Summary: South Korea’s AI Basic Act takes effect on January 22, 2026, becoming the first comprehensive national AI law to reach full effect. It applies to activities affecting the Korean market, defines high-impact AI at 10^26 floating-point training operations, and requires labels or watermarks for commercial generative AI output. Penalties can reach KRW 30 million, with at least a year of guidance-first enforcement.
On January 22, 2026, South Korea’s AI Basic Act enters into force. Layered with the final revisions the National Assembly pushed through in the closing days of December 2025, it becomes the first comprehensive national AI law anywhere to reach full effect — beating the EU AI Act’s complete application in August 2026 by the better part of a year. With two weeks on the clock, any team whose product touches the Korean market should already be checking whether it sits inside the jurisdiction.
Officially the Framework Act on the Development of AI and the Creation of a Foundation for Trust, the statute passed the National Assembly on December 26, 2024, was promulgated in January 2025, and folds 19 separate AI bills into one. Korean coverage from the year’s opening days observed that the last-minute revision tightened risk controls while leaning harder into industrial promotion.
A Law That Reaches Beyond Korea’s Borders
The jurisdiction refuses to stop at the border. Any AI activity affecting the South Korean market — domestic or overseas — falls under the act, with national defense and security uses as the lone exception; providers lacking a Korean presence must appoint a domestic representative. The revision also opens a door for universities and private companies, not just the science ministry, to establish AI research institutes and accumulate sovereign AI capability, and obliges public institutions to give AI products and services priority in procurement. Governance receives an upgrade alongside: the Presidential Council on National AI Strategy gains formal legal standing and expanded authority as the central coordinating body — harmonizing ministries, setting major policy, and monitoring implementation in place of non-binding guidance.
High-Impact AI, Defined by Compute
The revised act hands down an unusually concrete threshold: models trained with at least one hundred septillion (10^26) floating-point operations qualify as high-impact AI, with individual cases then assessed by application area, risk to fundamental rights, and severity. The examples span healthcare, hiring, loan screening, and biometric analysis — which places a health-assistant feature like ChatGPT Health, if offered in Korea, squarely inside the review perimeter. Operators owe risk-management plans and continuous monitoring of social impact, while regulators may demand data submissions and run on-site inspections.
Labeling and Watermarks for Generative AI
The duties do not halt at high-impact systems. Every commercial AI service provider must disclose AI-generated content through notices or technical measures such as watermarks, with the crosshairs on deepfakes, misinformation, and deceptive advertising. The practical translation for product teams: build labeling into the interface and the output pipeline from day one — it is not a decoration bolted on after launch.
Light Penalties, Long Grace Period
Violations — unlabeled content, no domestic representative appointed — draw fines of up to KRW 30 million (about $21,000), with imprisonment attached to some provisions. Placed beside the EU AI Act’s turnover-based penalties, the figure reads as gentle; commentators have said outright that the ceiling carries no deterrence for mid-sized and large companies. And for at least a year past the effective date, enforcement runs guidance-first: officials will favor guidance, discounts on safety and reliability certifications, and expert consulting over punitive sanctions. Light fines, however, do not mean low standards — the real invoice for cross-border teams is the build-out cost of disclosure, risk management, and impact assessment.
Industry Worries, and a Checklist
The nonprofit Startup Alliance sounded a warning: the fuzzy high-impact definition threatens to land hardest on early-stage startups — should routine tools like machine translation or photo editing trip watermarking duties, compliance costs will chill experimentation, investment, and market entry. It pressed for obligations to be phased in by company size and deployment risk rather than raw compute. Criticism from the opposite flank argues the statute tilts toward industrial promotion over enforceable individual protections — civil servants deploying AI are shielded from personal liability absent intent or gross negligence — and questions enforcement while the implementing decrees remain under revision.
The working checklist for developers and product teams: inventory which features reach Korean users, check whether training compute crosses the 10^26 FLOP threshold, attach content labels to generative outputs, line up a domestic representative absent a Korean entity, and track the enforcement decree as it lands.