International Organisation

From dialogue to delivery: making AI governance work in Southeast Asia

Disclaimer: The opinions expressed and arguments employed herein are solely those of the authors and do not necessarily reflect the official views of the OECD, the GPAI or their member countries.

Artificial intelligence (AI) is transforming how people learn, governments deliver services and economies grow. Governments in Southeast Asia are working to harness these opportunities while putting in place the governance needed to manage emerging risks.

Over the past year, the OECD, with support from Japan and the United Kingdom, has worked closely with Southeast Asian partner countries through a series of AI governance dialogues and workshops, including the OECD AI Policy Roundtable in Tokyo, Japan (2025), co‑creation workshops on the OECD AI Policy Toolkit in Bangkok, Thailand (2025 and 2026), and the OECD-Southeast Asia Policy Roundtable on Shipbuilding, Steel and AI in Tokyo (2026).

A clear message emerges from these exchanges: the real challenge is no longer defining principles but making them work in practice.

Moving from ideas to action

The OECD AI Policy Toolkit is a practical resource that helps governments map their AI policy landscape, identify priorities, and draw on concrete international experience to inform policy design and implementation. It is intended to support domestic policy efforts and peer learning across countries, rather than to rank countries or assess AI performance.

Against this backdrop, discussions held in Bangkok last Spring highlighted a strong interest in moving from principles to practice, with participants focusing on implementation questions: How can the Toolkit help governments identify priorities for policy development? Can it support coordination across ministries? How can it be adapted to different national contexts?

The feedback was clear: countries are not only looking to identify gaps, but also to prioritise actions and sequence implementation. Participants saw value in the Toolkit as a resource to support decision making, cross-government coordination and practical policy development.

Navigating uncertainty

AI technologies are evolving rapidly, requiring governments to make policy and governance decisions even as evidence on their impacts continues to emerge. Discussions around AI safety illustrate this challenge. The International AI Safety Report 2026 highlights both rapid progress in AI capabilities and risks that are already visible in areas such fraud, synthetic media and cybersecurity. At the same time, safeguards do not always behave as expected once systems are used outside controlled environments.

Countries are responding in practical ways. Singapore has taken an approach that builds on existing legislation, complemented by voluntary guidance and assurance tools such as AI Verify, alongside ongoing testing. Rather than relying on fixed rules, the focus is on iterative updating of approaches as more is learned.

The Hiroshima AI Process Reporting Framework: all hands on deck

Effective AI governance requires the active involvement of AI stakeholders beyond government, particularly developers and deployers of AI systems, given their respective roles and responsibilities. The Hiroshima AI Process (HAIP) Reporting Framework which invites organisations developing and deploying advanced AI systems to report on their risk management practices in line the Hiroshima AI Process International Code of Conduct for Organisations Developing Advanced AI Systems is emerging as a reference in this context.

By helping to operationalise internationally agreed AI principles and fostering consistency and comparable voluntary disclosures on AI risk management practices, the HAIP Reporting Framework aims to strengthen trust through greater transparency  while generating valuable learning that can help address gaps in methodology, coverage, and incentives. Going forward, three issues will be paramount: expanding participation, keeping pace with rapid technological change, and ensuring interoperability across emerging AI governance and reporting approaches to promote coherence and reduce duplication.

When language and culture become key policy issues for AI

Language- and culture-related relevance proved one of the most tangible challenges discussed.

AI systems still perform unevenly across languages and this gap matters in practice. Mistranslations, loss of context or weaker safety filters in non‑English prompts can affect how systems behave. AI systems that do not work reliably in local contexts will impact adoption given struggles to earn users’ trust as well as the value delivered to the region’s economies and societies.

This is prompting a shift, not only towards building new models, but also towards more targeted approaches: adapting systems, improving datasets and strengthening testing and evaluation. Examples from around the region, such as SEA-LION.AI, a family of open multilingual models developed by AI Singapore and adapted to Southeast Asian languages and contexts; MERaLiON, a speech-centric multilingual and multimodal model developed by A*STAR’s Institute for Infocomm Research (I²R); and Indonesia’s Sahabat-AI, an open LLM ecosystem designed to better support Bahasa Indonesia and regional languages such as Javanese and Sundanese. Initiatives led by Japan are showing the value of sharing data, computing resources and models for multilingual and multicultural LLM development – which considers multimodal and cultural content (e.g. images, video, sign language, narrative structures) alongside text.

