2023: The Hiroshima AI Process and the Rise of International AI Diplomacy

2023: The Hiroshima AI Process and the Rise of International AI Diplomacy

By 2023, artificial intelligence had entered a new phase.

The debate was no longer primarily about whether AI would become important.

It already was.

The emergence of generative AI had demonstrated that increasingly capable AI systems could reach millions of people rapidly and become integrated into education, business, software development, research, media and public life.

But the technology was developing faster than international governance mechanisms.

Different countries were developing different policies.

Companies were releasing increasingly capable models.

Researchers were developing new evaluation techniques.

Civil society organisations were raising concerns about misinformation, discrimination, privacy and human rights.

At the same time, a new category of concern was becoming more prominent: the risks associated with highly capable and general-purpose AI systems.

This created a difficult international policy problem.

AI systems operate across borders.

AI companies operate across borders.

Data flows across borders.

AI infrastructure and supply chains cross borders.

The effects of AI-generated information can cross borders almost instantly.

Yet most regulation is still national or regional.

In 2023, the G7 attempted to address part of this problem through the Hiroshima AI Process.

THE BEGINNING OF THE HIROSHIMA AI PROCESS

The Hiroshima AI Process was launched in May 2023 following the G7 Hiroshima Summit under Japan’s presidency.

The initiative emerged at precisely the moment when generative AI was transforming the policy landscape.

G7 leaders recognised both the transformative potential of advanced AI systems and the need to manage their risks.

The process was designed to examine international governance issues related particularly to generative AI and advanced AI systems.

This was significant for a simple reason.

International AI governance was moving from broad principles toward coordination around increasingly capable technologies.

The OECD Principles had provided a common policy vocabulary.

UNESCO had expanded the ethical and human-rights dimension.

The Hiroshima Process attempted to build a more practical international framework around advanced AI.

THE INTERNATIONAL DIMENSION

The G7 was not creating a global AI government.

It was attempting to coordinate major economies around common expectations.

This distinction matters.

The G7 does not represent every country.

It does not possess universal legislative authority.

Its statements do not automatically become binding international law.

But G7 members include several of the world’s major AI developers, technology markets, research ecosystems and regulatory powers.

Coordination among these countries could therefore influence how advanced AI systems are developed and deployed internationally.

The Hiroshima Process was consequently both a governance initiative and a diplomatic initiative.

AI policy was becoming AI diplomacy.

WHY GENERATIVE AI CHANGED THE DIPLOMATIC AGENDA

Generative AI created several governance challenges that were difficult to contain within national borders.

A model developed in one jurisdiction could be accessed by users around the world.

An application built in another jurisdiction could rely on that model.

Synthetic content could be distributed internationally.

AI-generated misinformation could affect societies far beyond the country where it was created.

Software generated by AI could be used globally.

Cyber-related capabilities could cross borders.

This meant that domestic regulation alone could not address every risk.

International cooperation therefore became increasingly important.

The Hiroshima Process represented one of the early attempts to build such cooperation around the new generation of generative and advanced AI systems.

THE FOUR QUESTIONS OF INTERNATIONAL AI GOVERNANCE

The emergence of the Hiroshima Process reflected four increasingly important questions.

First, what principles should organisations developing advanced AI follow?

Second, what information should they provide about their systems and risks?

Third, how should governments and companies cooperate on AI safety?

Fourth, how can different national governance systems remain interoperable?

These questions would become increasingly central to international AI governance.

FROM PRINCIPLES TO ORGANISATIONAL RESPONSIBILITY

One of the most important features of the Hiroshima Process was its focus on organisations developing advanced AI systems.

This represented an important evolution.

Earlier AI principles often described desirable characteristics of AI.

The Hiroshima Process increasingly translated those principles into expectations for organisations.

The focus shifted toward what developers should actually do.

This included responsibilities around risk identification, assessment, mitigation, transparency and information sharing.

The underlying idea was straightforward.

