Independent verification: what makes a country ready to govern AI?

An Editorial Note from Fathom

Independent evaluation is moving from the margins to the center of the AI governance conversation. Anthropic, OpenAI, and Google DeepMind have each, in their own way, pointed toward third-party evaluation of models. Independent Verification Organizations (IVOs) are explicitly named in the FRONTIER Act, the first federal blueprint for independent AI evaluation, and in enacted state laws, most notably California's SB 813, which Fathom sponsored and establishes the state’s framework for independent verification organizations. Policymakers and industry stakeholders have arrived at the same insight: companies cannot grade their own homework.

At Fathom, we’ve spent a lot of time making the case for IVOs in the United States. But the idea doesn’t stop at the US border, and neither does AI, or the risks and opportunities it brings. Independent evaluation is far more valuable when the systems that provide it are interoperable across countries. The US conversation matters, but it needs to go global.

There is broad consensus that advanced AI needs more oversight than exists today, and jurisdictions are experimenting with different models, mechanisms, and institutions. But what does flexible, interoperable AI governance, including independent evaluation, actually look like, and what determines whether a model can succeed from one country to the next?

At the highest levels of the global AI policy debate, independent research reveals that there are clear opportunities for governance, but such opportunities are being overshadowed by false dichotomies. Too often, potential governance mechanisms have died trying to thread the needle between economic development, innovation, and risk-fit governance. Strong conceptual buy-in, in other words, is not the same as readiness: converting that buy-in into on-the-ground adoption is complicated by a consistent set of political conditions that vary enormously from one capital to the next.

What we can say is that those conditions follow a pattern consistent enough to assess across jurisdictions. This report maps that pattern: the five factors that determine whether a government sees independent evaluation as an opportunity or an unnecessary constraint, what policymakers themselves told us about the constraints they work under, what determines if the framework travels well between countries, and how AI governance leaders can turn it from a diagnosis into a roadmap.