South Korea's AI Basic Act Takes Effect With Safety Rules, Industry Support and a Grace Period

A new AI framework begins with a year of guidance

South Korea's AI Basic Act took effect on Thursday, introducing six chapters and 43 articles covering artificial intelligence development, services, safety and industry support. As of January 25, 2026, the government is presenting it as the world's first comprehensive AI framework to begin operating, while giving businesses at least one year of guidance before investigations and administrative sanctions.

The law reaches from labels on realistic content generated by AI to safeguards for decisions affecting employment, loans and public services. Foreign companies can also face obligations: businesses meeting thresholds that include global annual revenue of 1 trillion won, domestic sales of 10 billion won or at least 1 million daily users in South Korea must appoint a domestic representative.

South Korea wants the framework to protect citizens while helping the country become one of the world's top three AI powers. Its approach combines support for research, startups and skilled workers with disclosure requirements, human oversight and administrative fines capped at 30 million won.

Kim Kyeong-man, deputy minister of the office of AI policy at the Ministry of Science and ICT, described trust as the purpose of the legislation:

We are approaching this from the most basic level of global consensus. The goal is not to stop AI development through regulation. It's to ensure that people can use it with a sense of trust.

Startups are less certain about how that balance will work. A Startup Alliance survey found that only 2% of startups focused on AI had a formal compliance plan, while approximately half said they did not fully understand the requirements. The immediate task for the government is therefore practical: explain which businesses and systems are covered, and what they must change.

How the law reached January 2026

The National Assembly passed the AI Basic Act on December 26, 2024, with overwhelming bipartisan support. The vote came during political turmoil surrounding President Yoon Suk Yeol's declaration of martial law and subsequent impeachment, showing broad agreement across political parties on the need to address AI development and its risks.

The legislation allowed a year of preparation before taking effect in January 2026. During that period, detailed rules and guidance for individual sectors were expected to explain how its broad provisions would work. Earlier implementation plans anticipated subordinate legislation in the first half of 2025. Subordinate legislation means the detailed rules issued under an act to put its requirements into practice.

• December 26, 2024: The National Assembly passed the act with overwhelming bipartisan support.
• First half of 2025: The earlier timetable anticipated subordinate legislation explaining regulatory details.
• Thursday before January 25, 2026: The AI Basic Act took effect after the preparation period.
• At least the first year of operation: Authorities plan to focus on consultation and education rather than investigations and administrative sanctions.
• August 2026: The core bulk of the European Union's AI Act is scheduled to begin a phased rollout.

The preparation period and the enforcement grace period serve different purposes. The first preceded the law's commencement; the second gives businesses time to adjust after it takes effect. The ministry is considering extending the grace period if conditions in domestic and overseas industries warrant it. No final decision on an extension has been announced.

Which uses count as high impact?

The act gives particular attention to AI that can substantially affect human life, safety or fundamental rights. An automated system screening a loan application or job candidate illustrates the concern: its output can shape a person's access to money or work.

The legislation identifies 11 areas for this category of AI:

• Energy supply systems.
• Production processes for drinking water.
• Healthcare service systems.
• Medical device development and use.
• Management of nuclear materials and facilities.
• Analysis of biometric information for criminal investigations.
• Employment and loan assessments.
• Transportation systems and management.
• Government decisions affecting public services.
• Student assessment in education.
• Other areas substantially affecting human life, safety and rights.

Companies operating covered systems must be able to explain the logic behind their AI decisions when asked, allow human intervention when necessary and tell users that AI is involved. These duties aim to give people an understandable account of a consequential decision and a route for a person to step in.

South Korean officials say no domestic services currently meet the criteria for this category. They identify fully autonomous vehicles at Level 4 or higher as a possible future example. In that setting, the AI's role in operating the vehicle makes safety and the ability to intervene central concerns.

The act also includes Fundamental Rights Impact Assessments, or FRIAs, among the measures associated with consequential AI, as does the EU framework. These assessments examine how an AI system could affect people's basic rights. The precise regulatory burden depends on detailed implementing rules and guidance for the relevant sector.

Powerful models face a different test

South Korea separates AI with major effects on people's lives from what it calls high performance AI. The first category concerns the use of a system. The second concerns technical characteristics of the model itself, including the cumulative computing work used to train it.

Training computation is the amount of calculation used while an AI model learns. A threshold based on that measure can bring an exceptionally powerful model under additional safety rules, regardless of the particular service in which it is later used. Officials say no existing model, in South Korea or elsewhere, currently meets the criteria for regulation under this provision.

