Paul K.H. Kim ·
On 22 January 2026, Korea’s AI Basic Act took effect, together with its detailed implementing decree. Korea calls it the world’s first comprehensive AI law to come fully into force. Taiwan’s own AI Basic Act took effect eight days earlier, but it sets out principles with no penalties. Korea’s has fines. They are on hold until early 2027.
What the law asks of companies
- Tell people when a service uses generative or “high-impact” AI. High-impact is Korea’s term for AI that can significantly affect lives, safety or rights.
- Label AI-generated output, especially realistic deepfakes.
- Manage the risks of high-impact AI and of the most powerful AI systems.
- Appoint a local representative: a person or firm in Korea who answers for a foreign company’s AI service. This applies to foreign firms above certain sizes: revenue of at least ₩1 trillion, AI revenue of at least ₩10 billion, or at least a million daily users in Korea.
The fines, and the clock
The fines are small: at most about $21,000 (₩30 million). They are suspended for a grace period of “at least one year”, which ends around 22 January 2027 unless the government extends it. By early November 2026 we found no announcement extending it.
So far, the reach is narrow
In January 2026 the science ministry said that, for now, “high-impact” AI covered essentially only highly automated self-driving cars, and that no model in use was powerful enough to face the extra safety duties for the most advanced systems. Those duties apply to models trained with at least 10²⁶ operations, and no Korean model is yet that large.
An amendment that took effect on 21 July 2026 clarified the labelling duty and added a reward: an AI certification that gives firms an advantage in public procurement. Companies say it is still unclear whether text an AI helped write, and a person then edited, must be labelled.
Where the law meets the triangle
The local-representative rule is the one foreign companies feel. Five months after the law took effect, only one foreign AI company, Anthropic, had completed the filing with the ministry. Google’s application had been sent back for missing information, and OpenAI, Meta and Microsoft had not finished the procedure. In August, Anthropic was still the only one. A ruling-party lawmaker has proposed requiring big foreign firms to name their own Korean subsidiary rather than a law firm or agency.
Washington is watching. In April 2026, the US Trade Representative said it was seeking “outcomes” on digital regulation with Korea, the EU and Australia.
What we expect in 2027
Start-ups say the rules are vague and want more time. Civic groups say the law is too weak. With small fines, a narrow definition of high-impact AI and American trade pressure, the law’s real force in 2027 may come less from penalties than from two simple duties: tell people when they are dealing with AI, and have someone in Korea who answers for it.
That is why one of the book’s ten predictions for 2027 is that no company will be fined under the Act during 2027 (65% confidence). The signpost to watch is the ministry’s decision in January 2027 on whether to extend the grace period. If it lets the period lapse and publishes enforcement guidelines, lower your confidence. The easiest breach to prove would be a large foreign AI service that still hasn’t named a local representative.
What to do now if you run an AI service in Korea: check whether you meet the local-representative thresholds; add clear notices where users meet generative AI; label realistic synthetic images, audio and video; and keep a record of how you judged whether any of your systems are “high-impact”.
This article is general information, not legal advice.
Sources: US International Trade Administration, South Korea AI Basic Act; Financial News, 21 January 2026; MTN, 25 June 2026; Korea Times, 23 April 2026. The book has a note for every figure.