Skip to content
South Korea's AI Basic Act Took Effect on 22 January 2026 With Labelling Rules, a Safety Duty for the Largest Models and a Grace Period of at Least One Year Before Fines
AI & Technology

South Korea's AI Basic Act Took Effect on 22 January 2026 With Labelling Rules, a Safety Duty for the Largest Models and a Grace Period of at Least One Year Before Fines

Illustration by tuput

English

Act No. 20676 and its Enforcement Decree came into force on 22 January 2026. They set up a presidential AI committee, support for data, start-ups and data centres, and duties on transparency, high-impact AI and the very largest models. The Ministry of Science and ICT says fines wait at least a year.

· · 10 min read

South Korea’s Basic Act on the Development of Artificial Intelligence and the Creation of a Foundation for Trust came into force on 22 January 2026, together with its Enforcement Decree. Korea’s Ministry of Science and ICT (MSIT) says the law does two jobs at once: it builds the machinery for a national AI industry, and it sets duties on companies that use generative AI, “high-impact” AI and the largest models. The ministry also says no fines will be imposed for at least a year.

This is a plain description of the law as the ministry and the Act’s English text give it, not legal advice. Dates and figures are as of 12 November 2026.

How the law came about

The National Assembly passed the Act in December 2024 with agreement between the ruling and opposition parties, according to MSIT. The Act (Act No. 20676) is dated 21 January 2025 in the English text kept by the Korea Legislation Research Institute, and its addenda set entry into force one year after promulgation.

Drafting the rules beneath it took most of 2025. MSIT says it formed a working group of about 80 private-sector experts in January 2025 and held more than 70 consultation sessions. A draft Enforcement Decree appeared in September 2025, went to legislative notice on 12 November 2025, and the Cabinet approved it on 20 January 2026. MSIT also ran more than 20 briefings for Korean and foreign companies.

MSIT describes the guiding idea as a “minimum regulation principle”: few obligations on business, and many measures to help the sector grow.

What the Act builds for the industry

The first half of the Act is about promotion. It gives a legal basis to a National Artificial Intelligence Committee under the President (MSIT’s release calls it the Presidential Council on National Artificial Intelligence Strategy). The Act’s text says the President chairs it, it has up to 45 members, and most of them must be outside experts. The same legal basis covers expert standing committees, a secretariat and Chief AI Officers.

The Act and the Decree then list support for AI research and development, training datasets, adoption of AI by firms and public bodies, new businesses, AI talent and AI data centres. The Decree also lets the government designate AI clusters, and it allows public bodies and public universities to open their facilities to companies for testing and performance evaluation of AI.

Telling people when AI is in use

The transparency duty sits in Article 31 of the Act. An operator offering a product or service built on high-impact or generative AI must tell users beforehand. An operator providing generative AI must label that output was generated by it. Where an AI system produces sound, images or video that are hard to tell from the real thing, the operator must make that fact clearly recognisable, though an artistic or creative work may be labelled in a way that does not interfere with enjoying it.

The Decree and MSIT’s transparency guidelines, released on 22 January 2026, supply the detail:

  • Deepfakes (AI-manipulated content that may cause social concern) need clear labelling that users can easily recognise.
  • Other AI-generated content, such as animation and webtoons, may carry an invisible watermark, or a pop-up or interface notice that it is AI-generated.
  • Content that stays inside a service, such as a chatbot reply on screen, can use lighter labelling. Content that users download or share must carry a visible or audible marker such as a watermark, or machine-readable information such as metadata, after a text or voice notice.
  • The duty falls on operators who provide AI products and services to users, and that includes foreign firms serving users in Korea. It does not fall on people or firms that use AI tools in their own work. MSIT’s example is a film company that makes and distributes a film with AI video tools: it counts as a user.

MSIT calls watermarking “a minimum safeguard” against misuse such as deepfakes and says leading international companies already apply it. Tuput’s piece on deepfakes and elections covers one setting where such labels matter.

High-impact AI and the largest models

The Act defines high-impact AI as a system likely to have a significant effect on, or pose a risk to, human life, physical safety and fundamental rights, used in listed areas. MSIT’s release names ten: energy, drinking water, healthcare, the nuclear sector, crime investigation, recruitment, credit assessment, transport, public services and education. Whether a human makes the final decision matters too. According to MSIT, a system in which a person makes the final call is treated as controllable and falls outside the definition.

