What can we learn from the failure of the Baruch Plan?

The Baruch Plan didn't fail because the idea was wrong. It failed because of how it was negotiated.

In June 1946, the United States — holding a nuclear monopoly — proposed to surrender that advantage to international control through the Baruch Plan. The plan passed the UN Atomic Energy Commission 10-0 (with two abstentions). It died only at the Security Council, where the Soviet veto killed it.

For eight decades since, we've lived with the consequences: an arms race, a cold war, and nuclear weapons 3,000 times more powerful than Hiroshima. If we're alive today, it's largely thanks to the coordination of security agencies and a significant dose of luck — not treaty architecture.

We cannot afford to repeat those mistakes with AI.

Why the Baruch Plan Failed

Three critical failure modes doomed the original plan:

1. Insufficient Diplomatic Bandwidth

The 1946 negotiations lacked the communications infrastructure needed for trust-building at speed. Delegates met periodically, exchanged formal documents, and navigated linguistic barriers through limited translation services. By the time serious negotiations gained momentum, the Soviets were months from their own bomb. The window closed before real dealmaking could occur.

Today's AI timelines are even more compressed. We cannot afford Geneva-pace diplomacy when frontier labs operate at Silicon Valley speed.

2. No "Realist" Treaty-Making Model

The Baruch Plan operated within the UN Security Council's veto-based system — guaranteeing paralysis when any permanent member objected. The Soviets, fearing they would be outvoted by liberal democracies in the proposed governance structure, exercised their veto. The process provided no path around this obstruction.

Traditional multilateralism — with its veto-based deadlocks, endless negotiations, and toothless accords — has repeatedly failed. Climate agreements took decades and remain largely unenforceable. Nuclear treaties prevented small countries from developing small arsenals but left superpowers' world-ending stockpiles untouched.

3. A Political Coalition Too Narrow

President Truman supported the plan, but key advisors (including some military leaders) quietly undermined it. Without a broader coalition of aligned influencers, the proposal lacked the political momentum to survive institutional resistance.

Our Two-Part Solution

We believe that realizing a "Baruch Plan for AI" — properly and in time — requires addressing both failure modes simultaneously:

Part 1: Ultra-High-Bandwidth Diplomatic Communications

The critical first step is US-China bilateral negotiations with radically enhanced diplomatic bandwidth. This means:

Secure Infrastructure for Ultra-High Bandwidth Diplomatic Communications. Recent leaks of high-stakes diplomatic calls — including the November 2025 Witkoff-Ushakov transcripts intercepted via WhatsApp — expose critical vulnerabilities in current US diplomatic communications.

Our proposal includes rapidly deployable, mutually trusted containerized Secure Compartmented Information Facilities (SCIFs), manufactured under joint US-China-neutral oversight, that enable the spontaneous, high-frequency, leak-proof communications that high-stakes AI treaty negotiations demand.

AI-Powered Real-Time Translation. Astounding advances in real-time AI-powered translation, if made trustworthy for accuracy and confidentiality, could enormously increase the bandwidth of communications among negotiating teams.

Physical Proximity and Intensity. Hundreds of negotiators and special envoys working 3 weeks per month in secure facilities — with parallel teams in DC, Beijing, and a neutral venue like Singapore. This enables the trust-building, social bonding, and continuous multi-level channels that high-stakes dealmaking requires.

Part 2: A "Realist" Global Constitutional Convention

We propose a treaty-making model based on proven historical success — the intergovernmental constituent assembly that produced the US Constitution in 1787.

OpenAI CEO Sam Altman suggested in March 2023 that this process should be the "platonic ideal" for building global AI governance. We agree — and we're working to turn that ideal into reality.

Key features of a Realist Global Constitutional Convention on AI:

Time-bound with hard deadlines. 2-3 months maximum for the convention phase. No UN-style eternal negotiations that outlive the negotiators. When the clock runs out, there is a vote seeking consensus, then supermajority, then (if needed) simple majority.

Veto-free by design. Supermajority rules mean no single nation — not even the US — can torpedo humanity's future over a pet grievance. The UN Security Council's veto killed the Baruch Plan; we won't make that mistake again.

Weighted voting reflecting power realities. GDP, population, and technological expertise determine influence. The US and China together command roughly 25-35% of votes — enough to lead, never enough to dictate alone. This makes it a realist constitutional convention, not a utopian exercise.

Subsidiarity principle. Power stays local where possible, global only where necessary. This prevents fears of a monolithic world state while ensuring coordination where it truly matters.

Unprecedented intensity. Inspired by Philadelphia 1787, where fifty-five delegates convened for four months to forge the US Constitution, we envision similar intensity amplified by 21st-century technology. Dedicated facilities with 2,000+ delegates working full-time, leveraging secure AI infrastructure for consensus-building, real-time treaty modeling, and automated consistency checks.

