OpenAI's CEO wants the UN for AI, but with America holding the veto pen and his company writing the rulebook.
The Summary
- Sam Altman proposed an international AI safety framework modeled after nuclear watchdog organizations, positioning it as necessary for global AI governance
- The Financial Times frames it more cynically: a US-led order that conveniently advantages the American AI oligopoly, including OpenAI itself
- The proposal arrives as AI capabilities race past any existing regulatory framework, with the winners trying to write the rules before anyone else can
The Signal
Altman's pitch sounds reasonable on paper. Model AI governance after nuclear oversight bodies, create international cooperation, prevent catastrophic outcomes. The nuclear comparison works because both technologies have existential risk profiles. Both can end civilizations if mishandled. Both require coordination that transcends borders.
But here's what the nuclear framework actually looked like: the US and USSR built arsenals first, then created treaties that locked in their advantage. The Nuclear Non-Proliferation Treaty didn't stop proliferation. It created a club of nuclear haves and have-nots. The haves got to keep their weapons. The have-nots got inspections and sanctions.
"A US-led world order would hand a big advantage to an American oligopoly that includes his own company."
Now apply that to AI. OpenAI, Anthropic, Google, and Meta have multi-billion dollar head starts. They've already trained models on datasets nobody else can replicate. They've built infrastructure at a scale that requires nation-state resources. An international framework led by the US doesn't level the playing field. It cements the current leaders' position while making it harder for anyone else to catch up.
The timing is the tell. You don't propose international oversight when you're behind. You propose it when you're ahead and want to stay there. When Standard Oil controlled 90% of US refining, Rockefeller suddenly became very interested in industry standards and safety regulations. Funny how that works.
The proposal's likely structure:
- Mandatory model audits (expensive, favors companies with compliance budgets)
- Compute threshold limits (protects those who already scaled)
- "Safety standards" set by committee (whoever sits on the committee defines "safe")
The crypto angle matters here. If AI governance follows the nuclear playbook, it will inevitably try to regulate the compute layer and the training data. That means controlling who can build agents, who can train models, who can deploy AI at scale. It's the opposite of Web4's promise: open protocols, permissionless building, distributed intelligence.
The Implication
Watch who ends up on this framework's governing board. If it's the existing AI oligopoly plus some academics they fund, you'll know exactly what this is: regulatory capture dressed up as safety. The nuclear comparison should worry anyone building in Web4. Nuclear oversight didn't democratize atomic energy. It concentrated it.
For builders: assume the window for permissionless AI development is closing. If you're working on open-source models, decentralized training, or agent frameworks outside big tech's walled gardens, the next 18 months matter. The people who want to regulate AI are the same people who currently dominate it. That's not a coincidence.