3.1 The OECD AI Principles

The First Intergovernmental AI Standard

The OECD Recommendation on Artificial Intelligence was adopted by the OECD Council on 22 May 2019 and holds the distinction of being the first intergovernmental standard on AI. Although OECD recommendations are legally non-binding ‘soft law,’ adherents—originally the OECD member countries plus several partner economies—make a political commitment to implement them. Over 45 governments have now adhered, which is precisely why the AIGP exam treats the OECD principles as the common vocabulary underneath almost every later framework. When you see terms like human-centered, robustness, or accountability echoed in the EU AI Act, the NIST AI RMF, or a national strategy, the lineage usually traces back to this document.

The Recommendation was updated in May 2024. The revision modernized the definition of an AI system to keep pace with generative and general-purpose AI, and it added emphasis on safety, information integrity (addressing mis- and disinformation), and environmental sustainability. For the exam, remember that the structure did not change—there are still five principles and five recommendations—but the language now explicitly contemplates generative AI risks.

The Five Values-Based Principles

The first half of the Recommendation sets out five principles for responsible stewardship of trustworthy AI. These are addressed to all AI actors:

A common exam trap is confusing transparency and explainability (principle 3) with accountability (principle 5). Transparency is about disclosure and comprehension; accountability is about who is answerable when things go wrong. Another frequent distractor treats ‘privacy’ as a standalone principle—on the OECD list it sits inside the human-centered values and fairness principle. Note too that these principles are framed to apply across the entire AI system lifecycle, from design and data collection through deployment, operation and decommissioning, rather than to a single moment of development. That lifecycle framing is a recurring theme you will meet again in NIST and ISO/IEC materials, so anchoring it here pays dividends later.

The Five Recommendations to Policymakers

The second half is directed at governments and gives five national-policy recommendations:

• Investing in AI research and development — fund R&D, including trustworthy AI and open datasets.
• Fostering an inclusive, AI-enabling digital ecosystem — build the data, compute and technology infrastructure and support for adoption.
• Shaping an enabling, interoperable governance and policy environment — use agile, experimental approaches such as regulatory sandboxes.
• Building human capacity and preparing for labour-market transformation — equip people with skills and support fair transitions.
• International co-operation for trustworthy AI — work across borders on interoperable standards and responsible stewardship.

Notice the pattern: the principles speak to AI actors (developers, deployers, operators), while the recommendations speak to policymakers. Keeping that audience distinction straight is often enough to answer several questions.

Why the OECD Principles Matter

The strategic significance is that the OECD principles created global convergence. In June 2019, just weeks after adoption, the G20 endorsed a set of AI principles drawn directly from the OECD text, extending their reach to major non-OECD economies such as China, India, Brazil and Indonesia. This gave the world a shared normative baseline before any binding law existed.

The OECD also stood up supporting infrastructure: the OECD.AI Policy Observatory, a live AI Incidents Monitor, and a network of experts that maintains a widely cited definition of an AI system—the same definition that influenced the EU AI Act’s scope. For an AIGP professional, the takeaway is that the OECD framework is not merely historical trivia; it supplies the conceptual scaffolding (human-centered values, robustness, accountability, a lifecycle view) that you will encounter, in slightly different wording, throughout the rest of the exam. Understanding the OECD baseline makes the NIST characteristics of trustworthiness and the EU AI Act’s risk-based obligations feel like variations on a familiar theme rather than unrelated systems.

Finally, keep the soft-law nature in mind. The OECD Recommendation does not create enforceable obligations or penalties. Its power is normative and diplomatic—it aligns expectations and seeds harder instruments. That is a deliberate design choice, and the exam may test whether you can distinguish a non-binding recommendation from a binding regulation like the EU AI Act.

The OECD Definition and Its Downstream Reach

One technical contribution deserves special attention because it recurs across the exam: the OECD definition of an AI system. The 2024 update refined it to describe a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. This language was chosen partly for interoperability—so that different jurisdictions could regulate the same object. The EU AI Act adopted a definition closely aligned with the OECD’s, and other bodies followed, which is exactly the kind of convergence the OECD set out to produce.

It is also useful to place the OECD instrument alongside its peers. The G20 endorsement (2019) broadened political reach; UNESCO’s Recommendation on the Ethics of AI (2021) added a values-and-ethics layer adopted by nearly 200 states; and the Council of Europe’s Framework Convention on AI (2024) created the first legally binding treaty on AI and human rights. On the exam, do not confuse these: the OECD Recommendation is soft law, UNESCO is an ethics recommendation, and the Council of Europe Convention is a binding treaty. The OECD’s enduring role is to be the shared, technically grounded baseline that these other instruments build upon and cross-reference.