At the same time, important constraints remain. High-quality training and evaluation data for many local languages are often limited, making it more difficult to develop and assess AI systems that perform reliably in local contexts. Existing benchmarks also tend to focus on English-language tasks,  meaning that strong performance on widely used tests may not accurately reflect how models perform in other languages. Finally, since many countries have limited influence over the design of large global models, the ability to test and adapt them for local needs becomes particularly important. Singapore’s emphasis on multilingual testing and red-teaming reflects this focus on ensuring that AI systems perform reliably and safely in local settings.

AI compute capacity as a governance lever

Behind these issues sits a fundamental constraint: AI compute capacity.

Discussions in Bangkok framed AI compute not just as infrastructure, but as a broader ecosystem, encompassing skills, energy and access to cloud services. This has direct implications for what countries can realistically do, whether developing AI systems, adapting them to local needs or overseeing their use.

Thailand, for example, described an approach that links investments in AI compute with research, public‑sector use and skills development, seeking to create mutually reinforcing capabilities across the ecosystem. Vietnam outlined a more gradual pathway, starting with model training and evaluation before moving to more resource‑intensive applications.

These examples illustrate how national AI ambitions are shaped by the compute resources and capabilities available in practice. Countries with limited influence over global model design may place greater emphasis on local capabilities for testing, evaluation and adaptation. As governments increasingly rely on technical assessments to understand the capabilities, limitations and risks of AI systems, access to compute becomes relevant not only for innovation but also for governance, helping countries generate evidence, scrutinise AI systems and make informed decisions about their deployment and use.  

What it looks like on the ground

Two sessions, on education and healthcare, brought many of the above-mentioned threads together.

  • On education, the OECD presented evidence from the OECD Digital Education Outlook 2026. AI can support learning when it is used with a clear purpose, for example, to encourage collaboration or reflection, but can lead to weaker outcomes when used in the absence of appropriate pedagogical frameworks. In vocational education, AI is already being used to inform curriculum design and skills mapping, for example by helping identify emerging skills needs and analysing labour-market information. This raises familiar questions about data quality and accountability, particularly when AI-generated insights are used to inform education and training decisions.
  • On healthcare, Thailand presented Thai School Lunch, using AI to support nutrition planning in public schools through transparent decision criteria, and KidDiary, which integrates child development, health and nutrition data to help identify potential risks through staged alerts and human oversight. Timor‑Leste, in turn, shared plans to use AI to strengthen disease surveillance, while highlighting the practical challenges posed by limited infrastructure and language resources and the initial steps being taken to address them.

Across these examples, outcomes are shaped less by the technology itself than by how it is embedded, monitored and combined with existing institutional arrangements, governance processes and service delivery systems over time.

From dialogue to sustained cooperation

OECD-Southeast Asia AI engagement is gaining momentum. Building on the inaugural OECD-Southeast Asia AI Dialogue workshop held in September 2026, countries reconvened at a second workshop, Harnessing Responsible AI across Southeast Asia, held in November 2026 as part of the Second OECD-Southeast Asia Policy Roundtable on Economic Security, organised by the Government of Japan and the OECD Tokyo Centre. Under the Global Partnership on AI (GPAI), with support from Japan, the dialogue is evolving into a regular platform for exchange through online sessions and annual in-person meetings, with the next gathering planned in Indonesia. These discussions will be complemented by implementation support, backed by the United Kingdom and drawing on the OECD AI Policy Toolkit and other resources available through OECD.AI.  

The aim is to continue comparing approaches, sharing experience and deepening regional cooperation around common challenges, from safety and risk management to measures reflecting language and cultural specificities, compute and infrastructure needs and sectoral uses.

A shared direction

The discussions held in Bangkok suggested a clear direction. Countries are working from a common set of principles while adapting them to different contexts, constraints and priorities.

What matters now is how those principles are translated into practical tools, processes and institutional choices, and how countries continue to learn from one another through implementation.

That process is already underway.

Follow us on LinkedIn
Receive the OECD's artificial intelligence newsletter!
Sign up with Linkedin

Disclaimer: The Organisation cannot be held responsible for possible violations of copyright resulting from the posting of any written material on this website/blog.