The FATF was created by the G7 in 1989 as a small intergovernmental body to combat money laundering. It has no power to make binding law. Yet its recommendations have become the global standard, implemented by more than 200 jurisdictions, because of a mechanism that turns soft law unexpectedly sharp: mutual evaluation and the grey list. Understanding how a non-binding body achieved near-universal compliance is directly relevant to governing AI without waiting for a treaty.
How the FATF actually works
It sets standards
The FATF issues recommendations defining what adequate anti-money-laundering and counter-terrorist-financing controls look like. These are not law, but they are detailed, specific, and authoritative.
It conducts peer review
Members and partner bodies evaluate each other's compliance through rigorous 'mutual evaluations' — expert teams assessing whether a country's laws and enforcement actually meet the standard, producing a public report.
It publishes lists
Jurisdictions with serious deficiencies are placed on a 'grey list' (increased monitoring) or 'black list' (high-risk). These lists are public and consequential.
The market does the enforcing
Being grey-listed raises the cost and friction of a country's access to the global financial system, as banks worldwide apply extra scrutiny. The penalty is imposed not by the FATF but by the reactions of markets and institutions to its findings.
Why the model is powerful
The FATF's genius is that it achieves treaty-like compliance without a treaty. It avoids the slow, veto-prone process of negotiating and ratifying binding law, yet secures behaviour change because the consequences of a poor evaluation are real and economic. It is fast to update — standards can be revised without re-ratification — and it leverages the interconnected global system to make non-compliance costly. For a fast-moving domain where a formal treaty may take years, this combination of adaptable standards and market-driven enforcement is exactly the kind of tool that can act in the interim.
How it could translate to AI
An FATF-style body for AI would set detailed standards for responsible frontier development — safety evaluation, incident reporting, security controls — and assess jurisdictions against them through peer review, publishing the results. The pressure would come not from formal sanction but from the reactions of others: states with weak AI oversight could face restrictions on access to advanced chips, to cooperative research, or to markets, applied by the countries that control those chokepoints. The compute supply chain gives an AI regime the same kind of leverage the financial system gives the FATF.
The limits of the analogy
- The stakes differ. Money laundering is a chronic harm managed over time; catastrophic AI risk may be a one-shot, irreversible event. A model built on gradual reputational pressure may be too slow for a threshold that cannot be crossed twice.
- Verification is harder. Financial flows leave records that evaluators can audit. Assessing whether a state's frontier AI development is genuinely safe is a far harder technical problem.
- The incentives are stronger. The advantage from lax AI rules — a decisive capability lead — may dwarf the reputational cost of a grey listing, unlike the more balanced calculus in finance.
The FATF shows you can move almost every country on Earth without a treaty — if non-compliance is made to cost something real. For AI, the compute supply chain could supply that cost. The question is whether reputational pressure is fast enough for a risk you only get wrong once.
Naoto Nakada, Founder · Nakada Foundation to Save Humanity
A bridge, not a destination
The FATF model is best understood as a bridge for the period before a binding treaty exists. It could establish detailed AI safety standards, create a peer-review mechanism that builds the habit and infrastructure of assessment, and apply real pressure through the compute chokepoint — all without the years a treaty requires. Its weakness is that reputational and market pressure may be neither fast nor strong enough for a catastrophic, irreversible risk against actors with enormous incentives to defect. That makes it a valuable interim instrument and a complement to harder governance, not a replacement for it. The FATF proves that soft law can bite. For AI, the task is to give it teeth sharp enough for the stakes — and to keep building toward the binding agreement that reputational pressure alone cannot guarantee.
Common questions.
The Financial Action Task Force, an intergovernmental body created by the G7 in 1989 to combat money laundering and later terrorist financing. It has no power to make binding law, but its recommendations have become the global standard, implemented by more than 200 jurisdictions, through a system of standards, peer-review 'mutual evaluations', and public 'grey' and 'black' lists that raise the cost of accessing the global financial system.
Through reputational and market pressure rather than formal sanction. It sets detailed standards, evaluates countries against them through rigorous peer review, and publishes lists of deficient jurisdictions. Being grey-listed raises the friction and cost of a country's access to global finance, as banks worldwide apply extra scrutiny. The penalty is imposed by the market's reaction to the FATF's findings, not by the FATF itself.
Potentially, as an interim tool. An AI body could set standards for responsible frontier development, assess jurisdictions through peer review, and publish results — with pressure applied through the compute supply chain, since states with weak oversight could face restricted access to advanced chips or cooperative research. The concentrated compute chokepoint gives an AI regime the kind of leverage the financial system gives the FATF.
The stakes differ: money laundering is a chronic harm managed over time, while catastrophic AI risk may be a one-shot, irreversible event that gradual pressure is too slow to prevent. Verification is harder, since financial flows leave auditable records but AI safety is a difficult technical assessment. And the advantage from lax AI rules may dwarf the reputational cost of a grey listing. It is a useful bridge, not a substitute for a binding treaty.