For the Future of AI Governance, Look to Asia

This year, the world’s two regulatory superpowers—Europe and the United States—each stepped back from advancing binding rules on advanced artificial intelligence (AI). In late July, the EU’s Digital Omnibus on AI entered into force. The general-purpose AI rules survived intact, but the Act’s centerpiece high-risk obligations for standalone systems were deferred from August 2026 to December 2027, a delay driven by the unreadiness of the standards and tools needed to assure compliance. In June, the White House issued an executive order on frontier model security that works through voluntary engagement with developers rather than binding obligations, following a December order aimed at limiting conflicting state AI rules.

Reasonable people disagree about the wisdom of each choice. The practical consequence, however, is striking: for at least the next eighteen months, the most consequential decisions about AI governance will be made not in Washington or Brussels but in Seoul and Tokyo.

These four capitals are not the only players. California’s frontier AI statute has been in force since January, with incident reporting and whistleblower protections operating now, notwithstanding the federal push against state rules. The UK’s AI Security Institute (AISI) leads the science of model evaluation, and Beijing enforces the world’s most extensive rules on AI-generated content.

The largest forces acting on AI governance are still centered in Washington and Beijing. Their choices about open and closed model release will shape the trajectory of AI more than any statute does. Whether those weights—the trained parameters that constitute the model itself—are published for anyone to download, or kept in-house and reachable only through the developer’s own servers, determines which safety protocols can be enforced at all, and against whom. A closed model leaves one company to regulate. A published one can be copied and stripped of its safety training by anyone, with no way to recall it. Neither capital has settled that question. What neither is producing, however, is a framework other governments can adopt: Washington by design, and Beijing because its rules are built for a market and a political system that others will not copy.

None of the major AI players has assembled what Seoul and Tokyo now have in place: national AI frameworks (binding and horizontal in Korea, promotional and administrative in Japan) where statute, implementing rules, and new institutions are all moving at once.

In January, South Korea’s Framework Act on Artificial Intelligence entered into force, making Korea one of the first countries anywhere to operate a comprehensive AI law that offers a cross-sectoral, horizontal statutory framework. Admittedly the statute is a frame, not a full picture. Its enforcement decree took effect alongside the act, fixing the main parameters. The operative substance is being drawn now, in the ministry notices, safety-assurance guidelines, and supervisory practice that Korea’s Ministry of Science and ICT is developing through 2026 and into 2027. Nonetheless, the statute addresses how high-impact AI classification works in practice, what safety evaluation actually requires, and what operational mandate the recently established Korea AI Safety Institute will hold.

Japan is moving along a different track toward the same destination. Its AI Promotion Act of May 2025 and the National AI Basic Plan adopted by the Cabinet in December 2025 reflect a distinctive governance tradition. The statute sets national objectives and creates a prime minister-led AI Strategy Headquarters rather than imposing obligations on developers, and the operative detail arrives through ministry guidance rather than enforceable rules. The Basic Plan commits the government to expanding the AI Safety Institute Japan, which now conducts frontier model evaluations.

Why should anyone outside Northeast Asia care? Because these two countries sit at chokepoints of the physical infrastructure on which the entire AI stack depends. Korean firms produce the overwhelming majority of the high-bandwidth memory that frontier AI training runs require. Korea is also building sovereign foundation models: frontier-scale systems trained and controlled domestically, in Korean, rather than licensed from developers abroad.

Japanese firms dominate semiconductor-grade silicon wafers and the inspection equipment on which advanced lithography depends. Japan also leads the world in the manufacturing of the industrial robotics through which AI will increasingly act in the physical world. Standards set at supply chain chokepoints carry weight far beyond their borders.

Korea’s reach is legal as well as economic: its new law applies to foreign AI systems that affect the Korean market and requires major foreign developers to appoint domestic agents. The rules being written in Seoul apply, in practice, to every frontier developer on earth, because all of them serve Korean users. Japan’s expectations travel through the administrative guidance attached to one of the world’s largest technology markets.

To be clear, Seoul and Tokyo should not be treated as a bloc. It is tempting, especially from a distance, to speak of a single “Asian model” or to prescribe a joint framework. The reality is more interesting and more useful. Korea has chosen comprehensive statute law with implementation coming through decrees. Japan has chosen promotion legislation with institutional and administrative depth. Their industrial bases differ, their political economies of AI differ, and each manages its own distinct set of strategic relationships. These are two separate, sovereign experiments in governing frontier technology, run by two of the world’s most capable administrative states. Where their approaches converge, the convergence is more meaningful precisely because it is chosen rather than imposed. For example, both countries’ safety institutes participate in the international AISI network and Korea’s decrees match California’s on some important dimensions.

What is being decided in these two capitals is unglamorous yet important. The laws address incident reporting: how governments learn that something has gone wrong inside a frontier system or about the organization deploying it in time to act. Both statutes include valuation methodology: how you rigorously test models that operate in Korean and Japanese, when nearly all the world’s evaluation science has been built in English. Both focus on institutional capacity: whether new safety institutes acquire the technical staff, legal authority, and industry access to hold whatever standards are written. These are the load-bearing walls of AI governance. They are being framed now and once set they will be expensive to move.

For the American research and policy community, three implications follow. First, the center of gravity of implementable AI governance has shifted, at least temporarily, toward Asia. Researchers who want their safety work to shape practice in 2026 and 2027 should understand that the decree drafters in Seoul and the planners in Tokyo are among their most consequential audiences. Second, engagement now involves translation, not export. Frameworks that land in Seoul or Tokyo and read like someone else’s regulatory model, however technically sound, will struggle against strong and entirely legitimate sovereign technology agendas. Both governments have staked national strategy on building domestic AI capability: Korea on sovereign models and memory manufacturing, Japan on physical AI and semiconductors. A safety proposal that appears as an external constraint undercuts that strategy. The same proposal, framed as what makes Korean and Japanese AI exportable and trusted, buttresses it. Arguments grounded in export credibility, supply chain trust, and national competitiveness are not spin; they are the genuine terms on which safety and prosperity align in both countries. Third, the window is short. The institutional architecture being built will likely set the terms of AI governance in both countries for years, with important implications for other regulators.

The United States will return to this table. Washington’s posture toward frontier AI has shifted before and will shift again. The EU’s clock resumes in December 2027. When the regulatory superpowers re-engage, they may find that the working models of practical AI governance, the ones with operating experience behind them, were built in Asia. The wiser course is not to wait and find out, but to engage now, seriously and respectfully, with the two democracies doing the work.

Henry Haggard is senior advisor at ADEN, the AI Diplomatic Engagement Network, and served for 25 years in the U.S. Department of State, including as Minister Counselor for Political Affairs at the U.S. Embassy in Seoul. Chris McKinney is Chief Operating Officer at ADEN and a former senior U.S. diplomat.

Thumbnail credit: Pexels