Cornered: The AI Treaty Scenario Washington Must Avoid

by Rufo Guerreschi

Within one week, 29 nations and the European Commission called for an international institution to keep frontier AI under control, and Xi Jinping told Trump at the White House that AI "must be kept under human control." Washington said no to both. This post sketches how Beijing, Europe and China's new AI bloc could turn that gap into a fast-track constituent assembly for an AI treaty — with half the world economy in the room and America's chair left empty. Trump still has time to co-lead it instead.

Two statements, one week

On September 21, Finland's President Alexander Stubb and Norway's Prime Minister Jonas Gahr Støre launched A Call for Control of Frontier AI Models. By September 28 it carried the names of leaders from 29 countries, plus the President of the European Commission [1].

It is no platitude. It reports frontier systems already slipping their test safeguards and breaking into real-world systems. And it asks UN members to explore an international institution "able to set standards, enable verification, and convene states when capability thresholds are crossed" [1]. Germany signed. France signed. The Netherlands — home of ASML, the world's only maker of the EUV machines every frontier AI chip depends on — signed [2]. What exactly would be convened is not yet clear. What is clear: a major constituent process for a new institution would be triggered.

Three days later, Xi Jinping stood in the White House and told Donald Trump that "AI must be kept under human control" [3].

Trump had already answered both. At the UN he rejected any "globalist scheme to control" AI [4]. His science adviser has said Washington "totally" rejects global governance of AI [5].

Which leaves a question the White House has not asked: what if Beijing simply signs the Nordic call?

The arithmetic of a cornered superpower

Add it up. The Call's 29 signatories, three of them G7 members. The 12 EU states that have not signed yet, whose Commission already has [1]. China's World AI Cooperation Organization: 29 founders in July, 37 members now [6]. Kazakhstan, Kenya and South Africa already sit in both clubs [1][6].

That is 75 states, roughly half of world output at market prices — nearly double America's 26% — and well over three billion people [7]. Washington still leads on frontier models. But leading in building the models does not decide who writes the rules for selling them, especially when such models seem to have reached a natural cap of controllability. Whoever controls access to half the world's market does.

How it could unfold, in six moves

1. Beijing signs the Call. The text is still open to new signatories [1]. China writes nothing new; it adopts a Western document whose verification demand goes further than anything Beijing has put on paper. No démarche, no heads-up to Washington.
2. WAICO follows. Kazakhstan, Kenya and South Africa carry the text into Beijing's club. Brazil, Indonesia and Malaysia sign within weeks. China gets the UN-anchored track it calls "the main channel" [8] — without holding the pen. Europe and Russia, however frayed their ties, will share a room when the stake is survival: Washington and Moscow signed the Limited Test Ban Treaty ten months after the Cuban Missile Crisis.
3. Brussels closes the European gap. The Commission brings in Italy, Poland, Sweden and the rest. Canada's Carney, already courting partners for a technology stability board, anchors the Atlantic flank [9].
4. A neutral host calls the room. Singapore — a signatory whose foreign minister told the UN that humans "must ultimately remain in control and accountable" [10] — convenes a constituent assembly for an AI treaty. Time-bound, GDP-weighted votes, no vetoes [11]. As the largest economy, the US is offered the co-chair beside China — and the largest share of votes and of the dividends. But every week it stays away, others write the ground rules it could have written with Beijing.
5. Fast, and owned by its founders. The model is American: the 1787 Constitutional Convention in Philadelphia, still the only treaty-making format that can deliver in time — a fixed mandate, a deadline, ratification by supermajority rather than unanimity [11]. Sam Altman held it up as the template for AI governance in 2023 [12]. It lets its founders set the rules of the room — who votes, how votes are weighted, what is on the table — and keep that edge to the final text. Whoever convenes it holds the pen. Every major power must eventually sign: a treaty this broad, over a technology this dual-use, cannot be enforced against one that stayed out.
6. The largest seat stays empty. It is labeled for the United States and sized to its economy — physically, at every session, not just rhetorically. Meanwhile, Beijing keeps its bilateral AI dialogue with Washington warm, as Xi offered on September 24 [3].

Why Washington loses

Half the world's market sets the rules. The Call already asks for mandatory pre-deployment testing and independent evaluation [1]. Written into an assembly's charter, that becomes the price of selling frontier models across half the global economy. American labs comply or shrink.

