The OECD: adherents, observatory, and the definition that conquered the world
Lesson 2 of 5 in Global Soft Law: OECD, UNESCO, and the G7 Hiroshima Process.
On 22 May 2019 the OECD Council adopted the Recommendation on Artificial Intelligence — the first intergovernmental AI standard anywhere. Its structure is deliberately simple: five values-based principles for anyone building or operating AI, and five recommendations telling governments what to do about it (invest in research, build the ecosystem, shape enabling policy, prepare the workforce, cooperate internationally). A month later the G20 — including China and Russia — endorsed AI principles drawn from the OECD text at the Osaka summit. Soft law had gone from think-tank exercise to G20 language in under a year; the EU AI Act would need five.
But the Recommendation’s most consequential export is a single sentence: the definition of an AI system. Refined in 2023–24 to say a machine-based system that infers, from the input it receives, how to generate outputs, it became the anchor text for the EU AI Act, the Council of Europe treaty, and statutes and frameworks well beyond. When two regulators in different jurisdictions can agree on what an AI system is, everything downstream — crosswalks, mutual recognition, multinational compliance — gets cheaper. That is definitional interoperability, and it is the OECD’s quiet superpower.
2019 original
Built for the pre-generative world. Five principles addressed to "AI actors", five recommendations to governments. The headline risks in view: bias, opacity, labour disruption. The definition spoke of systems making "predictions, recommendations, or decisions" — classifiers and scorers, the AI of its day.
Adherence began with OECD members plus a handful of partners; the G20 endorsement in June 2019 multiplied its reach without adding a single formal adherent.
2024 revision
The generative-AI patch, adopted May 2024. The revised text confronts what 2019 could not have seen: mis- and disinformation from generative systems appears by name; safety language is strengthened (including the ability to override or safely decommission systems); responsible business conduct and information integrity enter; rule of law joins the values list; and the updated definition — outputs now include "content", systems may "adapt after deployment" — replaces the old one.
Why it matters operationally: laws that anchored to the OECD definition inherit the update’s logic. The 2024 revision is the rare case of soft law being maintained like software — versioned, patched, and kept compatible with the statutes built on top of it.
Adherence is the OECD’s membership trick. The Recommendation is open to non-members, and 47 governments have adhered — including Brazil, Singapore, and Ukraine, none of them OECD members, plus the EU in its own right. Adherence costs nothing in sovereignty and buys a seat in the club that writes the shared vocabulary. Watch the pattern: countries adhere years before they legislate, and when they finally do legislate, the statute speaks OECD.
Then there is the machinery. The OECD.AI Policy Observatory (launched February 2020) is the field’s reference database — live repositories of national AI policies and strategies across dozens of jurisdictions, backed by an expert network that feeds the analysis. Its sharpest instrument is the AI Incidents Monitor (AIM), which scans global news coverage to catalogue real-world AI incidents and hazards. AIM matters for a reason beginners miss: soft law cannot fine anyone, but it can make harm visible and comparable, and visible harm is what moves legislatures.
Key terms: OECD AI Principles, ai incidents monitor, policy observatory, gpai partnership, definitional interoperability
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.