Updated May 2024
The OECD AI Principles are the first intergovernmental AI standard, adopted in May 2019 and updated in May 2024 to account for generative AI. Five value-based principles, five recommendations to governments, and no binding force whatsoever.
People dismiss this tier. I understand why — it's the tier of declarations and summits and communiqués, and if your job is compliance then a document nobody can enforce looks like noise.
I'd argue it's the most consequential tier on the board, and I can prove it with one example.
The proof
The EU AI Act needed a definition of "AI system." This is a law carrying penalties of up to €35 million or 7% of global turnover, and that definition decides who falls inside it.
The EU did not draft one.
It took the OECD's, as updated in May 2024.
So a non-binding recommendation from an economic cooperation body is now the trigger condition for a seven-percent penalty. Every argument about whether your system is in scope runs through language written by people with no enforcement power at all, several years before anyone was arguing about it.
That's what this tier does. It doesn't bind you. It writes the sentences that the things which bind you are made of.
I've watched a version of this play out before in a different industry. In revenue recognition and billing, the accounting standards bodies would publish something advisory, everyone in the market would ignore it as a theoretical exercise, and three years later it was in the audit. The people who read it early weren't being thorough. They were buying time.
The OECD, briefly
- Inclusive growth, sustainable development and well-being
- Human rights and democratic values
- Transparency and explainability
- Robustness, security and safety
- Accountability
- R&D investment
- Ecosystem development
- Governance environment
- Human capacity
- International cooperation
Honestly, the principles themselves are what you'd expect. The machinery around them is more useful.
The OECD.AI Policy Observatory maintains a national policy database — genuinely the best free source for "what is country X actually doing." The AI Incidents and Hazards Monitor and the Common Reporting Framework for AI Incidents are becoming the reference points for incident taxonomy, which matters more than it sounds: several regulatory regimes now require incident reporting without specifying a format. When somebody eventually specifies one, it'll look like this.
UNESCO
The Recommendation on the Ethics of Artificial Intelligence was adopted in November 2021 by all 193 member states. That makes it the most broadly subscribed AI ethics instrument in existence, which is a remarkable thing to have achieved and is also why it reads the way it does.
What distinguishes it is the tooling rather than the text. The Readiness Assessment Methodology is a national diagnostic that dozens of states have actually used. The Ethical Impact Assessment works at project level, and if you need an impact assessment methodology and aren't ready for ISO/IEC 42005, it's a credible place to start — and it's free.
The United Nations
Two mechanisms, both established by General Assembly Resolution A/RES/79/325 in August 2025.
The Independent International Scientific Panel on AI has forty members serving in a personal capacity, selected from more than 2,600 candidates and appointed in February 2026 for three-year terms. It's co-chaired by Yoshua Bengio and Maria Ressa. Its preliminary report launched on 1 July 2026 and warned, in fairly direct terms, that current safeguards aren't keeping pace with capability growth.
The Global Dialogue on AI Governance is the intergovernmental forum that receives those reports. The first session ran in Geneva on 6 and 7 July 2026 with participants from nearly 170 member states. The second is scheduled for New York in May 2027.
For most organisations this is context rather than compliance, and I'd treat it as an early-warning system rather than a reading list. The Panel's risk taxonomy is the sort of thing that shows up in a statute in 2029.
The G7 and the summits
The G7 Hiroshima Process produced guiding principles and a code of conduct for organisations developing advanced AI. Its most concrete output is the HAIP Reporting Framework, launched in February 2025 and hosted by the OECD — a common voluntary transparency template for developers. A Friends Group of 49 governments and regions beyond the G7 supports the code.
Then the summit series: Bletchley Park in November 2023, Seoul in May 2024, Paris in February 2025, Delhi in February 2026.
The summits are where frontier safety norms get negotiated in public before they turn up in law. Watch the Seoul commitments — sixteen companies signed up to publish safety frameworks in May 2024, and the shape of those commitments now appears in the EU's GPAI Code of Practice and in California's SB 53.
That's a two-year path from voluntary pledge to statutory obligation, in public, with the drafting visible the whole way.
How I'd actually use this tier
Don't read it for compliance. There's nothing to comply with.
Read it for vocabulary.
When a definition, a risk taxonomy or a reporting format appears in a soft law document, there's a reasonable chance you'll meet it again in a binding statute three to five years later. By then it's too late to influence and expensive to retrofit — but if you saw it coming, you built for it when it was cheap.
That's the whole value of Tier 8. Not what it requires of you. What it tells you about the shape of what's coming.
Things people ask me
Are the OECD AI Principles legally binding?
No. They're a non-binding recommendation. The EU AI Act nonetheless took its definition of an AI system directly from them, which gives them more practical force than their status suggests.
What is the UN Global Dialogue on AI Governance?
An intergovernmental forum established in 2025 that receives reports from the UN's Independent International Scientific Panel on AI. First session in Geneva, July 2026; second in New York, May 2027.
Why should a company care about non-binding AI principles?
Because they supply the definitions, risk taxonomies and reporting formats that binding law later adopts — usually with several years of warning for anyone paying attention.