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

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

The history of AI governance can be divided into several important phases.

The early discussions were largely about ethics.

Then came questions about privacy, bias, transparency and accountability.

Governments began developing national AI strategies.

But something important happened in 2019.

AI governance became explicitly intergovernmental.

On 22 May 2019, the OECD Council adopted the Recommendation of the Council on Artificial Intelligence.

The OECD AI Principles did not create a global AI law.

They did something different.

They established a common international policy framework around what trustworthy and responsible AI should look like.

That distinction is important.

The 2019 OECD Principles became one of the foundations on which later international AI governance efforts would build.

WHY DID THE OECD ACT?

By 2019, artificial intelligence had already become a major economic and policy issue.

Governments were investing in AI research.

Companies were deploying machine-learning systems across increasingly important sectors.

Public institutions were exploring AI for government services.

At the same time, concerns were growing around privacy, discrimination, transparency, safety and accountability.

The policy problem was becoming increasingly international.

AI systems could be developed in one country, trained using data from multiple jurisdictions, deployed through global digital platforms and used by people around the world.

National policies therefore could not completely address the governance challenge by themselves.

There was a growing need for common principles.

The OECD was well positioned for this role because it had decades of experience developing international policy recommendations on areas such as economic policy, digital policy, privacy, competition and responsible business conduct.

AI represented a natural extension of this work.

FROM NATIONAL STRATEGIES TO INTERNATIONAL PRINCIPLES

The 2010s saw governments increasingly develop national AI strategies.

Different countries had different economic priorities, institutional structures and regulatory traditions.

Some focused heavily on innovation.

Others emphasised social protection, privacy or human rights.

Some placed strong emphasis on national competitiveness.

Others focused on public-sector applications and responsible technology.

This diversity created a potential problem.

If every country developed completely different expectations for AI, international technology development could become increasingly fragmented.

Companies operating across borders could face incompatible requirements.

Researchers could encounter different standards.

Governments could struggle to cooperate.

The OECD Principles offered a way to establish a shared baseline without attempting to create a single global regulatory system.

This was one of their most important characteristics.

THE FIVE PRINCIPLES

The OECD AI Principles were built around five broad areas.

The first concerned inclusive growth, sustainable development and well-being.

The idea was that AI development should contribute to positive economic and social outcomes rather than simply increasing technological capability.

The second focused on human-centred values and fairness.

AI should respect the rule of law, human rights, democratic values, diversity and non-discrimination.

The third addressed transparency and explainability.

People should have meaningful information about AI systems and their interactions with them, while organisations should provide appropriate information about the factors influencing significant outputs.

The fourth concerned robustness, security and safety.

AI systems should function appropriately throughout their life cycle, with mechanisms for addressing risks and unintended consequences.

The fifth principle focused on accountability.

Organisations and individuals involved in AI development and deployment should be responsible for the proper functioning of AI systems and for respecting the relevant principles.

Taken together, these principles created a relatively simple but influential governance architecture.

AI should not merely be capable.

It should also be trustworthy.

TRUSTWORTHY AI AS A POLICY CONCEPT

The concept of trustworthy AI became increasingly important during this period.

Trustworthiness was not presented simply as a matter of public relations.

It involved technical, institutional and social dimensions.

An AI system could be technically impressive but still raise questions about fairness.

It could be accurate but insufficiently secure.

It could be useful but deployed without adequate accountability.

It could perform well in testing but fail under conditions different from those used during development.

The OECD framework therefore helped reinforce an important principle:

AI governance cannot be reduced to model performance.

The quality of an AI system depends partly on the environment in which it is developed and used.

This includes data, testing, deployment conditions, human oversight, organisational responsibility and mechanisms for monitoring.

THE IMPORTANCE OF HUMAN-CENTRED VALUES

The OECD Principles also connected AI governance with broader democratic and human-rights values.

This was significant because AI governance could otherwise have developed primarily as a technical standards exercise.

The OECD approach recognised that AI systems operate within societies.

Their impacts can affect individuals, workers, consumers, organisations and public institutions.

Questions about fairness, dignity, autonomy and democratic values therefore cannot be separated completely from technological design.

This created a bridge between traditional technology policy and broader public policy.

AI governance was becoming part of a larger conversation about how technology should interact with society.

TRANSPARENCY WITHOUT A SINGLE EXPLANATION STANDARD

Transparency became another major element of the OECD framework.

But transparency in AI is more complicated than simply opening the source code of a model.

Different users require different forms of information.

A regulator may need technical documentation.

A developer may need information about model limitations.

A deployer may need information about performance and risk.

An affected individual may simply need to know that an automated system was involved in an important decision and understand how to challenge the outcome.