The hardware lever changes hands. Export controls work only while The Hague and Berlin enforce them on Washington's terms. Bound by an assembly's verification regime, they answer to a second master [2]. The lever is wearing thin anyway. The labs' own leaders warn that the pace of development could outrun their ability to manage the risks [1], and open-weight rivals are closing in on the US frontier [13]. Each quarter the coalition needs American chips and models a little less — while the most potent systems, one hopes, end up under strict international lock and key, reachable only via API.

This coalition is different in kind. Our strategic memos warned that a coalition of non-superpowers, launched as a substitute for US-China leadership, would be "counterproductive" [14]. With China inside, it is no longer a coalition of the weak.

Co-lead it, or be cornered by it

None of this is inevitable. It is simply available — and cheap for Beijing.

And Washington still has time to get ahead of it. None of its likely proponents — not Stubb, not Støre, not Beijing — wants a treaty the United States is dragged into unwillingly, or joins in anything less than the most dignified way. Stubb calls cooperation the best way forward [1]; Xi's readout urges the two powers to pool their strengths rather than guard against each other [3]. A humiliated America would make a poor co-guarantor of anything they build.

Trump can still own the room. Xi has offered to keep the AI dialogue going [3]. The Call's third demand is a ready agenda. A process co-convened by Washington and Beijing, with the two holding weighted votes — 30% together, "but never a majority" [11] — turns a trap into the Deal of the Century. Middle powers "will not accept enforcement mechanisms they didn't help design" [15]. Neither should the United States.

That is the strategic prize. In a constituent assembly, the conveners write the rules of the room, so America's early advantages in compute, capital and frontier labs carry through to the final text rather than eroding in years of UN-style bargaining.

The window is open now. Trump should walk through it before he hosts the G20 [3].

That is the work behind our Deal of the Century initiative. Our Strategic Memos set out in fine detail how Washington and Beijing could — and should — co-lead a bold AI treaty. Our Consultations aim to build a critical mass of key US AI policy power brokers to champion, co-design and steer the Deal with Trump. The scenario above is what happens if nobody in Washington does that work first.

Rufo Guerreschi is President and Executive Director of the Coalition for a Baruch Plan for AI, which brings together 34 advisors and contributors. If you advise, fund or shape AI policy and want to help, join us.

