OECD urges Southeast Asia to turn AI principles into practical governance

Southeast Asian governments are shifting their AI governance focus from defining principles to putting them into practice, according to an OECD.AI article. Discussions supported by Japan and the United Kingdom highlight implementation, local relevance, safety reporting and access to computing resources as central challenges.
- Southeast Asian governments are prioritising implementation over principle-setting.
- The OECD AI Policy Toolkit supports policy mapping, priority-setting and coordination.
- Local-language testing, data and evaluation remain major AI governance needs.
- Compute capacity affects both AI development and governments' ability to oversee systems.
- Thailand and Timor-Leste presented public-sector AI applications and plans.
From principles to implementation
The OECD, with support from Japan and the United Kingdom, has worked with Southeast Asian partner countries through AI governance dialogues and workshops over the past year. These included the OECD AI Policy Roundtable in Tokyo in 2025, co-creation workshops on the OECD AI Policy Toolkit in Bangkok in 2025 and 2026, and the OECD-Southeast Asia Policy Roundtable on Shipbuilding, Steel and AI in Tokyo in 2026.
The discussions found that governments are looking for practical ways to apply AI policy, rather than simply define broad principles. The OECD AI Policy Toolkit is designed to help governments map their policy landscape, identify priorities and use international experience in policy design and implementation. The article says it is intended to support domestic policy efforts and peer learning, not to rank countries or assess AI performance.
Participants in Bangkok focused on how the Toolkit could support priority-setting, coordination between ministries and adaptation to national contexts. Their feedback pointed to a need to identify gaps, sequence actions and support decision-making and cross-government coordination.
Safety, transparency and local-language performance
Governments are making decisions while evidence about AI's effects continues to develop. The International AI Safety Report 2026 highlights rapid progress in AI capabilities, along with visible risks involving fraud, synthetic media and cybersecurity. The OECD.AI article also says safeguards can behave differently when systems are used outside controlled environments.
Singapore is described as combining existing legislation with voluntary guidance and assurance tools such as AI Verify, supported by ongoing testing. Its approach emphasises updating policies as more is learned rather than relying only on fixed rules.
The Hiroshima AI Process Reporting Framework is emerging as a reference for organisations developing and deploying advanced AI systems. It encourages voluntary reporting on risk-management practices in line with the Hiroshima AI Process International Code of Conduct for Organisations Developing Advanced AI Systems. According to the article, wider participation, adaptation to technological change and interoperability with other reporting approaches will be important.
Language and culture are also practical governance issues. Uneven performance across languages can produce mistranslations, lost context and weaker safety filters in non-English prompts. The article highlights SEA-LION.AI, MERaLiON and Indonesia's Sahabat-AI as examples of efforts focused on Southeast Asian languages and contexts. It also points to limited local-language data and English-focused benchmarks as continuing constraints.
Compute capacity and real-world applications
The discussions treated AI compute capacity as more than infrastructure. It includes skills, energy and access to cloud services, all of which affect whether countries can develop systems, adapt them to local needs or oversee their use. Thailand described linking compute investment with research, public-sector use and skills development. Vietnam outlined a more gradual route, beginning with model training and evaluation before pursuing more resource-intensive applications.
This makes compute relevant to governance as well as innovation. Governments need technical assessments to understand AI systems' capabilities, limitations and risks, while access to computing resources can help them generate evidence, scrutinise systems and make informed deployment decisions.
Examples from education and healthcare show how these issues appear in public services. The OECD Digital Education Outlook 2026 says AI can support learning when used for a clear purpose, such as encouraging collaboration or reflection, but may produce weaker outcomes without suitable pedagogical frameworks. Thailand presented Thai School Lunch, which uses transparent decision criteria for nutrition planning in public schools, and KidDiary, which combines child development, health and nutrition data with staged alerts and human oversight. Timor-Leste shared plans to use AI to strengthen disease surveillance.