Under Articles 33 to 35, an operator must review in advance whether its system is high-impact and may ask the minister to confirm. For high-impact systems the Act requires a risk management plan, a plan to explain results to the extent technically possible, user protection plans, human oversight, and documents that show what was done. Impact assessments on fundamental rights are something operators “shall endeavor” to carry out.

A separate duty applies to the very largest models. Article 32 asks operators of systems whose training used at least a compute standard set by decree to identify, assess and reduce risks across the system’s life, set up a way to respond to safety accidents, and report the results to the minister. MSIT says the Decree sets three conditions that must all be met: cumulative training compute above 10^26 floating-point operations, state-of-the-art technology, and a risk of broad and significant impact on fundamental human rights. The Daily Star, citing The Korea Herald, notes that this is a contrast with the EU, whose rules turn on how an AI system is used; Korea’s extra safety duty turns on a measure of compute, so few systems are expected to be covered.

Foreign companies and the size of the fines

An operator with no address or business office in Korea must name a domestic agent if it meets thresholds set by the Decree. The US International Trade Administration lists them as prior-year total revenue of at least 1 trillion won, prior-year AI-service revenue of at least 10 billion won, or at least one million average daily domestic users over the three months before year end. The agent handles submissions to the ministry and requests for confirmation of high-impact status. Cooley adds that it answers government inquiries.

Article 43 sets administrative fines of up to 30 million won (Cooley converts that to about US$21,000) for three things: failing to give the advance notice under Article 31(1), failing to appoint a domestic agent, and failing to comply with a ministry order to stop or correct a violation. The Daily Star reports that a company which fails to label generative AI content could face penalties of up to 30 million won. The Act’s own text routes labelling duties through a different step first: Article 40 lets the minister investigate suspected breaches and order them to stop or be corrected, and a fine follows if that order is ignored.

The grace period

MSIT says it will allow a grace period of at least one year from 22 January 2026. During it, fact-finding investigations and penalties are deferred, with investigations only in highly exceptional cases such as a death, a rights violation or other serious social harm. The one-year minimum therefore runs to at least 22 January 2027. The duties in the Act already apply; what is paused is enforcement.

The ministry paired the pause with an AI Basic Act Support Desk staffed by experts who helped draft the rules, with confidential and anonymous consultations, and with a working group on improving the Act that it said would start in February 2026. It also said the guidelines would be refined during the year.

The length of the pause was debated before the law began. At a National Assembly audit on 29 October 2025, according to the Asia Business Daily, one lawmaker asked for about three years for start-ups and ventures, and the minister said the government was considering a grace period of more than one year for fines. The ministry’s January announcement committed to at least one year. The sources reviewed for this article report no amendment that changes it.

Industry reaction in the first week was mixed. The Korea Times reported on 29 January 2026 that companies were still working out what counts as high-impact AI, which has no numerical test, and that a professor at Korea University suggested using the grace period to revise the law.

”First” or “second”?

Korea.net and the Korea Times call the Act the world’s first full AI law. MSIT’s own release calls it “the second AI regime enacted in the world”. The European Commission says the EU AI Act entered into force on 1 August 2024 and became applicable on 2 August 2026, with some exceptions, so the order depends on whether passage, entry into force or full application is counted. This article does not rank them. For the EU’s dates and tiers, see the EU AI Act explainer.

How India’s guidelines take another route

India has put its framework in a document rather than a statute. The Ministry of Electronics and Information Technology (MeitY) unveiled the India AI Governance Guidelines under the IndiaAI Mission on 5 November 2025, according to a report by ANI in The Tribune. The 68-page document has a preface by the Principal Scientific Adviser, Prof. Ajay Kumar Sood, and rests on seven sutras: Trust, People First, Innovation over Restraint, Fairness and Equity, Accountability, Understandable by Design, and Safety, Resilience and Sustainability. The guidelines say the sutras were adapted from the Reserve Bank of India’s FREE-AI Committee report.

The guidelines state that “a separate law to regulate AI is not needed given the current assessment of risks” and name existing laws on information technology, data protection, consumer protection and criminal matters as the tools to use, with amendments considered where gaps appear. They favour voluntary measures and what they call techno-legal solutions, meaning technology that helps give effect to policy. Their recommended machinery includes an AI Governance Group supported by a Technology and Policy Expert Committee, technical work by the AI Safety Institute, a national database of AI incidents, an India-specific risk framework and regulatory sandboxes. Everything in them is a recommendation, and sector regulators keep enforcement.

