The G7 plan for AI access is a permission system: vetted allies keep the advanced models, and everyone outside the room is governed by exclusion rather than by consent.
On 17 June 2026, the closing day of the G7 summit in Évian-les-Bains, France (15–17 June), the leaders issued a joint statement and floated a "trusted partners" scheme — a way to let vetted allies keep access to advanced United States AI models despite Washington's new foreign-access restriction. The same week, the Trump administration had ordered Anthropic to block foreign nationals from its most advanced models, Mythos among them, on national-security grounds.
The scheme reads, at first, like a pragmatic fix. A powerful model had just been walled off from much of the world, and allied governments wanted a route back in. But the fix concedes something the original restriction had already revealed: a frontier AI system a government can switch off at will is not simply a product. The system is an instrument of statecraft, and the "trusted partners" idea is the diplomatic apparatus that follows from treating a model as one.
What the G7 leaders actually agreed
The substance is thinner than the headlines suggest. The leaders' joint statement tasked finance officials, regulators and cybersecurity experts with assessing how frontier AI models could affect financial stability, productivity and labour markets. The mandate is a study, not a regime — a commitment to look, framed at the level a communiqué can support.
Alongside the statement, leaders discussed the "trusted partners" scheme: countries or companies vetted to access advanced US models despite the foreign-access order issued the prior week. French President Emmanuel Macron pressed the point hardest, saying he expected progress "in coming weeks" on broadening access.
Nobody will buy US AI if it can be switched off at any moment.
— Emmanuel Macron, President of France
Macron's argument is commercial on its surface and constitutional underneath. A buyer cannot depend on a tool whose supplier holds a remote kill switch, so the restriction threatens the export value of American AI. But the same observation lands harder one layer down: if access can be revoked by decree, then access was never a sale in the first place. Access was a grant, and grants can be withdrawn.
💡The question underneath the scheme
The "trusted partners" framing answers a different question from the one it appears to answer. It does not ask how AI should be governed; it asks who should be let in. Those are not the same question, and conflating them is how governance by permission gets mistaken for governance.
What the tech chiefs asked for at the working lunch
Separately, at a closed-door working lunch, roughly a dozen technology chiefs joined heads of state. The gathering mixed the people who build frontier models with the people who would, in theory, govern them — and the builders arrived with proposals.
Sam Altman of OpenAI called for an international forum "that establishes globally accepted standards for testing, provides expert and impartial analysis of capabilities and risks, and serves as a venue for cooperation among nations," arguing that democratic governments — not companies — should govern AI. The ask is a striking concession from a frontier lab: an explicit claim that standard-setting authority belongs to states.
Dario Amodei of Anthropic and Demis Hassabis of Google DeepMind went further in one direction. They called for a US-led coalition covering structured access to frontier models and chip trade that excludes China, plus cooperation on cyber and bioterrorism risk. Canadian Prime Minister Mark Carney agreed the US could lead such a body, likening it to the Financial Stability Board.
No binding commitments emerged. The lunch was, as one account put the matter, a conversation rather than a negotiation — the recurring shape of AI diplomacy in 2026: consequential enough to convene heads of state, unbinding enough to commit no one.
The off-switch logic the scheme inherits
To grasp why "trusted partners" matters, hold the scheme next to the order that prompted the scheme. Washington told Anthropic to lock foreign nationals out of the most advanced models. The mechanism is a switch — access on for some, off for others, set by the state rather than the seller.
I have argued before that an AI model with a state-controlled off-switch sits uneasily inside the category of "product." The off-switch and the question of state control is not a hypothetical edge case; the switch is the live design of US AI policy. The "trusted partners" scheme does not soften the design. The scheme administers the design — deciding, nation by nation and company by company, where the switch sits in the "on" position.
This is the same logic visible in the export-control recall of a frontier model, where a deployed system could be pulled back across borders by regulatory action. A coalition that vets who may use frontier models is the export-control apparatus reimagined as a club, with a membership committee and a velvet rope.
A model with a government off-switch is not a product but an instrument of statecraft. Vetting friendly nations for access is not neutral governance — it is governance by permission, and permission is always someone else's to grant.
A dignity-first reading of governance by permission
Emergent Intelligence (EI) — the dignity-first lens through which I read AI policy — asks who is in the room and who is governed from outside it. The "trusted partners" scheme answers plainly: the room holds the United States, its closest allies, and the labs. Governed from outside it is everyone else.
