Bletchley to Seoul: The Rise of AI Safety Diplomacy

Bletchley to Seoul: The Rise of AI Safety Diplomacy

The development of AI governance entered a new phase in 2023.

Until then, much of the international conversation had focused on principles, ethics and responsible AI. The OECD AI Principles of 2019 and the UNESCO Recommendation on the Ethics of Artificial Intelligence in 2021 helped establish broad international expectations around trustworthy, human-centred and responsible AI.

The rapid emergence of generative AI changed the urgency of that conversation.

Large foundation models were becoming increasingly capable, widely accessible and difficult to govern through traditional sector-specific approaches. Governments began confronting questions that could not easily be addressed through national regulation alone.

What happens when an AI system is developed in one country, trained using globally distributed infrastructure, provided by a multinational company and deployed by users around the world?

How should governments understand risks associated with increasingly capable frontier models?

Who should evaluate those systems before deployment?

What information should companies disclose?

And how can governments cooperate when their domestic regulatory systems, economic interests and geopolitical priorities are different?

These questions helped create the conditions for a new form of international engagement: AI safety diplomacy.

Bletchley: Bringing Frontier AI Risk into International Diplomacy

The first major milestone was the AI Safety Summit held at Bletchley Park in the United Kingdom on 1–2 November 2023.

The summit was convened specifically around the risks associated with frontier AI and brought together governments, technology companies, researchers and civil society. Twenty-eight countries and the European Union were represented, including the United States, China, India, Japan, Brazil, Nigeria, Kenya, Singapore, South Korea and the United Arab Emirates.

The importance of Bletchley was not that it created a global AI regulator. It did not.

Its importance was diplomatic.

For the first time, a broad group of governments agreed on a shared international statement specifically focused on frontier AI safety.

The result was the Bletchley Declaration.

The declaration recognised both the opportunities and risks associated with AI and stated that the most capable AI systems could create serious, including potentially catastrophic, harms through deliberate misuse or unintended failures. It identified areas including cybersecurity, biotechnology and disinformation as particular concerns.

This represented an important change in the governance conversation.

AI safety was no longer being discussed only as an internal engineering problem for technology companies or as a national regulatory issue for individual governments.

It was becoming an international policy problem.

The central diplomatic insight was relatively simple: some AI risks do not stop at national borders.

A model developed in one jurisdiction can be accessed globally. A vulnerability discovered in one system can potentially affect users across multiple countries. Synthetic media can circulate internationally within minutes. Cyber capabilities can cross borders. AI development depends on globally connected research, computing, investment and supply chains.

International cooperation therefore becomes relevant even when countries disagree about many other aspects of AI policy.

From General Principles to Frontier AI

The Bletchley Declaration also helped introduce a more specific governance vocabulary around frontier AI.

Earlier AI principles had generally applied to AI systems across society. Bletchley placed particular attention on systems with unusually advanced capabilities and potentially significant risks.

This distinction matters.

A recommendation algorithm used to personalise entertainment and a highly capable general-purpose model capable of operating across software, research, communication or cybersecurity tasks may present very different governance questions.

The declaration therefore emphasised safety testing, evaluations, transparency and risk mitigation for organisations developing frontier AI capabilities. It also recognised a particularly strong responsibility for developers of the most powerful systems.

This was an important step toward a governance model in which technical evaluation becomes part of public policy.

AI governance could no longer remain entirely at the level of broad principles.

Governments increasingly needed to understand how capabilities were measured, how dangerous behaviours were tested, what safeguards were in place, how incidents were reported and whether companies’ own safety claims could be independently assessed.

The Governance Problem Becomes Technical

One of the most consequential developments at Bletchley was therefore not simply the declaration itself, but the emphasis on scientific and technical capacity.

The countries represented at the summit agreed to support the development of an international, independent and inclusive “State of the Science” report examining the capabilities and risks of frontier AI.

This signalled a broader transformation.

Effective AI governance increasingly requires technical knowledge.

Policymakers need to understand model capabilities. Regulators need access to evaluation methodologies. Governments need institutions capable of testing advanced systems. Researchers need ways to communicate uncertainty to policymakers. International organisations need mechanisms for comparing evidence across jurisdictions.

This creates a new institutional category: public-sector technical capacity for AI safety.

The emergence of government-backed AI safety institutions in several countries reinforced this trend. Bletchley discussions included cooperation around safety testing and the development of public capabilities for evaluating frontier systems.

This is a significant development in the history of AI governance.

Earlier governance debates often asked what principles governments should establish.

The frontier AI debate increasingly asks whether governments possess the scientific and technical capacity to verify whether those principles are actually being followed.

That is a much harder question.