If increasingly capable AI creates new risks, organisations developing these systems should have concrete responsibilities for identifying and managing those risks.

This helped move AI governance toward lifecycle risk management.

THE INTERNATIONAL GUIDING PRINCIPLES

In October 2023, G7 leaders welcomed the Hiroshima Process International Guiding Principles for Organisations Developing Advanced AI Systems.

The principles addressed areas including appropriate risk identification and mitigation, responsible release, information sharing, security, transparency and accountability.

The framework was designed for organisations developing advanced AI systems.

This was significant because it recognised that governance needs to reach upstream.

If a foundation model is later integrated into thousands of applications, governance cannot begin only at the final application layer.

Some responsibilities need to exist at the model-development stage.

This is an increasingly important principle in modern AI governance.

THE CODE OF CONDUCT

The Hiroshima Process also produced an International Code of Conduct for Organisations Developing Advanced AI Systems.

The Code of Conduct translated broad expectations into a more operational form.

It encouraged organisations to take concrete actions to identify and manage risks associated with advanced AI systems.

This included expectations around risk assessment, testing, transparency, incident reporting and responsible development.

The Code was voluntary.

That limitation is important.

It was not equivalent to a binding international treaty.

It did not create a global enforcement authority.

Its influence depended partly on adoption by organisations and governments.

But voluntary frameworks can still matter.

They can establish expectations.

They can create common reporting practices.

They can support international interoperability.

They can provide a baseline for future regulation.

And they can help governments develop more concrete governance requirements.

FROM ETHICAL PRINCIPLES TO DUE DILIGENCE

The Hiroshima Process therefore represented another conceptual transition.

The language of AI ethics had often focused on values.

The language of the Hiroshima Process increasingly introduced organisational due diligence.

This means asking organisations to identify risks, assess them, take measures to mitigate them and provide information about how those risks are managed.

This is closer to established approaches in other areas of governance.

Companies are already expected in many contexts to manage risks related to cybersecurity, privacy, financial integrity and corporate responsibility.

AI governance was increasingly moving in the same direction.

Responsible AI was becoming a management process.

THE IMPORTANCE OF RISK IDENTIFICATION

Risk identification became particularly important because advanced AI systems can create multiple categories of risk.

Some risks arise from ordinary system failures.

Others arise from misuse.

Some emerge only after deployment.

Others may appear when models are integrated into new applications.

This means that risk management cannot be a one-time activity.

An organisation may evaluate a model before release and still discover new problems later.

The model may be updated.

Its users may discover new capabilities.

Its deployment environment may change.

New attack techniques may emerge.

New downstream applications may create unexpected risks.

Governance therefore needs to continue after release.

This helped strengthen the idea of continuous AI risk management.

THE IMPORTANCE OF RED-TEAMING AND TESTING

The Hiroshima framework also reflected the growing importance of testing.

Advanced AI systems increasingly require deliberate attempts to identify weaknesses before and after deployment.

This can involve red-teaming, adversarial testing and other forms of evaluation.

The underlying governance principle is important.

Organisations should not simply assume that their systems are safe because they performed well during development.

They should actively search for ways the system could fail or be misused.

This represents a significant change in the culture of AI governance.

Safety becomes an empirical activity.

THE TRANSPARENCY QUESTION

The Hiroshima Process also emphasised transparency.

But transparency in advanced AI is difficult.

AI models can contain billions or even trillions of learned parameters.

Their training processes can involve enormous datasets.

Their behaviour can vary across contexts.

A simple technical description may not be sufficient for meaningful accountability.

Transparency therefore increasingly needs to include information about:

How systems are developed.

What risks have been identified.

What evaluations have been performed.

What limitations are known.

What safeguards exist.

How incidents are handled.

The governance challenge is to make transparency useful rather than merely symbolic.

INFORMATION SHARING

International AI safety also depends on information sharing.

