과 신뢰 기반 조성 등에 관한 기본법), as amended by Act No. 21311 | National Law Information Center, Ministry of Government Legislation | 2026-01-20 | https://www.law.go.kr/LSW/lsInfoP.do?lsiSeq=282791
  - Enforcement Decree of the Framework Act (인공지능 발전과 신뢰 기반 조성 등에 관한 기본법 시행령), consolidated text as amended by Presidential Decree No. 36580 | National Law Information Center, Ministry of Government Legislation | 2026-08-18 | https://www.law.go.kr/LSW/lsInfoP.do?lsiSeq=288781
  - AI Basic Act takes effect as the foundation for becoming a top-three AI power (인공지능 3대 강국의 기틀인 「인공지능기본법」 시행) | Ministry of Science and ICT | 2026-01-21 | https://www.msit.go.kr/bbs/view.do?sCode=user&mId=307&mPid=208&bbsSeqNo=94&nttSeqNo=3186786
  - Framework Act on the Development of Artificial Intelligence and the Creation of a Foundation for Trust (English translation) | Korea Legislation Research Institute | accessed 2026-10-20 | https://elaw.klri.re.kr/eng_mobile/viewer.do?hseq=73499&type=part&key=18
related: [india-synthetic-media-labeling-rules, eu-ai-content-labeling-code, china-ai-labeling-enforcement-cases]
---

South Korea's AI Basic Act (인공지능기본법) entered into force on 22 January 2026, in the version amended two days earlier. Its formal name is the Framework Act on the Development of Artificial Intelligence and the Creation of a Foundation for Trust. Its Enforcement Decree, the decree that implements the Act, took effect the same day.[^1, 2, 4]

Article 31 requires AI business operators to tell users in advance when a product or service runs on high-impact or generative AI. Operators must also mark generative AI output as AI-generated. When operators use an AI system to provide sound, images or video that are hard to distinguish from real ones, they must disclose this in a way users can clearly recognize.[^1, 4]

Under Article 23 of the decree, a label on generative AI output may be perceptible to people or machine-readable. A provider that chooses a machine-readable label must also tell users at least once that the content is AI-generated, in text or by voice. Users must be able to check a disclosure on realistic synthetic media easily, by sight or hearing or with software.[^2] The Ministry of Science and ICT (MSIT) says invisible watermarks are allowed for outputs that are not deepfakes, such as animation and webtoons (web comics).[^3]

MSIT may investigate suspected labeling breaches and order corrections. Fines of up to 30 million won apply to operators that skip the advance notice or ignore such an order. A labeling breach alone is not fined.[^1] MSIT says it will hold off on investigations and fines for at least a year. It says that during that period it will investigate only in exceptional cases, such as loss of life.[^3]

Separate safety duties, including reporting results to MSIT, apply to the most advanced systems. These are systems that could seriously affect people's lives, safety and rights and that were trained with at least 10²⁶ floating-point operations, a measure of the computing used in training. The decree leaves the method of counting these operations to a ministerial notice.[^1, 2]

:::analysis
The decree accepts machine-readable labels but specifies no format. Unless MSIT's guidelines or notices settle one, a label written by one service may be unreadable to another platform or to an outside checker. Because the duty falls on AI business operators, a missing label does not show that a file is authentic. Safety results go to MSIT and need not be published, so the law gives outsiders no way to check which systems cross the compute threshold.
:::

:::checked
On 20 October 2026 we read the Act and decree in Korean on the National Law Information Center. We also read the Korea Legislation Research Institute's English translation of the Act and MSIT's press release of 21 January 2026. English names follow that translation. We did not review MSIT's guidelines or notices, check for changes to the grace period or examine how any service labels its output.
:::