Britain sat at Yalta in 1945. On November 19, Donald Trump, Xi Jinping and Vladimir Putin met in Shenzhen. That same day, Britain was at a different table. Its prime minister joined Ukraine, European leaders and American envoys in Berlin to discuss a possible peace and the guarantees it would require.
A day later, Britain and the European Union announced plans for a digital partnership that could include AI safety and security. The three meetings bring a question into focus that will matter as artificial general intelligence becomes a source of power. Who gets to write the rules that others will have to live under?
In August, I argued that the United States and China should initiate a limited, verifiable agreement on frontier AI, open to other countries from the start. The historical model was Robert Schuman’s decision to begin with coal and steel, the material foundations of industrial power. An agreement on frontier AI could begin with similarly concrete obligations. Report serious incidents. Allow independent evaluations of safeguards. Set expiration dates for emergency powers.
Munich supplies a second lesson. In 1938, the powers assembled there agreed to transfer Czechoslovak territory in Czechoslovakia’s absence. I propose a Munich Clause for AGI. Countries whose security, access to computing capacity or legal obligations would be materially affected by an agreement should be able to help write its rules from the start. That gives countries across the Global South a role in shaping the agreement and a reason to uphold it.
The same principle applies within democracies. When an AI system materially affects a person’s life, that person should be able to find out who authorized the decision, obtain human review and seek a remedy. Those protections must reach ethnic, religious and other minorities in practice. Future generations must retain the right to revise the rules they inherit.
The United States has a strategic interest in that outcome. It can translate its technological lead into agreements that other governments have a stake in enforcing. And the American debate can begin during the presidential primaries. Here is one question I would put to any candidate. Would you give an independent authority the power to audit an AI deployment your own administration had approved, and to suspend it if the safeguards failed?
