In 1946 the United States was the only country with nuclear weapons, and for a short time it seemed possible that no one else would ever need to build them. In June, the American financier Bernard Baruch presented a plan to the newly formed United Nations. Atomic energy, in all its dangerous forms, would be placed under an international authority. That body would control the mining, the materials, and the technology worldwide, inspect for violations, and ensure the science was used for power and not for bombs. Once the system was in place, the United States would give up its atomic arsenal.
It was, on paper, an extraordinary offer: the sole nuclear power proposing to surrender its monopoly to international control rather than exploit it. It failed. Within a few years the Soviet Union had the bomb, and the arms race that the plan was meant to prevent ran for the next four decades and never fully ended.
Why it collapsed
The failure was not a matter of nobody caring. It came down to conditions that governance of a strategic technology has to satisfy and that the Baruch Plan could not.
- Trust and sequencing. The plan required the Soviet Union to accept intrusive international inspection and forgo its own weapons program first, while the United States kept its bombs until the system was fully working. Moscow, unsurprisingly, would not accept a scheme that locked in American advantage during the transition.
- Enforcement and the veto. The plan proposed removing the great-power veto for atomic matters, so violations could be punished. The Soviet Union, unwilling to be outvoted on its own security, refused. Real enforcement and national sovereignty collided, the same tension examined in our piece on the sovereignty objection.
- The race had already started. Even as the plan was debated, both powers were pursuing the underlying capability. Governance was chasing a competition that was already underway.
The parallels to AI, stated carefully
The resemblance is uncomfortable. A transformative and dangerous technology arrives. For a brief period the field of players is small and the situation is, in principle, still governable. A proposal exists to bring the technology under collective control before an unrestrained race locks in. And the effort runs into mutual distrust, the difficulty of verified enforcement, and the fact that the competition is already moving. Change atomic to AI and the shape is familiar, which is why the Baruch Plan haunts serious discussions of AI governance.
The Baruch Plan is the story of a window that was open, briefly, and then was not. The lesson is about timing.
Two lessons, not one
The easy reading is despair: international control of a strategic technology was tried at the best possible moment and failed, so why expect better for AI. That reading is incomplete.
The first real lesson is about the window. The Baruch Plan failed partly because it came slightly too late, after distrust had hardened and both sides had committed to the race. The window for governing a technology is widest before the competition is entrenched, and it closes. For AI, that window is now, not after the race has fully set, which is the urgency behind everything the Foundation does.
The second lesson is about design. The Baruch Plan foundered on verification and enforcement, on the impossibility, in 1946, of trusting compliance without a way to check it. That is precisely the problem that modern tools, from verification regimes to hardware-based monitoring, exist to solve, and it is why the later arms-control treaties that came after the Baruch Plan's failure succeeded where it could not. The plan did not prove governance is impossible. It showed which problems have to be solved, and that they must be solved early. Our answer to both is our plan.
Common questions.
The Baruch Plan was a 1946 United States proposal to the United Nations to place atomic energy under international control. An international authority would control the mining, materials, and technology of atomic power worldwide, inspect for violations, and ensure the science was used peacefully, after which the United States, then the only nuclear power, would give up its weapons. It was an offer by the sole atomic power to surrender its monopoly to international control, and it was not adopted.
For several linked reasons. It required the Soviet Union to accept intrusive inspection and forgo its own weapons program while the United States kept its bombs until the system was working, which Moscow would not accept. It proposed removing the great-power veto for atomic matters so violations could be punished, which the Soviet Union rejected as a threat to its security. And the arms race had effectively already begun, so governance was chasing a competition that was already underway.
The situations rhyme. A transformative and dangerous technology emerges; for a short period the set of players is small and the situation is still governable in principle; a proposal exists to bring it under collective control before an unrestrained race locks in; and the effort runs into mutual distrust, the difficulty of verified enforcement, and a competition that is already moving. That structural resemblance is why the Baruch Plan is frequently cited as a cautionary precedent in debates about governing AI.
No. It carries two lessons. One is about timing: the plan failed partly because it came slightly too late, after distrust had hardened and both sides had committed to the race, which shows that the window to govern a technology is widest before competition is entrenched and then closes. The other is about design: it foundered on the 1946 inability to verify and enforce compliance, exactly the problem that modern verification tools exist to solve and that later arms-control treaties did solve. The plan showed which problems must be addressed, and that they must be addressed early, rather than proving governance impossible.