On labelling, the guidelines describe watermarks as a popular method for showing that content was generated or modified by AI, note limits such as bypassing by bad actors, and recommend that an expert committee develop global standards and that the proposed AI Governance Group review India’s rules on content authentication. The Principal Scientific Adviser summed up the approach as “Do No Harm”, as ANI reported, and MeitY’s Secretary S. Krishnan tied it to India’s digital public infrastructure (the shared platforms the guidelines name as Aadhaar, UPI and DigiLocker), and the guidelines describe that base as central to the approach. The IndiaAI Mission supplies the compute and models side of the same plan, and India’s digital public infrastructure is covered separately.

Set side by side, the two countries answer the same questions with different tools. Korea wrote a framework Act with a decree and a fine schedule; India has guidelines, existing statutes and voluntary measures. Korea sets a compute number for extra safety duties; India’s document proposes an India-specific risk framework built from evidence of harm. Both name deepfakes and content provenance as a concern: Korea through labelling rules in force since 22 January 2026, India through a recommended review of its content authentication rules.

Share
Copied!

Sources & further reading

  1. Ministry of Science and ICT (Korea): The AI Basic Act Comes into Force to Lay the Foundation for Korea to Become an AI G3 (press release, 22 January 2026)
  2. Ministry of Science and ICT (Korea): MSIT Releases Guidelines on Ensuring AI Transparency (press release, 22 January 2026)
  3. Korea Legislation Research Institute: Framework Act on the Development of Artificial Intelligence and the Creation of a Foundation for Trust (Act No. 20676, English translation)
  4. Korea Ministry of Government Legislation, law.go.kr: Framework Act on the Development of Artificial Intelligence (English, enforcement date 22 January 2026)
  5. Center for Security and Emerging Technology, Georgetown University: Framework Act on the Development of Artificial Intelligence and Establishment of Trust (translation, 9 July 2025)
  6. Korea.net: AI Basic Act takes effect as world's first set of laws for AI (22 January 2026)
  7. U.S. International Trade Administration: South Korea AI Basic Act (market intelligence, 29 May 2026)
  8. Cooley LLP: South Korea's AI Basic Act, Overview and Key Takeaways (27 January 2026)
  9. The Korea Times: Unclear guidance, vague terms in AI Basic Act leave businesses in limbo (29 January 2026)
  10. The Daily Star: South Korea rolls out new law to regulate AI
  11. The Asia Business Daily: Vice Prime Minister Bae Kyunghoon on a grace period for the AI Basic Act (29 October 2025 audit)
  12. European Commission: AI Act (policy page)
  13. MeitY, Government of India: India AI Governance Guidelines, Enabling Safe and Trusted AI Innovation (PDF hosted by PIB)
  14. The Tribune (ANI): MeitY unveils India AI governance guidelines to ensure safe, inclusive and responsible AI adoption (5 November 2025)

Researched and written with the help of AI tools and edited for accuracy. Provided for general information and discussion only, not professional advice. See our editorial standards and disclaimer. Spotted an error? Tell us.

#south korea#ai basic act#ai regulation#ai labelling#high-impact ai#india ai governance guidelines#ministry of science and ict

Enjoyed this? Get the next one.

One good read at a time, straight to your inbox. No spam, unsubscribe anytime.

More in AI & Technology
MeitY Released India's AI Governance Guidelines on 5 November 2025: Seven Sutras, Six Pillars and a Plan That Builds on Laws India Already Has
What India's first national AI governance framework recommends, who wrote it, and which of its ideas have since turned into rules, committees and working papers.
The EU AI Act Has Applied in Stages Since 2 February 2025, and Its Rules for High-Risk Systems Now Wait Until 2 December 2027 and 2 August 2028
Which parts of the EU's AI law already bind companies, which have been pushed back, and what each tier asks for, read from the Commission's own text and pages.
An AI-Made Biden Voice Placed 9,581 Calls Before New Hampshire's January 2024 Primary, and the FCC, the EU and India Have Since Written Different Rules for AI Fakes in Elections
Two election incidents and what the courts and regulators did about them, then a dated look at what the Election Commission of India, MeitY, Brussels, YouTube, Meta and X now require when AI makes a political fake.
← all articles