If one organisation discovers a serious vulnerability, other organisations may need to know about it.

If a particular misuse technique becomes widespread, governments and companies may need to coordinate responses.

If evaluation methods improve, researchers and developers may benefit from sharing methodologies.

This creates an international public-interest dimension.

AI safety information can have value beyond the organisation that discovers it.

The Hiroshima Process therefore contributed to the emerging idea that responsible AI governance requires structured information sharing.

INCIDENT REPORTING

Incident reporting is another important part of this evolution.

Traditional technology governance often learns from failures.

Aircraft safety improved partly because aviation developed systems for investigating incidents.

Cybersecurity has similarly developed processes for vulnerability disclosure and incident response.

AI governance is increasingly moving toward a comparable logic.

If an advanced AI system produces a serious harmful outcome, the incident should not simply disappear.

There needs to be a mechanism for documenting what happened, understanding why it happened and learning from it.

This creates institutional memory.

It also helps policymakers understand whether existing governance mechanisms are working.

THE CHALLENGE OF VOLUNTARY GOVERNANCE

However, the Hiroshima Process also reveals an important limitation.

The Code of Conduct is voluntary.

Organisations are encouraged to commit to it.

But there is no single global regulator capable of enforcing it.

This raises a fundamental question.

Can voluntary governance keep pace with increasingly powerful AI?

There is no simple answer.

Voluntary frameworks can move faster than legislation.

They can adapt more easily to technical change.

They can establish common practices before formal regulation exists.

They can also support international coordination where legally binding agreements are difficult to negotiate.

But voluntary frameworks depend on participation.

Their credibility depends on implementation.

And their effectiveness may vary across organisations.

This is why voluntary and mandatory governance increasingly need to be understood as complementary rather than automatically competing approaches.

THE G7 AND GOVERNANCE INTEROPERABILITY

Another important issue addressed through the Hiroshima Process was interoperability.

Different countries were developing different AI governance systems.

The EU was moving toward a comprehensive legal framework.

The United States was developing a combination of executive, regulatory, standards-based and sectoral approaches.

China was developing its own regulatory and standards architecture.

Other countries were creating national frameworks.

If these systems became completely incompatible, international AI development could become more fragmented.

The Hiroshima Process therefore contributed to the search for common principles and interoperable approaches.

This does not mean that countries need identical laws.

It means that different governance systems may need enough common ground to interact effectively.

THE ROLE OF THE OECD

The OECD played an important role in supporting the Hiroshima Process.

It conducted a stocktaking of G7 approaches to generative AI and helped inform the process through analysis of opportunities, risks and policy priorities.

The OECD’s 2023 report described the process as an effort toward a G7 common understanding of generative AI.

This was institutionally significant.

The OECD already had the 2019 AI Principles.

The Hiroshima Process could therefore build on an existing international policy framework while addressing the new governance challenges created by generative AI.

The relationship between the two illustrates how international governance can evolve incrementally.

One framework creates principles.

Another develops more specific expectations.

Later mechanisms can focus on implementation and monitoring.

FROM G7 TO A WIDER COMMUNITY

The Hiroshima Process was initiated within the G7.

But the G7 recognised that AI governance could not remain limited to seven economies.

The October 2023 leaders’ statement explicitly called for outreach and consultation beyond the G7, including developing and emerging economies.

This is important for the Global South.

AI systems are global technologies.

Their impacts do not stop at G7 borders.

Developing countries are major users and increasingly important markets.

Their populations can be directly affected by AI-generated content, automated decision-making and AI-enabled services.

Yet their representation in international AI governance has historically been uneven.

The Hiroshima Process therefore raised a broader question.

Can an international AI governance system be legitimate if the countries most affected by AI do not have meaningful participation in shaping it?

THE GLOBAL SOUTH QUESTION

This question is especially relevant to developing countries.

Many countries may not have the capacity to build frontier models.

But they still need to govern the use of frontier models.

This creates an asymmetry.