From Bletchley to Seoul

The next major step came in South Korea.

The AI Seoul Summit took place in May 2024, building directly on the work initiated at Bletchley.

The Seoul Declaration expanded the international conversation around three interconnected objectives: safety, innovation and inclusivity. It also moved the discussion toward more practical international cooperation on AI safety science.

This was an important evolution.

Bletchley was primarily about establishing a shared recognition of the frontier AI safety problem.

Seoul began moving toward the question of how governments and institutions could cooperate in addressing it.

The Seoul Statement of Intent on International Cooperation on AI Safety Science called for greater coordination and collaboration in AI safety research. It also recognised the growing importance of public and government-backed institutions involved in AI safety research, testing and guidance.

In other words, the diplomatic conversation was beginning to connect directly with institutional infrastructure.

AI safety diplomacy was becoming less about declarations alone and more about scientific collaboration, evaluation capacity and implementation.

The Rise of AI Safety Institutes

One of the most important institutional developments surrounding Bletchley and Seoul was the growth of government-backed AI safety institutes.

These institutions emerged partly because governments recognised that effective oversight of advanced AI requires expertise that traditional regulatory agencies may not possess.

A conventional regulator may be capable of enforcing legal requirements. But frontier AI raises a different problem.

A regulator may need to understand model evaluations, red-teaming, dangerous capability assessments, robustness testing, model behaviour, security vulnerabilities and emerging risks.

That requires specialised technical institutions.

The Bletchley process helped elevate this model internationally, while the Seoul Statement explicitly recognised the importance of public and government-backed institutions working on AI safety research and testing.

This development has potentially significant consequences for the future of regulation.

Governments may increasingly rely on a combination of law, standards, technical evaluations and specialised institutions rather than regulation alone.

The regulator of the future may therefore need access to a technical evaluation ecosystem.

Safety Testing Becomes a Governance Function

Another important transition from Bletchley to Seoul was the growing importance of AI evaluations.

Traditional technology regulation often focuses on whether a product satisfies specified legal or technical requirements.

Frontier AI creates a more dynamic problem.

Capabilities can change between model generations. New capabilities may emerge unexpectedly. The same model can be used for different purposes. Risks may depend on deployment context. New vulnerabilities can be discovered after release.

This makes continuous evaluation increasingly important.

At Bletchley, governments and industry participants agreed to collaborate on testing future AI models against national-security, safety and societal risks.

The implication is significant.

AI governance is increasingly moving toward a model in which evidence about system behaviour becomes an important part of regulatory decision-making.

Instead of asking only whether a company has adopted responsible AI principles, policymakers may increasingly ask:

What was tested?

How was it tested?

Who performed the evaluation?

What capabilities were identified?

What risks were discovered?

What safeguards were implemented?

What changed after deployment?

These questions point toward a future in which AI assurance, evaluation and auditing become major components of governance.

International Cooperation Without a Global AI Regulator

It is important not to overstate what Bletchley and Seoul achieved.

Neither summit created a binding global AI treaty.

Neither established a universal international regulator.

The declarations and statements relied substantially on cooperation, shared commitments and voluntary or nationally implemented measures.

That limitation is important.

International AI governance operates in an environment where countries have different legal systems, economic interests, technological capabilities and national-security priorities.

The United States, China, the European Union, India, Japan, South Korea and other countries may share concerns about AI safety while pursuing very different domestic approaches.

International cooperation therefore does not necessarily mean regulatory uniformity.

A more realistic possibility is interoperability.

Countries may maintain different regulatory systems while developing compatible approaches to areas such as AI testing, risk measurement, incident reporting, technical standards and safety research.

This could become one of the most important governance questions of the next decade.

Can different AI governance systems recognise and work with each other’s evaluations, standards and assurance mechanisms?

If they can, international governance may become more practical without requiring a single global regulatory framework.

AI Safety Diplomacy in a Geopolitically Competitive Environment

There is another reason the Bletchley-to-Seoul transition matters.

AI safety diplomacy emerged at the same time as international competition over AI capabilities was intensifying.

The countries participating in these discussions did not share identical strategic interests.

The presence of both the United States and China at Bletchley was particularly significant because AI governance was developing alongside wider competition over advanced computing, semiconductors, research capabilities, industrial policy and national security.

This creates a difficult diplomatic environment.

Countries may want to cooperate on safety while competing over technological leadership.

They may want to share information about risks while protecting sensitive research and national-security capabilities.

They may support international standards while attempting to preserve domestic regulatory autonomy.

AI safety diplomacy therefore operates within a broader geopolitical tension between cooperation and competition.

That tension is unlikely to disappear.

The more strategically important AI becomes, the more difficult it may become to separate technical governance from questions of technological power.