The OECD approach therefore helped move the discussion away from a simplistic idea of transparency toward a more practical governance question:

What information is necessary for people and institutions to understand, evaluate and appropriately challenge AI systems?

This question would become increasingly important as AI systems became more complex.

ROBUSTNESS AND SAFETY

Another important contribution of the OECD Principles was their treatment of robustness, security and safety.

This was significant because AI failures are not always caused by malicious behaviour.

A system can fail because of poor-quality data.

It can behave unexpectedly in unfamiliar environments.

It can produce unreliable outputs.

It can be vulnerable to manipulation.

It can perform differently after deployment than during testing.

This means that AI governance must consider the full lifecycle of a system.

Development is only one stage.

Testing, deployment, monitoring and incident response are also part of responsible AI governance.

This lifecycle perspective would become increasingly important in later AI regulation.

ACCOUNTABILITY

The OECD Principles also established accountability as a core governance concept.

This may appear obvious, but accountability becomes complicated when many actors are involved.

A modern AI system can involve:

A model developer.

A data provider.

A cloud or compute provider.

An application developer.

A deploying organisation.

A public authority.

An end user.

A third-party auditor.

When something goes wrong, responsibility may not be located in a single institution.

The OECD Principles therefore helped establish a policy direction in which organisations involved in AI would need to understand and manage their responsibilities.

This idea would later become central to AI risk management, auditing and regulatory compliance.

FROM PRINCIPLES TO POLICY

The importance of the OECD Principles was not that they immediately solved these problems.

They did not.

Their significance was that they created a common policy language.

Governments, companies, researchers and international organisations could discuss AI governance using concepts such as:

Human-centred values.

Fairness.

Transparency.

Robustness.

Safety.

Accountability.

These concepts became increasingly common across national strategies and international frameworks.

The result was a degree of convergence.

Different jurisdictions continued to develop different approaches, but they increasingly used overlapping vocabulary.

That created an important foundation for future cooperation.

THE OECD PRINCIPLES WERE NOT A GLOBAL AI LAW

It is important not to misunderstand what happened in 2019.

The OECD Recommendation was not a treaty imposing binding AI regulation on all countries.

It was not equivalent to a comprehensive AI law.

It did not establish a global AI regulator.

It did not create a single international enforcement mechanism.

Instead, it was a policy recommendation adopted through an intergovernmental organisation.

Its influence therefore came partly from its ability to establish shared expectations and provide a reference point for national policy development.

This distinction between principles and enforceable rules would become one of the central questions of AI governance in the following years.

THE G20 CONNECTION

The OECD framework also gained wider international visibility when the G20 adopted human-centred AI principles based on the OECD approach in 2019.

This mattered because the G20 brings together major advanced and emerging economies.

The movement from OECD principles toward broader G20 endorsement demonstrated that AI governance was becoming part of a wider international policy conversation.

AI governance was no longer confined to technology ministries or specialist policy communities.

It was increasingly connected to discussions about economic development, international cooperation and global digital transformation.

A NEW INTERNATIONAL LANGUAGE

One of the most important consequences of the OECD Principles was therefore linguistic as much as institutional.

Governance requires shared concepts.

Countries cannot easily cooperate if they use completely different definitions of responsible AI.

The OECD framework contributed to a shared vocabulary.

Trustworthy AI became a widely used policy concept.

Human-centred values became a common reference.

Transparency and explainability became established governance concerns.

Robustness and safety became recognised as lifecycle issues.

Accountability became a central responsibility principle.

This common language helped prepare the ground for later international initiatives.

THE LIMITATION OF PRINCIPLES

But principles have limits.

A principle can say that AI should be transparent.

It does not automatically determine what documentation a company must provide.

A principle can say that AI should be safe.

It does not automatically establish which tests must be conducted before deployment.

A principle can say that organisations should be accountable.

It does not automatically determine who should audit them.

This created the next major governance challenge.

How do principles become operational?

That question would dominate the next phase of AI governance.

FROM PRINCIPLES TO IMPLEMENTATION

The transition from principles to implementation requires institutions.

Governments need regulatory authorities.

Companies need internal governance systems.

Technical organisations need standards.

Auditors need methodologies.

Researchers need evaluation frameworks.

Organisations need mechanisms for monitoring and incident reporting.

Users need channels for complaints and redress.

In other words, responsible AI cannot exist only as a statement of values.

It needs infrastructure.

This distinction would become increasingly visible in the 2020s.

WHY 2019 MATTERED

The OECD Principles were important because they represented a transition.

Before this period, AI governance was increasingly discussed as an ethical and social challenge.

After 2019, AI governance increasingly became an international policy architecture.

The shift can be understood in three stages.

