The Limits of a Coalition of Democracies
Jordan Ryan
June 13, 2026
Image: Lightspring / shutterstock.com
In June 2026, Dario Amodei, the chief executive of Anthropic, made an important argument. In his essay Policy on the AI Exponential, he calls for governments to have the authority to prevent some Artificial Intelligence (AI) systems from reaching the market. His proposal would require third-party testing of powerful “frontier” models above a defined compute threshold and give the state power to block or reverse deployment where independent assessors judge a model to pose unacceptable risk in four areas: cybersecurity, biological weapons, loss of control, and automated research that could accelerate the other three. Coming from one of the industry’s most prominent safety advocates, this deserves close attention, not only for what it recommends but for what it implies.
For several years, Anthropic and its peers argued mainly for transparency: disclosure of safety practices, reporting of incidents, and visibility into emerging risks. In 2025, Anthropic supported transparency legislation in California, New York and Illinois on that basis. Amodei now maintains that transparency is no longer enough, and that binding regulation is required. His preferred analogy is aviation: frontier systems, like aircraft, should undergo independent testing before release and be grounded if they fail. When a developer with a strong commercial interest in avoiding regulation presses instead for stronger public oversight, the question is no longer whether frontier AI warrants governance; it is what kind.
It is here that Amodei’s essay is least convincing. He invokes the nuclear analogy repeatedly, warning that AI may reshape international security even more profoundly than nuclear weapons did. Yet when he turns to the international level, the institutional lesson of the nuclear age largely disappears. His proposed solution is a coalition of democracies that coordinates regulation, shares advanced chips and semiconductor technology among trusted partners, and restricts access for adversaries. Universal participation appears as a longer-term aspiration, not an immediate design principle. In practice, it is a governance model built around a closed group of states.
The risks Amodei identifies do not stop at coalition boundaries. A biological weapon designed with the assistance of a frontier model threatens populations regardless of their governments’ alignment. A catastrophic failure of control would not distinguish between members and non-members. Such failures could also destabilise fragile states, fuel violent conflict and undermine crisis-management efforts. Risks that are global in consequence require governance that is global in reach. That is a matter of effectiveness as much as fairness.
A system that excludes large parts of the world will struggle to achieve the two things any serious risk-management framework requires: verification and legitimacy. Verification requires visibility where the risks actually sit, not only where political trust already resides. Legitimacy requires that those expected to comply have had some part in shaping the rules. A coalition satisfies neither easily.
The nuclear precedent makes the omission sharper. The nuclear age produced alliances, but it also produced institutions. The Treaty on the Non-Proliferation of Nuclear Weapons and the safeguards system of the International Atomic Energy Agency were designed for rivals as well as friends. Imperfect and contested, they rested on a principle worth recovering: a technology of universal consequence requires institutions capable of commanding broad participation.
There is a serious counter-argument, which the essay does not make. Verification in the biological and cyber domains has always been difficult. The collapse of negotiations on a verification protocol for the Biological Weapons Convention in 2001, when the United States rejected the draft text and walked away from a decade of work, showed how hard it is to design inspection systems that are effective and politically acceptable. But that history points to investing in verification design, not to abandoning it. If verification proves difficult, that is a reason to begin early rather than postpone.
The developing world is largely absent from the essay’s map. It appears once, as a recipient of benefits that a coalition might choose to distribute. Yet many developing countries are among the most exposed to the biological and cyber risks that advanced AI presents. Health systems with limited surge capacity, critical infrastructure with weak cyber defences, and economies dependent on a few digital platforms are vulnerable. A system designed by advanced economies and presented to the rest of the world is unlikely to command the cooperation that effective governance requires. This carries a particular cost for democracies. A coalition that draws its authority from democratic legitimacy weakens that claim when it builds governance to exclude, and the standing of democracies as rule-makers rests on writing rules that others are able to participate in creating.
None of this diminishes what Amodei has done. He has conceded that transparency by itself is insufficient, and that governments should be able to constrain the deployment of dangerous systems, a marked departure from the industry’s position. The weakness lies in the international half of the proposal. If frontier AI carries risks on the scale he describes, governance cannot stop at the boundary of a single democratic coalition.
Four steps would begin to close that gap.
- Convene a dedicated United Nations General Assembly process on frontier AI verification before the next major leap in capability.
- Seat the states most exposed to biological and cyber risk at the centre of setting testing and reporting standards.
- Draw on decades of arms-control and non-proliferation experience to design verification mechanisms that can function during periods of rivalry and low trust.
- Lodge authority to certify, and where necessary restrict, the most dangerous systems in a body whose legitimacy extends beyond any single coalition, with the United Nations working alongside regional organisations and independent assessors.
Amodei has the scale of the challenge right. Coordination among democracies, though, is only the beginning. The harder task is to build institutions broad enough to govern dangers that fall on everyone, and legitimate enough that those exposed to them have helped to write the rules. That is the limit of a coalition of democracies. The nuclear age built institutions of broad reach the hard way, after the danger was already present; the age of AI has the chance to build them in advance.
Other articles from this author:
Governing the Ungovernable (3-minute read)
The Secretary-General This Moment Demands (10-minute read)
From Reform to Reinvention: Reimagining the United Nations for the 21st Century (10-minute read)
The UN’s Withering Vine: A US Retreat from Global Governance (3-minute read)