The two categories therefore cover different concerns. Employment, healthcare and transport appear in the act because of the consequences of AI decisions in those settings. Computing thresholds address advanced models through a separate technical test. Under current capabilities, the latter would cover only a very limited set of models.

The EU's approach places strong weight on risks linked to particular uses, such as healthcare, recruitment and law enforcement. South Korea also identifies sensitive uses, while applying technical thresholds to its separate category for powerful models. Keeping those categories distinct helps explain why the law can name 11 areas for oversight while officials say no current domestic service or existing advanced model meets the relevant criteria.

What users will see on AI services and content

Companies providing products or services using generative AI or AI in the category with major effects on people must notify users in advance. Generative AI produces content, such as text or images. Notices can take the form of pop-up alerts or prominent statements explaining that AI is part of the service.

Content that could be mistaken for something real faces a more visible disclosure requirement. Companies must provide clear visual or audible labels, including watermarks where appropriate. The rule targets deepfakes and other outputs that can mislead people about whether an image, recording or event is genuine.

A Ministry of Science and ICT official described the purpose of watermarking:

Applying watermarks to AI-generated content is the minimum safeguard to prevent side effects from the abuse of AI technology, such as deepfake content.

Creative works receive more flexible treatment. Games, animations and webtoons made using AI can use less intrusive disclosures, including labels in metadata. Metadata is information attached to a digital file that can describe its origin or contents without appearing directly in the image or story.

The difference matters to creators who want to preserve the appearance of a work while disclosing AI involvement. Where authenticity or safety depends on knowing whether something is real, the label must be visible or audible. For creative material, the disclosure can be placed where it does not interrupt the work.

Webtoon platforms and creators remain concerned about where the obligation begins. Partial use of AI, such as generating backgrounds, adding color or editing a work, creates uncertainty about what counts as content generated by AI. The law applies to companies and platforms offering creation tools powered by AI or distributing works generated by AI, rather than individual users simply employing those tools.

Who must comply, including companies overseas?

The act divides AI businesses into development operators and utilization operators. Development operators build and provide AI. Utilization operators use AI supplied by developers to offer their own products or services. Both categories face transparency obligations, with the main regulatory attention directed toward generative AI and uses with major effects on people.

South Korea's framework defines obligations broadly across these operators. The EU framework assigns different duties to different participants in the chain of companies that develop, supply and deploy AI. South Korea's simpler structure may make the rules easier to describe, while leaving practical questions about how duties apply across a diverse industry.

Foreign AI companies operating in South Korea or serving Korean users must designate a domestic representative when they meet specified criteria. The thresholds are global annual revenue of at least 1 trillion won, given as $681 million; domestic sales of at least 10 billion won; or at least 1 million daily users in the country.

OpenAI and Google currently fall within these requirements. A domestic representative provides a local point of contact for the authorities, allowing officials to communicate with a company whose main operations are overseas.

The obligation places large foreign providers within the same national framework that governs domestic businesses. For smaller companies, the more immediate questions concern their role as a developer or service provider, the kind of AI they use and the notices their customers must receive.

Correction comes before heavy punishment

South Korea's law uses administrative enforcement rather than criminal penalties. Its stated approach prioritizes compliance and corrective orders, with fines capped at 30 million won, given as approximately $26,200. Ignoring a suspension or corrective order can lead to a fine.

The framework identifies three specific violations subject to fines: failing to make the required notification about operations using generative AI or AI with major effects on people, failing to appoint a domestic representative when required, and failing to comply with a suspension or corrective order from the Minister of Science and ICT.

The account of the notification offence describes notification to authorities, while the transparency provisions separately require advance notice to users. Businesses need to address the relevant notification duties alongside the labels and explanations required for their products.

The EU allows much larger penalties, with different levels depending on the violation and a maximum reaching 7% of global annual turnover. Turnover means a company's revenue, so a fine based on worldwide turnover grows with the size of the business. South Korea instead sets a fixed ceiling in won.

The grace period further softens the immediate enforcement burden. For at least one year, authorities plan to avoid fact-finding investigations and administrative sanctions, concentrating on consultations and education. The government presents that approach as a way to establish basic safeguards without discouraging development.

Industry support is part of the legal framework

The Ministry of Science and ICT is the main authority implementing AI policy. The National AI Committee operates under the president and deliberates on broader policy directions. The act also creates an AI Policy Center and an AI Safety Institute to support implementation and address safety concerns.