Model-development capacity and governance responsibility are not necessarily located in the same countries.

A country may import AI technology while still being responsible for protecting its citizens.

This creates demand for local regulatory expertise.

It also creates demand for independent evaluation capacity.

If governments cannot independently assess AI systems, they may become dependent on vendors to explain their own technologies.

That can weaken regulatory capacity.

The Global South therefore needs not only access to AI.

It needs AI governance capability.

THE BANGLADESH RELEVANCE

For Bangladesh, this is an important strategic lesson.

Bangladesh is unlikely to need the same institutional structure as a frontier AI developer.

But Bangladesh still needs the ability to understand and govern AI systems developed elsewhere.

That includes the capacity to assess AI risks in education, financial services, healthcare, public administration, labour markets and information environments.

It also includes the ability to participate in international AI policy discussions.

This requires research institutions capable of translating global frameworks into national context.

It requires universities that can develop AI policy, governance and technical evaluation expertise.

It requires public-sector professionals who can understand AI systems sufficiently to regulate their use.

And it requires independent institutions capable of producing evidence rather than simply reproducing international policy language.

This is where organisations such as Atlas AI Institute could potentially contribute through research, assessment and capacity-building work.

THE HIROSHIMA PROCESS AS A GOVERNANCE BRIDGE

The broader significance of the Hiroshima AI Process is that it sits between two models of governance.

On one side is soft governance.

Principles.

Guidelines.

Codes of conduct.

Voluntary commitments.

On the other side is hard governance.

Legislation.

Regulation.

Enforcement.

Sanctions.

The Hiroshima Process belongs largely to the first category.

But it can help prepare the ground for the second.

Voluntary frameworks can create common terminology and expectations that later become incorporated into national laws or regulatory standards.

This is one reason why international soft law should not automatically be dismissed as irrelevant.

Its influence can occur indirectly.

THE IMPLEMENTATION PROBLEM

The next question is whether voluntary commitments can be verified.

If an organisation says that it manages AI risks responsibly, what evidence supports that claim?

Has the model been evaluated?

Were independent testers involved?

Were serious incidents recorded?

Were known limitations disclosed?

Were risks reassessed after major model updates?

Were downstream developers informed?

These questions transform a voluntary principle into an evidence problem.

And this is where AI governance is increasingly moving.

From commitments.

To implementation.

From implementation.

To verification.

FROM SELF-COMMITMENT TO MONITORING

The evolution did not stop in 2023.

The Hiroshima Process later developed mechanisms for monitoring application of the Code of Conduct.

The OECD launched a pilot monitoring framework in 2024 and later introduced a global reporting framework through which organisations could report on their practices against the Code.

This development is important because it illustrates how governance can evolve.

A voluntary code creates expectations.

A reporting mechanism creates evidence.

Comparable reporting creates the possibility of monitoring.

Monitoring can eventually inform stronger governance.

This is precisely the movement from principles toward implementation that defines the broader evolution of AI governance.

THE GOVERNANCE INFRASTRUCTURE IS GROWING

The Hiroshima Process therefore contributed to a broader ecosystem of AI governance infrastructure.

International principles.

Codes of conduct.

Risk-management frameworks.

Model evaluations.

Red-teaming.

Incident reporting.

Transparency mechanisms.

Standards.

Auditing.

Regulatory oversight.

These mechanisms do not replace legislation.

They complement it.

A mature AI governance system is likely to require multiple layers.

No single institution can manage every risk associated with advanced AI.

THE GEOPOLITICAL DIMENSION

The Hiroshima Process also demonstrated that AI governance is increasingly geopolitical.

The G7 was not discussing AI only as a consumer technology.

It was discussing advanced AI as a technology with implications for economic growth, democratic values, security and international cooperation.

This is important because AI governance increasingly overlaps with questions of technological power.

Who develops frontier models?

Who controls computing infrastructure?

Who has access to advanced chips?