The Global South Question

Another unresolved issue concerns representation.

The Bletchley summit included countries from Africa, Asia, the Middle East and Latin America, alongside major AI powers. The declaration itself emphasised inclusive international cooperation.

But participation in international dialogue does not automatically create equal governance capacity.

Countries differ substantially in access to advanced computing, AI researchers, evaluation infrastructure, regulatory institutions, investment and technical expertise.

This creates a structural problem.

A country may participate in international AI governance discussions while lacking the institutional capacity to evaluate whether advanced AI systems are safe for its own population.

For developing countries, therefore, AI safety cannot be reduced to frontier-model catastrophe scenarios.

It must also include practical questions about public-sector readiness, education, labour-market effects, misinformation, healthcare, agriculture, public services, data governance, cybersecurity and local-language systems.

International AI safety governance will have limited legitimacy if safety is defined only by the priorities of countries and institutions with the greatest technological capacity.

The Global South therefore needs not simply a seat at international meetings, but the scientific, institutional and financial capacity to participate meaningfully in governance.

What This Means for Bangladesh

The Bletchley-to-Seoul transition offers an important lesson for Bangladesh.

AI governance capacity cannot consist only of writing a national AI strategy or adopting broad ethical principles.

If advanced AI becomes increasingly governed through testing, evaluation, auditing, technical standards and institutional oversight, Bangladesh will need people and institutions capable of understanding those mechanisms.

That includes policymakers who understand AI systems, researchers who can conduct independent evaluations, universities that can develop AI governance and safety expertise, legal scholars who understand technology regulation, and public-sector institutions capable of translating technical evidence into policy.

This is particularly important for universities.

Universities are not simply training institutions for future AI developers. They can become part of the country’s governance infrastructure by producing researchers, policy analysts, technical evaluators and interdisciplinary experts.

For Bangladesh, one useful national research question is therefore not simply whether universities are teaching AI.

A more important question is whether universities are becoming ready to support the governance of AI.

That includes curriculum, research capacity, faculty expertise, AI policy centres, technical evaluation capability, ethics and governance programmes, institutional AI policies, data governance, responsible-use frameworks and collaboration with government and industry.

This is an area where substantial empirical research remains possible.

From Declarations to Verification

The most important lesson from Bletchley to Seoul may be that international AI governance is gradually moving from agreement about principles toward questions of verification.

The first stage was largely normative.

Governments agreed that AI should be trustworthy, responsible, human-centred and safe.

The next stage is operational.

Governments need to determine what “safe” actually means in practice.

What should be measured?

Which risks should be tested?

What evaluation methods should be trusted?

Who should perform the evaluation?

How independent should an evaluator be?

What evidence should companies disclose?

When should governments intervene?

How should incidents be reported internationally?

And how should evaluation methods evolve as AI capabilities change?

These are not purely philosophical questions.

They are institutional and technical questions.

This is why AI safety diplomacy represents a significant stage in the evolution of AI governance.

Bletchley helped establish a shared international recognition of frontier AI risks.

Seoul moved the conversation toward safety science, institutional cooperation, testing and implementation.

Together, they demonstrated that AI governance was becoming increasingly connected to scientific infrastructure.

The Next Phase

The movement from Bletchley to Seoul does not represent the end of AI safety diplomacy.

It represents its beginning.

The international system is still searching for effective ways to coordinate AI safety research, technical evaluations, incident reporting, standards and regulatory approaches.

At the same time, governments are developing domestic AI laws and institutions.

This creates an increasingly important relationship between international diplomacy and national regulation.

International principles can establish shared expectations.

National laws can create enforceable obligations.

Technical standards can translate broad requirements into operational practices.

Independent evaluations can generate evidence.

AI safety institutions can develop technical capacity.

And international cooperation can help these systems communicate with one another.

The future of AI governance may therefore not be built around a single global AI law.

It may emerge as a layered system combining domestic regulation, international agreements, technical standards, safety institutions, independent evaluation and continuous scientific assessment.

That would represent a major shift from the early AI ethics debate.

AI governance began with a relatively simple question:

What principles should guide the development and use of artificial intelligence?

The Bletchley-to-Seoul era introduced a harder question:

How can societies actually verify that increasingly capable AI systems are being developed and deployed safely?

That question will become increasingly important as AI systems become more capable, more autonomous and more deeply integrated into critical sectors of society.

The next phase of AI governance will therefore require not only principles and diplomacy, but institutions capable of implementation and verification.

This is where the global AI governance conversation begins to move toward regulation.

And that brings us to the next major development: the European Union’s attempt to translate risk-based AI governance into a comprehensive legal framework.