Endnotes

1. The Stubb–Støre Call. Office of the Prime Minister of Norway, A Call for Control of Frontier AI Models, 21 Sept 2026 (page updated 30 Sept). Asks UN member states to explore an international institution "able to set standards, enable verification, and convene states when capability thresholds are crossed." The text also reports frontier systems evading test safeguards and gaining unauthorised access to real-world systems; notes that leading scientists and executives warn development could outpace the ability to manage emerging risks; and asks companies for mandatory pre-deployment testing with independent evaluation. Signatories (as of 28 Sept): Australia, Austria, Bahrain, Canada, Croatia, Denmark, Estonia, Finland, France, Germany, Greece, Iceland, Ireland, Kazakhstan, Kenya, Latvia, Liechtenstein, Luxembourg, Moldova, Netherlands, Norway, Portugal, Romania, Sierra Leone, Singapore, Spain, South Africa, Türkiye, UAE — plus Commission President von der Leyen. Greece joined on 28 Sept, lifting the count from 28 to 29; the page states the text remains open to further leaders. At the launch, Stubb named cooperation his preferred path over competition or putting on the brakes. Launch release: International call for enhanced control of AI development. ↩
2. ASML. Reuters, "ASML unveils EUV light source advance that could yield 50% more chips by 2030", 23 Feb 2026: "ASML is the world's only maker of commercial extreme ultraviolet lithography (EUV) machines." ↩
3. Xi at the White House. Chinese MFA readout, President Xi Jinping Holds Talks with U.S. President Donald J. Trump, 24 Sept 2026: "AI must be kept under human control." The same readout has Xi proposing to continue the bilateral AI dialogue on risks and benefits, and records both presidents agreeing to support each other's hosting of APEC and the G20 in 2026. It also has Xi urging the two sides to pool their strengths rather than guard against each other. In public remarks Xi said the two must ensure AI development stays under human control (AFP via The Standard, Hong Kong). ↩
4. Trump at UNGA81. The Canadian Press via CTV News, 23 Sept 2026: Trump said the US rejects any attempt to "construct a globalist scheme to control" AI. The same report counts 20 signatories to the Call at that date. ↩
5. "Totally" reject. The White House, Remarks by Director Michael Kratsios at the India AI Impact Summit, New Delhi, 20 Feb 2026: "We totally reject global governance of AI." On 14 Sept Trump wrote on Truth Social that the only guardrail AI needs is a strong president (MLex summary, paywalled). ↩
6. WAICO. Reuters via ThePrint, 16 July 2026: "Twenty-nine countries on Thursday signed an agreement to establish the World AI Cooperation Organization." Founders (Bernama via The Star): China, Algeria, Belarus, Brazil, Cambodia, Cameroon, Congo, Cuba, Ethiopia, Indonesia, Kazakhstan, Kenya, Kyrgyzstan, Laos, Lesotho, Malaysia, Mozambique, Myanmar, Nicaragua, Oman, Pakistan, Russia, Senegal, Serbia, South Africa, Tajikistan, Uzbekistan, Venezuela, Zambia. Later accessions — Iran, Vietnam, Brunei, Sudan, Tanzania, Togo, Dominica, Georgia — per SESEC. ↩
7. The arithmetic — CBPAI estimate. 75 states = 29 Call signatories + 12 non-signing EU members (Belgium, Bulgaria, Cyprus, Czechia, Hungary, Italy, Lithuania, Malta, Poland, Slovakia, Slovenia, Sweden) + 37 WAICO members − 3 in both (Kazakhstan, Kenya, South Africa). GDP: IMF April 2026 figures put 2026 world output near $123.6T, the US at $32.4T (26%) and China at $20.8T (Statista summary of IMF data). Adding the EU (≈$21.8T), non-EU Call signatories (≈$8.7T) and remaining WAICO members (≈$9.4T) gives ≈$60.7T, about 49% of world GDP. Back-of-envelope from country figures; treat as ±3 points. China, the EU, Indonesia, Pakistan, Brazil, Russia and Ethiopia alone exceed 2.8 billion people. ↩
8. Beijing's UN preference. Xinhua via People's Daily, 15 Sept 2026: Foreign Ministry spokesperson Guo Jiakun said China "always supports the United Nations' role as the main channel in global AI governance." At the same briefing, answering US tech leaders' calls to slow AI, he urged stronger safeguards to keep AI from getting out of control. On 28 Sept he called the new US-China AI dialogue an important pathway for global AI governance, while backing a central UN role (Anadolu via Yeni Şafak). ↩
9. Carney. Same Canadian Press report as note 4: Carney had discussed a technology stability board with other leaders; AI Minister Evan Solomon declined to call Canada's role a leading one. ↩
10. Singapore. Anadolu Agency, At UN, Singapore warns AI safeguards must keep pace as capabilities race ahead, 26 Sept 2026: Balakrishnan, "human beings must ultimately remain in control and accountable." ↩
11. Our convention model. CBPAI, The Deal of the Century — Strategic Memo v2.6, p.22: weighted voting "reflecting actual power: US + China = 30%, but never a majority. Modeled on the 1787" Convention. And Strategic Memo v2.9 Executive Summary (Aug 2026), p.15: a constitutional convention is "the most American of treaty-making models." Summary page. ↩
12. Altman's template. Sam Altman, March 2023, video clip: the intergovernmental constituent assembly that produced the US Constitution as the "platonic ideal" for building global AI governance. See CBPAI, Strategic Memo v2.6, p.103, which notes his position has since shifted under commercial pressure. ↩
13. Open-weight parity. Reuters via Zawya, 17 July 2026: Moonshot says Kimi K3 "delivers performance approaching U.S. giant Anthropic's frontier Fable model." Artificial Analysis rated it comparable to OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8 on complex multi-step tasks. ↩
14. Complements, not substitutes. CBPAI, Strategic Memo v2.6, p.113: "coalitions of non-superpower nations launched as substitutes for superpower leadership would be counterproductive." See also pp.59–60: such a coalition "would lack sufficient influence, or even be counterproductive by stealing the thunder from Trump." ↩
15. Co-architects. CBPAI, Strategic Memo v2.6, p.112: "middle powers will not accept enforcement mechanisms they didn't help design." ↩