Publication of the OECD.AI Index / Reflections on Measuring the Implementation of AI Governance

I. The Challenge of Measuring Agreed-Upon Governance Frameworks

As international policy consensus regarding AI continues to make steady progress, the question arises as to how to objectively track the extent to which established principles are taking root in the legal systems, markets, and technological foundations of individual countries. The "OECD.AI Index Technical paper," published by the OECD in 2026, presents a composite index for quantitatively measuring the implementation status of the AI Recommendation—adopted in 2019 and revised in 2024—across member countries [Note 1]. This can be positioned as an attempt to facilitate a transition from the stage of formulating principles as ideals to the stage of tracking their implementation.

1. From Capability Assessment to Implementation Measurement

Many existing AI-related indices have focused on measuring a nation's technological superiority or its readiness for adopting new technologies. The Stanford "AI Index," the IMF's "AI Preparedness Index," and Oxford Insights' "AI Readiness Index" are typical examples, primarily focusing on competitiveness indicators such as the number of research papers, patent applications, infrastructure development, and venture capital investment [Note 1].

In contrast, the subject of measurement for this index is different. It asks to what degree the OECD AI Recommendation is being complied with—that is, to what extent internationally agreed-upon normative frameworks are actually being implemented in each country. It incorporates 28 variables corresponding to five policy pillars: AI research and development, foundational infrastructure, policy environment, employment and skills, and international cooperation. It evaluates both policy engagement, such as the establishment of AI safety institutes and the introduction of regulatory sandboxes, and actual outcomes, such as VC flows, computing infrastructure, and research results, within the same framework [Note 1]. As an attempt to visualize the distance between "law on the books" and "law in action," this design is also interesting from the perspective of legal policy studies.

2. Functioning as a Diagnostic Tool

This index explicitly states that it is not a trophy for policymakers to compete for top spots in international rankings, but rather a tool to function as a diagnostic for identifying imbalances in their own country's policy efforts [Note 1]. Even if a country puts forward impressive principles and strategies, it is difficult to achieve policy goals if the supporting telecommunications infrastructure and computing resources are not in place. Conversely, if a country invests heavily in R&D but has little involvement in forming international safety standards, its evaluation in terms of implementing the recommendations will be mixed. The index illuminates these imbalances by component.

In the 2023-2024 assessment, the overall scores for each country ranged from 0.17 to 0.66, with the United States, the United Kingdom, and Switzerland ranking at the top [Note 1]. The breakdown of each country's strengths is contrasting: the U.S. was driven by the scale of VC investment, the number of AI models developed, and infrastructure via GPU clusters; the U.K. showed strengths in digital government utilization and data infrastructure; and Switzerland recorded high figures in skill-related indicators, particularly in the knowledge sub-component [Note 1]. This data shows that top-tier countries do not necessarily follow a uniform strategy, and that different approaches can lead to success in implementing the OECD AI principles.

II. The Constraint of Economic Structure

1. What Differences in Normalization Methods Reveal

In this index, each variable is generally normalized per working-age population and converted to a scale of 0 to 1. This process is intended to remove bias caused by differences in population size and to make the achievements of each country comparable.

What I am particularly interested in regarding the original text is the comparison of normalization methods conducted as a sensitivity analysis. It is reported that when GDP is used as the basis for normalization instead of population, significant fluctuations occur in each country's score and ranking [Note 1]. While this is a matter of technical verification, its implications are not small. A structural bias—that the implementation status of AI governance is determined to a considerable extent not only by a country's stance or choice of institutional design, but also by its economic scale and level of wealth—is hidden within the choice of normalization method.

2. The Gap Between Normative Consensus and Implementation Capacity

International governance discussions tend to lean toward normative dimensions, such as the design of legal systems and the formulation of ethical principles. However, building a safe development environment, developing computing resources, and securing highly skilled technical personnel all require massive capital investment. The fact that rankings fluctuate significantly with GDP normalization numerically demonstrates the difficulty for economically less-developed countries to meet the requirements of international governance. There is a clear barrier of economic power between consensus on norms and the capacity to implement them domestically. Policy comparisons that overlook this point have structural limitations.

III. In Conclusion

The attempt of this index to translate AI principles into concrete policy actions and market observation data can be evaluated as providing a foundation for evidence-based policymaking. The original text itself acknowledges that the comprehensiveness of the data and the selection of variables require further refinement [Note 1], and I intend to continue monitoring the revision process. At the same time, the reality shown by this index—that the degree of governance achievement is also a function of economic power—reconfirms that when discussing how to make principles agreed upon in international rule-making forums effective, the perspective of differences in economic conditions cannot be omitted.

(Magazine) "AI and Law - Reflections"

* Please refer to the following for the table of contents

Note General Provisions Article 3, Paragraph 2, First Sentence
3.2 The copyright of digital content created by the creator belongs to the creator.