OECD AI Principles: What They Are and Why They Matter

In May 2019, something unusual happened in technology policy: 42 countries agreed on a common set of principles for artificial intelligence. No treaty, no enforcement mechanism — just agreement, on paper, about what responsible AI should look like. The OECD AI Principles became the first intergovernmental standard on AI, and their fingerprints are on nearly every AI law and framework written since.

I reference these principles constantly, not because they tell companies what to do day to day, but because they are the common ancestor of modern AI governance. Understand them and you understand where the EU AI Act, the NIST framework, and dozens of national strategies came from.

This guide explains the five principles, the accompanying policy recommendations, and their real-world influence. For the governance landscape they helped create, see my overview of AI governance.

Table of Contents

• What the OECD Is and Why It Acted
• The Five Values-Based Principles
• Principle 1: Inclusive Growth and Well-Being
• Principle 2: Human-Centered Values and Fairness
• Principle 3: Transparency and Explainability
• Principle 4: Robustness, Security, and Safety
• Principle 5: Accountability
• The Policy Recommendations for Governments
• Real-World Influence
• What the 2024 Revision Changed
• Frequently Asked Questions

What the OECD Is and Why It Acted

The Organisation for Economic Co-operation and Development is a club of mostly high-income democracies that develops economic and social policy standards. It does not pass laws. Its power is softer: when the OECD publishes a standard, member countries tend to build their national policies around it, and non-members often follow.

By 2019, AI policy was fragmenting — every country writing its own rules from scratch. The OECD’s AI expert group spent a year distilling shared values into principles that could anchor national strategies without dictating specific laws. The result was adopted in May 2019 and updated in 2024 to reflect generative AI and newer developments.

The Five Values-Based Principles

The principles are addressed to organizations and individuals developing, deploying, or operating AI — not just governments. Here is what each one says, in plain terms.

Principle 1: Inclusive Growth and Well-Being

AI should benefit people and the planet: economic growth, social well-being, and sustainable development. In practice, this means asking who gains from an AI system and who might be left behind — and designing so the benefits spread rather than concentrate.

This principle is the least operational of the five — nobody gets audited on “inclusive growth.” But it sets the tone: AI governance is not just about preventing harm, but about directing the technology toward broadly shared benefit.

Principle 2: Human-Centered Values and Fairness

AI systems should respect human rights, democratic values, and fairness. Concretely: they should not discriminate, should respect privacy, and should preserve human agency — people should remain in meaningful control of decisions that affect them.

This is the principle behind bias testing requirements, human-in-the-loop mandates, and the general insistence across AI law that consequential decisions keep a human meaningfully involved. When a regulation requires fairness assessments, it is operationalizing this principle.

Principle 3: Transparency and Explainability

People interacting with AI should know they are interacting with AI, and those affected by AI decisions should be able to understand the basis for them — to a degree appropriate to the context. A movie recommendation needs less explanation than a rejected loan application.

This principle drove the transparency duties now found in the EU AI Act (chatbot disclosure, deepfake labeling) and the broader push for model documentation. Note the careful phrasing: “appropriate to the context” — the OECD never demanded full technical explainability for every system, which would be impossible for modern deep learning.

Principle 4: Robustness, Security, and Safety

AI systems should work reliably, resist attacks, and fail safely. This covers the engineering disciplines: testing, validation, cybersecurity, and monitoring. A system that is accurate in the lab but brittle in deployment violates this principle.

In corporate governance terms, this is the principle behind red-teaming, robustness testing, incident reporting, and the general expectation that AI systems get the same engineering rigor as other safety-relevant software.

Principle 5: Accountability

Organizations and individuals developing, deploying, or operating AI should be accountable for its proper functioning, in line with these principles. Someone must answer for the system.

This is the principle I see cited most in boardrooms, because it translates directly into organizational design: named owners, documented decisions, and clear escalation paths. No accountability structure, no compliance with this principle.

The Policy Recommendations for Governments

Alongside the five principles for AI actors, the OECD issued recommendations for policymakers:

• Invest in AI research and development, including public-interest applications.
• Foster an inclusive AI ecosystem, with attention to small businesses and developing countries.
• Shape an enabling policy environment — interoperable frameworks rather than fragmented rules.
• Build human capacity — education and training for the AI transition.
• Cooperate internationally on standards, measurement, and incident sharing.

These are aspirations, not obligations. But they shaped how governments wrote their national AI strategies — nearly every one echoes this structure.

Real-World Influence

The principles’ influence is hard to overstate:

• G20 endorsement. The G20 adopted a human-centered AI agenda explicitly “drawing from” the OECD principles, extending their reach beyond OECD members.
• National strategies. Dozens of countries reference the principles in their AI strategies and laws.
• The EU AI Act. Its risk-based structure and emphasis on human oversight, transparency, and accountability track the principles closely.
• Corporate policies. Company AI principles — Google’s, Microsoft’s, and many others — rhyme with the OECD’s five, because everyone was reading the same document.
• Measurement. The OECD’s AI Policy Observatory tracks implementation globally, creating a feedback loop between principles and practice.

Critics note the principles are non-binding and vague — fair criticisms. But as a coordination device, they worked: they gave the world a shared vocabulary for AI governance before the hard laws arrived.

For teams implementing these ideas in practice, understanding how AI models are trained grounds the abstract principles in technical reality, and surveying AI tools shows how widely the principles’ concerns now apply.

What the 2024 Revision Changed

Five years after the original adoption, the OECD revised its principles in May 2024 — the first major update, and a telling one, because it shows how the governance conversation changed between the classifier era and the generative AI era.

The biggest shift is scope. The 2019 principles defined AI narrowly, around the machine learning systems of the time. By 2024, general-purpose and generative models had arrived, and the revision expanded the definition to cover them explicitly — including the responsibilities of organizations deploying models they did not train. That matters because it closes a governance gap: under the original text, a company could plausibly argue that a foundation model it merely licensed fell outside its accountability obligations. The updated principles make that argument much harder.

The transparency principle also got teeth. The original wording called for transparency about AI systems generally. The revision sharpens this into expectations about disclosure of AI-generated content, content provenance, and labeling — a direct response to synthetic media, deepfakes, and the collapse of “assume it’s real until proven otherwise.” Where 2019 asked for explainability as a design virtue, 2024 asks for mechanisms that let ordinary people know when they are interacting with or consuming AI output.

Misinformation and information integrity moved from the margins to the center. The revision strengthens provisions on protecting against AI-amplified misinformation, manipulative content, and risks to democratic processes — issues that barely registered in the 2019 drafting sessions but dominated the years after. Similarly, the safety principle was updated to address advanced model capabilities explicitly: not just “be robust,” but account for the risks that emerge when models can plan, use tools, and operate semi-autonomously.

The policy recommendations were restructured as well, and this is the part governments felt most directly. The updated text gives clearer guidance on regulatory sandboxes, on interoperability between national frameworks, and on measuring implementation — less abstract advice, more operational detail. The OECD’s AI Policy Observatory now tracks how adherents translate the principles into law, creating accountability for the coordination device itself.

One thing the revision did not do: make anything binding. Critics who wanted enforcement mechanisms were disappointed again. But the revision’s logic is that shared vocabulary must come before shared rules — you cannot harmonize regulation across forty-plus countries until they agree on what they are regulating. The 2024 update gave them that vocabulary for the generative era.

Practically, if your organization aligned its AI policy with the 2019 principles, the revision means three updates: expand scope to cover general-purpose and third-party models, add content-provenance disclosure, and revisit your misinformation controls. Those three items are the difference between a policy written for 2019’s AI and one that works for today’s.

Frequently Asked Questions