South Korea now has one of the world's first comprehensive AI statutes on the books, and it is choosing not to enforce the teeth of it for a year. That choice — deliberate, public, and paired with a mechanism to rewrite the rules before they bite — is the most interesting thing happening in AI governance right now, and it deserves more attention than the law's passage got.
What Actually Took Effect
The Act on the Development of Artificial Intelligence and Establishment of Trust — the AI Basic Act — took effect on January 22, 2026, the Ministry of Science and ICT (MSIT) confirmed. It builds three pillars: development and trust infrastructure (including a National AI Committee and an AI Safety Institute), industry support for compute and data, and oversight obligations for generative and "high-impact" AI.
The obligations themselves are real but modest. "High-impact AI" is defined as systems that significantly affect human life, safety, or fundamental rights — healthcare, energy, transportation, hiring, and biometrics — and triggers transparency, safety, and risk-management duties. Penalties top out at 30 million KRW (about US$21,000) for failures like not disclosing AI use or not appointing a domestic representative, per a Cooley analysis of the statute. By the standard of the EU AI Act's revenue-percentage fines, these are nominal.
The Grace Period Is the Policy
The headline is what MSIT won't do. The ministry has committed to a one-year grace period in 2026, during which, in its words, "fact-finding investigations and administrative fines will generally be deferred except in exceptional cases involving serious social harm, such as loss of life or human-rights violations," according to the U.S. government's market-intelligence brief. MSIT explicitly frames its posture as "minimum regulation."
Crucially, the grace period is not a passive pause. Seoul has stood up an AI Basic Act Institutional Improvement Task Force — a public-private body drawing on more than 40 experts from industry, academia, and civil society — to find gaps during the year and translate them into concrete amendments. The still-fuzzy line around what counts as "high-impact" AI, and the compute threshold for large-scale systems (reportedly around 10²⁶ FLOPs), are exactly the kinds of definitions the task force is meant to sharpen with operational evidence rather than guesswork.
This is regulation as iteration. The law sets the direction; the enforcement calendar buys time to calibrate the specifics against how the technology and the market actually behave.
Steelmanning the Critics
The case against this approach is not frivolous. A statute with deferred penalties and undefined thresholds offers firms little certainty about what compliance even looks like, and a grace period can read as a green light to deploy now and worry later — precisely when an unsafe hiring or medical model could cause real harm before any corrective order lands. Korean civil-society groups have already warned that the draft enforcement decree leaned too far toward industry and underweighted human-rights risk. Those are legitimate concerns, and "we'll fix it as we go" is only credible if the fixing actually happens.
But the alternative — locking in prescriptive rules before anyone understands the technology's risk surface — has a worse track record. The EU's AI Act is already being revisited and partially delayed under industry and member-state pressure, a sign that writing detailed obligations ahead of evidence produces rules that need patching anyway, just with more disruption in between. South Korea has front-loaded the patching into a structured feedback loop instead of pretending it got the definitions right on the first try.
Why the Restraint Fits Korea
The approach also tracks an unusual democratic mandate. Seventy percent of South Koreans say advancing science and medicine through AI is a bigger priority than protecting industries through regulation, MIT Technology Review reported on June 15, 2026, citing the Stanford AI Index. That is not a public clamoring to be shielded from AI; it is a public asking to be allowed to build with it. The same Stanford data ranks Korea third globally in the number of notable AI models — a genuine frontier position, not an aspirational one.
The spending matches the sentiment. President Lee Jae-myung, who took office in 2025 pledging to make Korea a "top three AI power" alongside the U.S. and China, more than tripled the country's AI investment to roughly 10 trillion won (about $6.7 billion) in the 2026 budget, approved by parliament in December 2025. A government betting that heavily on building AI has every incentive not to strangle it with premature enforcement — and, equally, every incentive to get the safety guardrails credible enough that the bet pays off without a backlash.
The Test Ahead
The risk in Seoul's model is execution, not philosophy. A grace period only works if the task force produces sharper definitions, narrower high-impact categories, and clear compliance pathways before enforcement begins in 2027 — and if the "serious social harm" carve-out is actually used when a deployment crosses the line. If the year passes with vague rules simply hardening into vague enforcement, the critics will have been right.
For now, though, South Korea is doing something most jurisdictions only claim to: regulating AI proportionately, in the open, with a stated willingness to revise. A comprehensive law that defers its own fines to learn how the technology behaves is not weakness. On the current evidence, it is the most defensible bet on the board.