Korea vs Japan vs Taiwan AI Basic Acts

Only Korea's imposes concrete, checkable obligations right now — a risk assessment for “high-impact” or generative AI systems and a Korea-based local representative for foreign developers with no Korean office, both live since January 2026. Taiwan's Act is formally in force but its substantive content (MODA's risk-classification framework, sector agencies' own codes of practice under Clause 16) has not been published, so nothing concrete has attached to any developer yet. Japan's Act has been in force the longest (most provisions since June 2025) but creates no enforceable duty beyond a non-binding “duty to cooperate” — legally the weakest of the three despite being the oldest.

Japan's and Taiwan's explicitly do not — both statutes were written without a penalty provision, confirmed independently for each (Future of Privacy Forum for Japan; Tech Policy Press for Taiwan). Korea's position is genuinely unresolved in CASRAI's sourcing: the U.S. International Trade Administration's summary of the Act, which is the primary reference CASRAI's Korea guide relies on, does not state a fine amount or describe an enforcement penalty structure, and CASRAI has not verified one from a source it trusts enough to state as fact. Treat “Korea has no fines” and “Korea has fines” as equally unverified claims until a primary Korean-language source or its finalized subordinate regulations can be checked directly.

Because “binding” and “operative” are different questions. Taiwan's Basic Act on Artificial Intelligence is a validly enacted, promulgated statute with the full force of law — the same legal status Korea's and Japan's acts have. What it lacks is content: it is structured as a framework act that delegates the actual rulemaking (risk tiers to MODA, sector rules to individual agencies under Clause 16) to instruments that have not yet been written. That is a real and useful distinction from Japan's Act, which is equally in force but was never designed to create binding compliance content in the first place — Taiwan's gap is temporary and rulemaking-dependent; Japan's is structural.

All three, but for different reasons and on different timelines. MSIT (Korea) is the only one of the three currently administering live obligations, so it is the one with immediate compliance relevance. MODA (Taiwan) is the one to watch for what happens next — its unpublished risk-classification framework is the document that will turn Taiwan's Act from a statement of intent into an operative regime, and Clause 16's sector-agency rulemaking means the effective regulator for a given company may end up being its sector's own ministry, not MODA itself. Japan's AI Strategy Headquarters, chaired by the Prime Minister, is worth watching less for rulemaking and more for enforcement posture — the January 2026 Grok/X Corp case is the first real test of what “administrative guidance” does when a company does not cooperate quickly, and its outcome will say more about Japan's model than the statute's text does.

No. NIKOLAI is CASRAI's own independent, unendorsed frontier-AI-safety dictionary, and every crosswalk row in it is a shadow mapping — CASRAI's own interpretive reading, not an official record of any jurisdiction's law — unless an organization files its own Mapping Declaration and that declaration clears editorial review. The closest-fitting element, Coverage Scope Threshold on Track N1, documents exactly the kind of if-then test Korea's “high-impact” category and Taiwan's forthcoming MODA framework represent, in the same family as the EU AI Act's Article 51 threshold and California SB 53's compute trigger. Checked directly before writing this: that element's crosswalk table carries rows for the EU, California, the US federal government, and six AI developers' own frameworks — and none for Korea, Japan, or Taiwan. Korea's threshold is the strongest current candidate for a future row, because it already exists as text (even if the exact quantitative test isn't public); Taiwan's and Japan's are not yet mappable, because Taiwan's framework doesn't exist in published form and Japan's Act was never designed to create one.

No, and the difference is structural, not just geographic. The EU AI Act and SB 53 both set a quantitative compute or revenue threshold that triggers specific, enumerated obligations for covered systems — the test and the consequence are both in the statute. None of the three East Asian Basic Acts works that way. Korea comes closest, with a concrete risk-assessment duty already in force, but its “high-impact” threshold itself isn't public yet. Japan and Taiwan are further still: Japan never adopted a threshold-and-consequence model at all, and Taiwan's threshold is explicitly deferred to a future MODA framework. A compliance team used to the EU's or California's model should not assume any of the three East Asian acts works the same way by default.