Two AI models in Asia-Pacific: Korea legislates, Singapore convenes

Europe is not the only place where AI rules became operational in 2026. On a single day — 22 January 2026 — two of Asia-Pacific's most watched jurisdictions put their approaches into effect, and they could hardly be more different. South Korea switched on a binding, risk-based statute; Singapore published a voluntary framework for the newest problem on everyone's desk: AI agents that act on their own. Together they map the two poles between which much of the region — and much of the world — is now choosing. For any organisation that sells beyond the EU, this layer sits alongside the international governance frameworks and the patchwork of US state laws.

Korea legislates: a binding statute with extraterritorial reach

South Korea took the legislative route. Its AI Basic Act and Enforcement Decree took effect on 22 January 2026, making the country the first in Asia, and the second jurisdiction after the EU, with a comprehensive, horizontally binding AI law: risk-based duties for high-impact and generative AI, transparency and human-oversight obligations, and a domestic-representative rule that reaches foreign providers above defined revenue and user thresholds — with administrative fines deferred during a grace period through 2026. Because that is a statute and not guidance, we treat it in full in a dedicated analysis: South Korea's AI Basic Act. What matters for the comparison here is the model: a law, with tiers, hooks and penalties on a timer.

Singapore convenes: governing AI that acts, by guidance

Singapore took the opposite route — and aimed at a newer target. On the same 22 January 2026, at the World Economic Forum in Davos, the Infocomm Media Development Authority (IMDA) launched the Model AI Governance Framework for Agentic AI: guidance for systems that do not merely generate output but plan, reason and take actions on a user's behalf. It is presented as the first governance framework dedicated to agentic AI, and it builds on Singapore's original Model AI Governance Framework from 2020 as a living document.

The framework is structured around four concerns: bounding and assessing the risks of an agent before deployment; keeping humans meaningfully accountable for what the agent does; building technical controls and processes around the agent's autonomy; and enabling end-user responsibility. The thread running through all four is the gap that makes agents different from ordinary models — an agent can chain actions, call tools and commit a user to consequences without a human in the loop for each step — so the framework concentrates on delegation boundaries, override points and clear accountability when something goes wrong.

Crucially, compliance is voluntary — but the framework is explicit that voluntariness does not dissolve liability: organisations remain legally accountable for their agents' behaviour and actions under existing law. The bet is that fast, practical guidance, iterated as the technology moves, will steer behaviour faster than a statute that takes years to pass and longer to amend.

Two models, one direction of travel

Read together, the contrast is the point. Korea legislates: a binding statute, risk tiers, extraterritorial hooks, penalties on a timer. Singapore convenes: voluntary guidance, fast iteration, accountability located in existing law rather than a new one. Neither is a copy of the AI Act, yet both borrow its grammar — risk, transparency, human oversight — confirming that the EU's vocabulary has become the regional lingua franca even where the enforcement philosophy diverges. Singapore's move also shows where the frontier is heading: the agentic question — who answers for what an autonomous system does — is the one the EU's own Article 50 transparency layer does not yet reach.

For a European organisation the practical lesson is concrete. If you serve Korean users at scale, the domestic-representative threshold may already bind you, grace period or not. If you deploy agentic AI in Singapore, the Model Framework is the yardstick a regulator and a counterparty will reach for first. The compliance map no longer stops at the EU's border.