AI Governance Design Index
Scores · index version 2.0

Five frameworks, side by side

Each framework is coded on the same twelve features. Every number links back to the anchor that produced it, so you can check the reasoning rather than take it on trust.

Draft coding — awaiting author review

All five frameworks are now coded against the twelve features. These are a first-pass draft: each cell has a justification and a primary source, but no cell has yet been checked by all three authors, nothing has been coded blind, and no inter-coder reliability testing has been done. Not yet findings — please do not cite until this notice is gone.

Frameworks included

Five landmark instruments spanning the spectrum from non-binding guidance to hard law, from universal membership to regional integration, and from purely voluntary to legally enforceable commitment.

  • OECD AI Principles

    Adopted 22 May 2019 · OECD/LEGAL/0449 · Recommendation, non-binding · 47 adherents

    The first intergovernmental standard on AI, adopted by the OECD Council by mutual agreement and adhered to by 47 countries including eight non-members. Its influence runs through adoption elsewhere rather than through enforcement. Coded as adopted in 2019; the May 2024 revision is recorded but, under rule 2, does not move a score — it changed the definition of an AI system rather than the instrument's design.

    Legal instrument OECD.AI overview

  • EU Artificial Intelligence Act

    Regulation (EU) 2024/1689 · adopted 13 June 2024 · in force 1 August 2024 · binding

    The first comprehensive binding AI law, applying risk-tiered obligations across the single market with penalties for non-compliance and a central AI Office. Coded as adopted. It was substantively amended by the Digital Omnibus on AI, Regulation (EU) 2026/1744, in force 27 July 2026, which deferred the high-risk application deadlines to December 2027 and August 2028 and expanded the AI Office's enforcement powers. Under coding rule 2 that amendment is recorded but does not move a score.

    Official Journal text Digital Omnibus amendments Implementation timeline

  • UNESCO Recommendation on the Ethics of AI

    Adopted 2021 · non-binding · the scored instrument for this row

    Adopted by all UNESCO member states in 2021, and the instrument in this pair that carries the operative machinery. Under coding rule 3 it is what gets scored. The Global Digital Compact, agreed within the Pact for the Future in 2024, is recorded as context rather than coded.

    UNESCO Recommendation Global Digital Compact

  • African Union Continental AI Strategy

    Endorsed by the AU Executive Council, 45th Ordinary Session, 18–19 July 2024 · non-binding

    The African Union's continental strategy for AI, setting shared priorities across member states without creating binding obligations or an enforcement mechanism. It provides for a five-year implementation plan with objectives targeted for 2030, and calls for an Advisory Board on AI, a regional AI Ethics Board and an expert group on peace and security — none of which it establishes itself.

    Continental AI Strategy

  • ASEAN Guide on AI Governance and Ethics

    Unveiled at the 4th ASEAN Digital Ministers' Meeting, 2 February 2024 · voluntary

    Regional guidance built on national-level implementation and consensus, with no central body, no obligations and no enforcement — the clearest contrast in the set to the EU approach. It describes itself as a living document to be periodically reviewed, and proposes an ASEAN Working Group on AI Governance that it does not establish. Coded as adopted in February 2024; the January 2025 generative-AI expansion is recorded, not scored.

    2024 Guide (PDF) Expanded guide, generative AI (PDF)

Axis profiles at a glance

Each axis runs 0–100, where 100 always means more of the property named. These are the unweighted means of the four features in each axis, with the two reverse-coded features inverted first.

OECD AI Principles

Adopted 2019

Power conc.38
Reversibility65
Efficiency53

EU AI Act

Adopted 2024

Power conc.79
Reversibility29
Efficiency60

UNESCO Recommendation

Adopted 2021

Power conc.49
Reversibility70
Efficiency40

AU Continental AI Strategy

Endorsed 2024

Power conc.41
Reversibility76
Efficiency41

ASEAN Guide

Adopted 2024

Power conc.24
Reversibility70
Efficiency39

The full matrix

Twelve features, five frameworks. Feature names link to their anchors on the index page. On the two reverse-coded rows, the small arrow shows the value that actually enters the axis average.

Scores 0–100 against the anchors in version 2.0 of the index. Higher always means more of the property named. Every framework is coded as adopted, at face value — see the four coding rules on the index page.
Feature OECD Principles EU AI Act UNESCO Rec. AU Strategy ASEAN Guide
Axis 1 · Power concentration
1.1Decision-making concentration 15 75 30 20 5
1.2Agenda-setting concentration 45 90 60 55 30
1.3Resource concentration 35 70 35 45 20
1.4Technical concentration 55 80 70 45 40
Power concentration 38 79 49 41 24
Axis 2 · Reversibility
2.1Exit possibility 100 0 95 100 100
2.2Amendment difficulty 60→40 45→55 30→70 35→65 40→60
2.3Sunset mechanisms 50 50 55 65 35
2.4Path dependency 30→70 90→10 40→60 25→75 15→85
Reversibility 65 29 70 76 70
Axis 3 · Efficiency
3.1Deliberation speed 80 45 50 65 75
3.2Binding implementation 10 75 15 5 0
3.3Optimised coordination 80 50 45 70 75
3.4Compliance verification 40 70 50 25 5
Efficiency 53 60 40 41 39

Power concentration Reversibility Efficiency Bar length under each number shows the score on a 0–100 scale.

Frameworks side by side

Grouped bar chart of three axis scores for five frameworks. The EU AI Act is far highest on power concentration at 79 and far lowest on reversibility at 29. The ASEAN Guide is lowest on power concentration at 24. Efficiency varies least, from 39 to 60.
Figure 1. Axis averages across the five frameworks. Power concentration and reversibility move sharply in opposite directions, and the EU AI Act sits at the extreme of both. Efficiency is the flat axis: only 21 points separate the most and least efficient framework, against 55 points on power concentration. Draft coding, awaiting author review.
Heatmap of twelve features by five frameworks, darker cells indicating higher scores. The EU AI Act column is dark across power concentration and light across reversibility; the four non-binding instruments are light across power concentration and dark across reversibility.
Figure 2. The full matrix as a heatmap, with reverse-coded features already inverted. Reading down a column shows a framework's design signature; reading across a row shows which features actually discriminate between frameworks and which do not. Draft coding, awaiting author review.
Horizontal bar chart of the EU AI Act's twelve feature scores grouped by axis, with axis averages of 79, 29 and 60.
Figure 3. The EU AI Act feature by feature, as the clearest single case. Reverse-coded features show the value entering the axis average, with the raw score in brackets. Draft coding, awaiting author review.
On the figures

All three are generated from the same numbers as the table above, so they cannot disagree with it. If you change a score, regenerate them — or the two will drift apart, which is the most common way an index like this loses a reader's trust.

How to cite these scores

Balbis, C., Sheikh, P. and Vincendeau, J. M. (2026). AI Governance Design Index, version 2.0, framework scores. Available at https://example.org/frameworks.html

Score changes after publication are recorded in the methodology notes.