First, societies identified problems created or amplified by AI.

Then governments and institutions articulated principles for responsible AI.

The next step was to determine how those principles could be implemented, measured and enforced.

The OECD Principles were therefore not the end of the first phase of AI governance.

They were a bridge between phases.

THE START OF A GLOBAL GOVERNANCE ARCHITECTURE

The period following 2019 would see a rapid expansion of AI governance initiatives.

UNESCO would develop a global Recommendation on the Ethics of Artificial Intelligence.

The G7 would develop the Hiroshima AI Process.

Governments would introduce or strengthen AI regulation.

The European Union would move toward a comprehensive AI law.

International organisations would expand work on AI standards, safety and policy coordination.

The Council of Europe would develop a legally binding international treaty framework on AI and human rights, democracy and the rule of law.

The vocabulary established during the earlier principles era would increasingly be connected to legal and institutional mechanisms.

This was the beginning of a major transformation.

THE EMERGING GOVERNANCE GAP

At the same time, technological progress was accelerating.

AI systems were becoming more capable.

Generative models were beginning to attract significant research attention.

Large-scale models were demonstrating increasingly general capabilities.

The governance frameworks of 2019 had been designed primarily around a world in which many AI applications were relatively specialised.

The coming generation of foundation models would challenge that assumption.

A general-purpose model could be integrated into many different applications.

A single model could influence education, software development, customer service, research, media, healthcare and public administration.

This meant that governance could no longer focus only on individual applications.

It increasingly needed to consider the underlying models and their broader lifecycle.

THE GLOBAL SOUTH QUESTION

The international development of AI principles also raised an important question.

Who participates in creating global AI governance?

International principles can provide useful common frameworks.

But countries differ substantially in institutional capacity, infrastructure, regulatory expertise and access to technical resources.

A governance framework designed primarily around the experience of highly digitised economies may not automatically address the realities of developing countries.

For many countries in the Global South, the challenge is not simply regulating AI.

It is also developing the capacity to understand, evaluate and deploy AI.

That requires researchers, regulators, technical experts, standards professionals, public-sector capability and educational institutions.

This distinction will become increasingly important as AI governance becomes more complex.

AI GOVERNANCE AS CAPACITY

The OECD experience therefore reveals a broader lesson.

AI governance is not only about having good principles.

It is also about having the institutional capacity to apply them.

A country may formally support trustworthy AI.

But if it lacks technical expertise, regulatory institutions, evaluation infrastructure or research capacity, implementation can remain weak.

This creates an important distinction between policy adoption and governance capability.

A country can adopt international principles without necessarily possessing the institutional infrastructure required to operationalise them.

For developing countries, this may become one of the central challenges of the next decade.

FROM OECD TO THE NEXT PHASE

The 2019 OECD AI Principles helped establish an international baseline.

They demonstrated that governments could agree on broad principles for trustworthy AI despite differences in political systems, economic structures and national priorities.

But they also exposed the next problem.

Principles are easier to agree on than enforcement mechanisms.

It is relatively easy to agree that AI should be safe.

It is much harder to agree on what safety means in practice.

It is relatively easy to support transparency.

It is much harder to determine what information should be disclosed.

It is relatively easy to support accountability.

It is much harder to determine who is responsible when a complex AI system fails.

This gap between principle and implementation would become one of the defining issues of AI governance in the 2020s.

CONCLUSION

The 2019 OECD AI Principles marked an important transition in the history of AI governance.

They did not create a global AI regulator.

They did not establish a single international law.

And they did not resolve the difficult questions surrounding AI risk.

Their importance was different.

They established a widely recognised intergovernmental framework for thinking about trustworthy AI.

They connected AI development with human-centred values, fairness, transparency, robustness, safety and accountability.

They helped create a common international vocabulary.

They demonstrated that AI governance could become an area of international policy coordination.

And they provided a foundation on which later governance initiatives could build.

The deeper significance of 2019 is therefore not simply that the OECD published a set of principles.

It is that AI governance began moving from an emerging ethical discussion toward an organised international policy field.

The next challenge was inevitable.

If AI principles were going to matter, governments and institutions would eventually have to determine how those principles could be translated into concrete governance systems.

That question would take AI governance into its next major phase.

FROM PRINCIPLES TO GLOBAL ETHICS.

In the next article, we move from the OECD’s intergovernmental policy framework to a broader global ethics framework.

In 2021, UNESCO adopted the Recommendation on the Ethics of Artificial Intelligence.

It represented an important expansion of the conversation—from trustworthy AI and policy coordination toward human rights, democracy, cultural diversity, environmental sustainability and global ethical governance.

That development would further reshape the international AI governance landscape.

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