Historical Precedents for Success

The intergovernmental constituent assembly model has a strong track record. Notable successes include:

  • Swiss Confederation (1848) — Transformed a loose alliance into a stable federation

  • US Constitutional Convention (1787) — Created American federalism in four months

  • German Federal Republic (1949) — Built democracy from occupation

  • European Union development — Treaty of Rome (1957), Maastricht Treaty (1993)

The 1787 US Constitutional Convention is particularly instructive. Five years after the Articles of Confederation proved inadequate against existential threats, a few states initiated a constituent process. The Constitution was ratified if endorsed by nine of thirteen states — not unanimity. This supermajority approach allowed progress despite holdouts.

Every Failure Mode Is Addressable Today

Bandwidth? Advances in secure digital communication, AI-enhanced translation, and remote deliberation enable unprecedented diplomatic intensity. We can operate at Silicon Valley speed, not Geneva pace.

Treaty-making model? The constituent assembly approach is proven. It's faster, veto-free, and creates space for the intense dialogue needed to converge disparate worldviews.

Verification? 1946 offered no way to confirm compliance without intrusive inspections. Today, satellite imagery, compute monitoring, and AI-assisted inspection make verification feasible.

Coalition? The persuasion campaign we're conducting specifically targets this failure mode, working to align key influencers before negotiations begin.

Responding to Academic Skepticism

Leading scholars have analyzed the Baruch Plan's failure and drawn cautionary lessons for AI governance. Their arguments deserve serious engagement — and serious rebuttal.

The Dafoe-Zaidi Analysis

In their influential 2021 paper "International Control of Powerful Technology: Lessons from the Baruch Plan for Nuclear Weapons," Allan Dafoe (now Head of Long-Term AI Strategy at Google DeepMind) and Waqar Zaidi provide the most thorough academic analysis of this history. Their core argument: the Baruch Plan failed because it was perceived by the Soviets as designed to "lock in U.S. advantage" rather than create genuinely equitable governance.

They note that the plan required the USSR to accept inspections and halt weapons development before the US would surrender its nuclear monopoly — terms no rival power could reasonably accept. The lesson they draw: "schemes for international control of dual-use technologies are only likely to succeed if all relevant stakeholders perceive them to be earnest and equitable."

Our response: We agree entirely with this diagnosis — which is precisely why our proposal differs fundamentally from the original Baruch Plan. Our treaty framework features:

  • Weighted voting capped at 25-35% for US + China combined — enough to lead, never enough to dictate

  • Joint development of enforcement infrastructure under neutral oversight, not unilateral US control

  • Simultaneous commitments rather than sequential surrender by one party first

The Dafoe-Zaidi paper itself acknowledges that "radical levels of cooperation become more feasible in light of truly existentially dangerous" technologies. That's exactly the situation we face with ASI.

The RAND Corporation Assessment

A May 2025 RAND report, "Insights from Nuclear History for AI Governance," reaches even more pessimistic conclusions. The authors argue that AI governance modeled on nuclear nonproliferation is unlikely to succeed because:

  1. No "AI stability" exists comparable to nuclear MAD — the strategic balance that made the NPT possible

  2. Too many actors — dozens of states and private companies, unlike the bilateral US-Soviet dynamic

  3. The Baruch Plan "might have been worse than nothing" — poisoning later cooperation attempts

Our response: These arguments misread both history and our proposal.

On "AI stability": The authors conflate conditions for success with preconditions for attempt. The NPT emerged 22 years after the Baruch Plan's failure — but that doesn't mean the Baruch Plan was wrong to try. Nuclear scientists in 1946 couldn't wait for "nuclear stability" to emerge; they understood the window was closing. We face the same urgency. The conditions for cooperation must be created, not awaited.

On "too many actors": This is precisely why we propose a US-China bilateral emergency accord first, followed by a broader multilateral convention. The RAND authors acknowledge that the NPT succeeded through "collusion" between the two superpowers, who then drew other nations into alignment. Our two-phase approach follows this exact logic.

On "worse than nothing": The RAND report suggests the Baruch Plan's failure made later cooperation harder. But this confuses correlation with causation. What made cooperation harder was the arms race that followed — not the attempt to prevent it. The counterfactual where Truman never proposed international control is not a world of easier cooperation; it's a world where the very concept of nuclear governance was never planted.

What The Skeptics Miss

Both analyses share a crucial blind spot: they treat the Baruch Plan's process failures as evidence that the goal itself was flawed.

The 1946 negotiations failed because:

  • Diplomatic bandwidth was catastrophically insufficient

  • The treaty-making process operated within the UN's veto-based structure

  • The timeline couldn't match the urgency

These are process problems with process solutions. Ultra-high-bandwidth secure communications, a veto-free constitutional convention model, and hard deadlines directly address what actually went wrong — not what skeptics assume went wrong.

The Dafoe-Zaidi paper itself endorses the Baruch Plan "as an inspiration" for AI governance. The question isn't whether international control is desirable — even the skeptics agree it is. The question is whether it's achievable. That depends on learning the right lessons from 1946.

Learn More