AI Sovereignty, Democracy, and the Club That Left Britain Out
The guardrails are not being voted on. They never were. What is new is who is no longer in the room.
There is a phrase that gets repeated in every AI policy summit, every white paper, every op-ed calling for "responsible governance": we must govern this technology democratically. It sounds unimpeachable. It is also, on inspection, mostly untrue — and worth being honest about why.
The guardrails are not being voted on
The frontier of AI — the models capable enough to matter geopolitically — sits inside a handful of companies, which sit inside one or two jurisdictions. The rules that govern what those models can do, who can access the chips that train them, and which countries get cut off from the supply chain are decided by executive action: export control lists, safety frameworks issued by agencies, bilateral arrangements between a small number of governments. None of this passes through a legislature in any meaningful sense, let alone a referendum. Calling it "democratic governance of AI" is closer to branding than description.
This is not a scandal so much as a pattern. It is the same pattern that produced the postwar order.
We have seen this movie before
Every consequential global regime of the last eighty years was built the same way — a small number of capable states deciding the rules, everyone else ratifying or complying:
- Bretton Woods (1944). Forty-four nations sat in the room, but the monetary architecture that followed was substantially the product of one negotiation — Harry Dexter White for the United States against John Maynard Keynes for Britain — with Washington winning most of the substance.
- The Nuclear Non-Proliferation Treaty (1968). It legitimized exactly five nuclear states — the US, the Soviet Union, Britain, France, China — and froze everyone else out of possession, permanently. That club has not admitted a new member since.
- The UN Security Council. The same five countries have held veto power since 1945. There is no credible path to expanding it.
- COCOM and its successor, the Wassenaar Arrangement. A small coalition of Western states decided, unilaterally, what technology could flow east — first to the Soviet bloc, later to whoever the coalition deemed a risk.
- The SWIFT financial architecture. A handful of allied governments effectively decide who remains inside the global payments system and who gets severed from it.
In every one of these, the language of shared, rules-based order coexisted with a reality of concentrated decision-making. The honest defence was never "everyone gets a vote." It was: the deciders remain accountable to their own electorates, and the arrangement can in principle be reversed. Whether that is a sufficient standard of legitimacy is the real argument — not whether the process is literally democratic, because it never has been.
The novelty: Britain is on the outside
Here is what is different this time. In every regime listed above, the United Kingdom was in the room. Nuclear club — founding member. UN Security Council — permanent seat. Bretton Woods — co-author of the framework, even while losing the argument. COCOM, Five Eyes, NATO's nuclear planning group — Britain was there by design, a function of the "special relationship" that guaranteed it a chair regardless of its relative economic weight.
AI is the first regime of this kind where that seat does not exist. For the first time since 1945, alliance history is not enough to buy Britain a place at the table where the guardrails are actually written.
AI is the first regime of this kind where that seat does not exist. Britain hosted the Bletchley Park AI Safety Summit. It has stood up an AI Safety Institute and positioned itself as a convenor of the conversation. But convening is not deciding. The UK has no frontier lab operating at the scale of OpenAI, Anthropic, or Google DeepMind's parent company under its effective jurisdiction, and no leverage over the semiconductor chokepoints — Taiwan's foundries, American export controls — that actually determine who can train the next generation of models.
For the first time since 1945, alliance history is not enough to buy Britain a seat at the table where the guardrails are actually written. What buys the seat now is compute and model ownership — infrastructure, not history. That is a genuinely new kind of exclusion, and it says something larger about where power has actually migrated in the twenty-first century: away from the diplomatic and military architecture of the last century, toward whoever owns the physical and computational substrate of intelligence itself.
Sovereignty as the fallback
There is a narrower, more honest version of "AI sovereignty" that does hold up under scrutiny — not democratic co-authorship of the global rules, but the right of a state to build or secure its own compute, its own data infrastructure, and its own models so that it is not simply a rule-taker in someone else's regime. That is a real project, and one open to far more countries than the five or six writing the actual safety and export frameworks. It is not a seat at the table. It is the ability to build a different table.
That distinction — global rule-setting versus sovereign opt-out — is probably where the debate should actually be fought. Everything else is vocabulary doing work that the institutions themselves are not.
If you read this far, there is more. Essays on geopolitics, capital, and culture — directly to your inbox. Subscribe free on Substack →
michelerovatti.com · Writing since 2012
The Dream of Never
Culture, Capital, and the Refusal of Limits
by Michele Rovatti
A sharp diagnosis of the growth imperative and the cultural logic that made it inevitable — from theology to AI.
📘 Get your copy on Amazon