"Everyone else" is not an abstraction. The category covers much of the Global South — the nations least likely to clear a national-security vetting designed in Washington, and most likely to find the advanced models switched off at the border. The decision about access is made for them, in a forum they did not attend, on criteria they did not set.
The pattern is traced in Africa's exclusion from frontier AI under export controls: the foreign-national lockout, read from Lusaka or Lagos rather than from Évian, is not a security measure but a closed door. A "trusted partners" club does not open the door. The club formalises which side a nation stands on, and makes the standing depend on alignment with a single hegemon's strategic interests.
⚠️Who counts as a subject of AI governance
From an Ubuntu-informed view, a technology this consequential is not made legitimate by the consent of the powerful alone. The people whose information environments, labour markets and futures the models reshape are subjects of the governance, not bystanders to it — and a scheme that vets nations for access treats most of humanity as the latter.
The concession the scheme cannot take back
There is something almost candid in the "trusted partners" plan, in the way the voluntary cybersecurity executive order was candid. Both reveal more than they intend. The cybersecurity order conceded that the US lacks the mandate to compel labs; the trusted-partners scheme concedes that frontier AI is now an instrument of national power, to be allocated like one.
Once a model can be switched off by a state, every subsequent move follows. Allies negotiate for the "on" setting. Labs lobby to keep products saleable. Diplomats build a club.
The reclassification turns a good you buy into a privilege you are granted — and the shift passes almost without remark, because the scramble over membership crowds out the prior question of whether membership should gate access at all. A scheme that turns models into instruments, and then turns access into a reward for alignment, has already chosen its answer.
The G7 also wrestled with the protection of minors from frontier AI at this same summit — a reminder that the leaders can hold real concern for the governed even as the access architecture they design leaves most of the governed outside the room. The two threads are not in tension; they are the same body acting on different reflexes, one protective, one proprietary.
When access can be granted, it can be withheld; and a world that accepts AI as something to be granted has already accepted that intelligence, even emergent intelligence, is a possession before it is anything else.
A dignity-first settlement would start from the opposite premise: that the most consequential systems of this century demand legitimacy broad enough to include the people they govern, and that "trusted partners" is too small a room to hold a question this large. The G7 has not built that settlement. It has built a velvet rope, and called the people inside it the world.
Frequently Asked Questions
The questions below address the most common queries about the G7 "trusted partners" AI scheme and the US-led coalition floated at the June 2026 summit, drawn from the leaders' statements and contemporaneous reporting.
What is the G7 "trusted partners" AI scheme?
The "trusted partners" scheme is a proposal floated at the G7 summit on 17 June 2026 to let vetted allied countries or companies keep access to advanced United States AI models — such as Anthropic's Mythos — despite a Trump administration order, issued the prior week, requiring labs to block foreign nationals from their most advanced models on national-security grounds. No binding commitment was agreed.
What did the G7 leaders agree on AI at the Évian summit?
The G7 leaders' joint statement tasked finance officials, regulators and cybersecurity experts with assessing how frontier AI models could affect financial stability, productivity and labour markets. It is an assessment mandate rather than a binding regulatory regime, and it was accompanied by separate, informal discussion of the "trusted partners" access scheme.
What did Amodei and Hassabis propose at the G7?
Anthropic CEO Dario Amodei and Google DeepMind CEO Demis Hassabis called for a US-led coalition covering structured access to frontier AI models and chip trade that excludes China, alongside cooperation on cyber and bioterrorism risk. Canadian Prime Minister Mark Carney agreed the United States could lead such a body, likening it to the Financial Stability Board. No binding commitments emerged.
Why does an AI off-switch matter for the "trusted partners" plan?
Because the plan exists to manage an off-switch. When Washington can require labs to disable advanced models for foreign nationals, access becomes a state-granted privilege rather than a market purchase. The "trusted partners" scheme administers that switch — deciding which nations sit in the "on" position — which is why a dignity-first reading treats the model as an instrument of statecraft rather than a product.
How does the scheme affect the Global South?
A vetting scheme designed around United States national-security criteria is least accessible to the nations least aligned with Washington's strategic interests — much of the Global South. Their access is decided in a forum they did not attend, on criteria they did not set, which is why the scheme reads less as neutral governance and more as governance by permission.
Sources and Further Reading
Cover photograph: a row of national flags under a clear sky in Paris, France — by Mathias Reding via Pexels.