The legislation supplies a legal basis for a Basic Plan for AI, a national roadmap for technology development. Its support measures cover research and development, technical standardization, development of data used to train AI and help for companies adopting AI technologies.

Small and medium-sized enterprises, startups and the development of skilled workers receive particular attention. The government's aim is to combine minimum regulation with support for competitiveness, rather than use restrictions alone to shape the industry.

A dedicated Framework Act on AI Help Desk offers confidential consultations to people developing AI, doing business with it or using AI tools. Experts from organizations including the AI Safety Institute and the National Information Society Agency help businesses determine whether their systems fall within the law and how to comply.

The ministry also plans a guidance platform and a dedicated support center during the grace period. These services are intended to turn broad legal requirements into practical answers for companies facing decisions about disclosure, system classification and oversight.

The National AI Committee's coordinating role matters because AI reaches across personal information, copyright, healthcare and defense. Different ministries hold expertise in different sectors. The framework calls for coordination so that rules reflect the characteristics of each activity while supporting a common national policy.

Startups question the cost of going first

The Startup Alliance survey points to a gap between the law's commencement and businesses' readiness. Only 2% of startups focused on AI said they had a formal compliance plan. Approximately half acknowledged that they did not fully understand the requirements.

Founders worry that vague language will encourage overly cautious development choices. They also question whether compliance costs arriving earlier than comparable requirements elsewhere will leave Korean businesses at a disadvantage against foreign competitors.

Lim Jung-wook, co-head of South Korea's Startup Alliance, expressed that frustration:

There's a bit of resentment — why do we have to be the first to do this?

President Lee Jae Myung has acknowledged industry concerns. During a recent meeting with aides, he urged policymakers to listen to businesses, give venture companies and startups adequate support, maximize the industry's potential through institutional support and manage anticipated side effects before they grow.

A ministry spokesperson said the government continues to review ways to reduce the burden on industry. Consultations, guidance and the possibility of a longer grace period are its main responses. For startups, the value of those measures will depend on whether they provide clear answers about the products companies are already building.

How South Korea fits into the global picture

South Korea describes its law as the first comprehensive AI framework to take effect, while also following the European Union in adopting a broad regulatory framework. The timing explains the distinction: the EU began implementing some AI rules in 2025, but the core bulk of its act is scheduled for a phased rollout beginning in August 2026.

The United States has no federal AI law. In December 2025, the White House issued an executive order challenging state laws it described as onerous and as obstructing AI innovation. China has taken a different route, adding provisions concerning ethics or risk monitoring to its existing cybersecurity law.

South Korea's approach draws on both private sector autonomy associated with the US model and the EU's focus on safety and reliability. Its combination of modest administrative penalties, targeted safety duties and industry support is intended to build competitiveness alongside public trust.

Officials describe the legislation as a starting point. Detailed rules and guidance for particular sectors will shape its operation, while coordination among ministries will become more demanding as AI spreads into more activities. International cooperation also matters because companies and AI services operate across borders, making compatible approaches to trustworthy AI increasingly valuable.

The immediate test is narrower and more concrete than South Korea's global ambitions. Businesses need to understand which category their systems fall into, users need meaningful notices, and authorities need to make the grace period useful. The act has begun operating; the next stage is putting its safeguards and support measures into everyday practice.

The Bottom Line

• The AI Basic Act took effect in January 2026 with six chapters and 43 articles covering AI governance, safety and industry support.
• The National Assembly passed it on December 26, 2024, with overwhelming bipartisan support.
• The law identifies 11 areas where AI can affect life, safety or rights and requires explanations, human intervention and user notices for covered systems.
• A separate technical test covers powerful models. Officials say no existing models meet that threshold and no current domestic services meet the criteria for the category with major effects on people.
• Realistic content generated by AI requires clear labeling; creative works can use less intrusive disclosures such as metadata.
• Foreign providers meeting revenue or user thresholds must appoint a domestic representative. OpenAI and Google fall within the requirements.
• Administrative fines are capped at 30 million won. The law imposes no criminal penalties.
• Authorities plan at least one year of consultation and education without investigations or administrative sanctions.
• Only 2% of startups focused on AI reported a formal compliance plan, while approximately half said they did not fully understand the law.
• Research support, training data, help for smaller businesses and talent development form part of South Korea's ambition to become a top three global AI power.