Who controls data?

Who sets standards?

Who determines safety thresholds?

Who evaluates the most capable systems?

These are increasingly questions of international power.

AI diplomacy is therefore becoming part of technology geopolitics.

THE LIMITS OF G7-LED GOVERNANCE

At the same time, the Hiroshima Process should not be interpreted as a complete global governance solution.

The G7 represents a specific group of countries.

Other major AI actors have different governance models and strategic interests.

China is outside the G7.

Many developing countries are outside the G7.

Emerging economies have different development priorities.

The international community therefore still faces a representation problem.

A governance framework created by a small group of advanced economies may provide useful standards, but global legitimacy requires broader participation.

This is why subsequent efforts to involve countries beyond the G7 are strategically important.

THE CENTRAL LESSON

The Hiroshima AI Process demonstrates that AI governance is becoming increasingly layered.

International organisations establish principles.

Governments develop laws and policies.

Companies create internal governance systems.

Technical communities develop evaluation methodologies.

Standards organisations create common practices.

International forums coordinate policy.

Civil society and researchers provide external scrutiny.

The future governance system will likely emerge from the interaction of all these layers.

There may never be one global AI law.

Instead, the world may develop an interconnected governance ecosystem.

THE MOVE TOWARD AI SAFETY DIPLOMACY

The Hiroshima Process also prepared the ground for another major development in 2023.

The same year saw the UK-hosted AI Safety Summit at Bletchley Park.

There, governments and other stakeholders focused more explicitly on the risks associated with highly capable AI systems.

The distinction between general AI governance and frontier AI safety was becoming clearer.

The Hiroshima Process focused heavily on principles, developer responsibilities and international governance.

The Bletchley process would place greater emphasis on frontier safety and international cooperation around advanced AI risks.

Together, these developments signalled the emergence of AI safety diplomacy.

CONCLUSION

The Hiroshima AI Process represented an important stage in the evolution of global AI governance.

It emerged at a moment when generative AI had made the limitations of existing governance approaches increasingly visible.

The process did not create a global AI regulator.

It did not establish binding international law.

And it did not eliminate disagreements between countries.

Its importance lies elsewhere.

It created a practical international framework for discussing the responsibilities of organisations developing advanced AI systems.

It moved the conversation from broad principles toward risk identification, evaluation, transparency, information sharing and organisational accountability.

It demonstrated that AI governance was becoming an international diplomatic issue.

It also revealed the central tension that will shape the next phase of governance.

International AI governance needs to move quickly enough to keep pace with technological change, but it also needs legitimacy, transparency, evidence and broad participation.

The Hiroshima Process was an important step in that direction.

But it left a deeper question unanswered.

What happens when AI systems become sufficiently powerful that voluntary commitments are no longer enough?

That question leads directly to the next stage of the global AI governance story.

THE NEXT PHASE:

From Hiroshima to Bletchley.

The next article examines the rise of international AI safety diplomacy in 2023, including the Bletchley AI Safety Summit and the emerging debate around frontier-model risks.

The central question will be:

Can countries cooperate internationally on the safety of increasingly powerful AI systems even when they compete strategically over the technology itself?

ATLAS AI GOVERNANCE SERIES:

“From Principles to Power: The Evolution of AI Governance”

Part 1 — What Is AI Governance—and Why Has It Become So Important?

Part 2 — Before AI Governance: What Did We Learn from Data Protection, Cybersecurity and Internet Governance?

Part 3 — Why Did AI Governance Become a Global Policy Issue in the 2010s?

Part 4 — 2019: OECD AI Principles and the Start of a New Era in AI Governance

Part 5 — 2021: UNESCO Recommendation and the Global Ethics Dimension of AI Governance

Part 6 — 2022–23: Generative AI and the New Governance Problem

Part 7 — 2023: Hiroshima AI Process and the G7

Part 8 — Bletchley to Seoul: The Rise of AI Safety Diplomacy