The OECD AI Policy Observatory collects country policy initiatives, research and indicators relating to artificial intelligence. Its purpose is to give policymakers a shared place to compare approaches and definitions. That is useful work — and easy to mistake for something it is not. A policy inventory shows what governments have adopted. It does not show whether rules are enforced, whether they work, or whether the people affected by AI systems would recognise the picture it paints.
Measuring activity is not measuring governance
Governance has at least three layers: what is written down, what is implemented, and what actually happens to people. An observatory is strong on the first, partial on the second, and largely blind to the third. Counting national AI strategies tells you that strategies exist. It cannot tell you whether a hiring algorithm was audited, whether a benefits system was explained to the person it rejected, or whether a model’s errors fell evenly across groups.
What a good governance measure would need
If the goal is to know whether AI is governed well, a few properties seem necessary. The measure would need to look at outcomes, not just policies. It would need to be comparable across countries without flattening very different legal traditions. It would need independent verification, so that self-assessment is not the only input. And it would need to cover the private sector, where most consequential AI is deployed, not only the public sector that reports to international bodies.
Why this is genuinely hard
None of those requirements is easy. Outcomes are contested: a system that speeds up decisions for most people may slow them down for the few who need exceptions. Comparability runs into different definitions of privacy, fairness and liability. Independent verification requires access that firms rarely grant. The difficulty is a reason to be precise about what any given indicator actually captures.
Used well, an inventory is a map of where to look. If a country has passed an AI law, the next question is how it is enforced and by whom. If it has published an ethics framework, the question is whether it changed any practice. The observatory gives you the first fact; the reporting and research it links to give you the rest. Treating the inventory as a scoreboard would misread its design.
The gap that remains
There is a real risk that measurement becomes a substitute for action: governments point to adopted principles, firms point to voluntary commitments, and the hard questions about enforcement and remedy stay unanswered. Closing that gap requires the unglamorous work of audit, complaint mechanisms and transparency — the parts of governance that produce evidence rather than announcements.
The enforcement gap
Between a published principle and a changed practice sits enforcement, and enforcement is where most governance frameworks are thinnest. A law without a regulator, a regulator without resources, or a remedy without a route to complain all leave the same result: rules that exist on paper. Measuring governance honestly means asking who can hold a deployer of AI to account, and what happens when they are wrong.
Credible measurement would combine three things: independent data, comparable indicators and a focus on outcomes. Independent data means not relying solely on self-reporting. Comparable indicators mean agreeing on definitions, which is slow and political but unavoidable. Outcome focus means tracking what happens to people — errors, appeals, remedies — rather than what has been announced. None of this is easy, and that difficulty is the honest finding.
The comparison problem
Comparing governance across countries runs into a deeper difficulty than data gaps. Legal systems differ in what they regulate and how they define harm. A framework built on individual rights, a framework built on risk management, and a framework built on administrative guidance may all be serious efforts, yet they produce very different indicators. A single scoreboard tends to reward the approach its designers understand best. The honest response is to compare within families of similar systems, and to read the detail rather than the rank.
Voluntary commitments and their limits
Much AI governance currently rests on voluntary commitments: firms publish principles, sign declarations and appoint ethics boards. These can be genuine, and they can also be a way to pre-empt binding rules. The test of a voluntary commitment is whether it changes behaviour when it is inconvenient — whether a model is delayed, a feature withdrawn, or a finding published against the firm’s interest. Without such moments, a commitment is hard to distinguish from marketing, and any measure that counts commitments rather than consequences will overstate how well AI is governed.
This report draws on the OECD AI Policy Observatory.
Image: Prompt by JPxG, model by Boris Dayma, upscaler by Xintao Wang, Liangbin Xie et al. · Public domain · via Wikimedia Commons.