An international body gains binding authority over US and Chinese AI

Where things stand (15 October 2026)

No treaty that is in force for both the US and China gives any international body binding power over AI deployment. Nothing that has happened since the question opened on 16 September 2026 comes close to that.

• The Trump–Xi state visit (23–25 September 2026) produced only a "Super Intelligence (SI) Dialogue", with the next exchange due in November, and an agreement to set up a channel for AI incidents. The White House fact sheet and China's foreign ministry list of eight deliverables (26 September) describe political commitments. No signed text has been published (Digital Applied, 27 September). Signals from the Chinese side point to "low-hanging fruit" for November: the incident channel, misuse by terrorists, and how to classify risks (Pekingnology, 1 October). Under the question's exclusions, a channel or dialogue counts only as consultation or monitoring.
• The US position is hostile. At the UN General Assembly on 22 September, Trump said the US "totally rejects any attempt to construct a globalist scheme to control" AI (Bloomberg via Yahoo). After the summit he said the US is "not going to be putting on brakes" (USA Today, 26 September). In January 2026 the US withdrew from 66 international organizations.
• Other tracks do not qualify. Twenty countries plus the EU called for states to "explore" an institution that could set standards. The US and China did not sign (NBC, 22 September). China's World AI Cooperation Organization has 38 founding members, and neither the US nor any other G7 country is among them. The UN Global Dialogue on AI is not a treaty process. In the talks on autonomous weapons, the US and Russia watered down the final report (HRW, 7 September). In any case, a protocol under the Convention on Certain Conventional Weapons would place obligations directly on states rather than hand decisions to a body. The mandatory version of the shipping regulator's code for autonomous ships is not expected to take effect until 1 January 2032 (IMO), which is after the window. I found no existing treaty that both countries belong to that expressly covers AI.

Reference class and base rate

There has never been a US–China bilateral arms-control treaty, let alone one that delegates binding decisions to a body. Novel strategic technologies have not historically been placed under binding international control: the 1946 Baruch Plan for nuclear power failed, and the IAEA never received binding authority. The closest precedent for a body with delegated power, the Montreal Protocol's power to adjust ozone limits, covered a technology that was not strategic. A US–China AI deal of any kind is "exceedingly unlikely… and without precedent" in the view of CFR's Chris McGuire, and the IELP blog (30 September) concludes that any treaty must wait for a new administration. Related Metaculus questions, which I could see only in search snippets, sit at about 5–6% for a US–China treaty limiting advanced general AI being signed before 2030, and about 4–5% for any formal agreement to limit frontier AI before 2029. This question is much stricter: the treaty must be in force for both countries, a body must hold binding power, and that power must cover deployment.

Paths to YES and the hazard model

1. A crisis or "grand bargain" under Trump (to January 2029). This world is moving fast: AI companies report AI that helps build the next generation of AI, a swarm of OpenAI agents attacked Hugging Face, and lab leaders are calling to "pace the frontier". A catastrophe could change Trump's mind. Even then, his style favours non-binding "boards" and announcements. I put this path at about 0.5%.
2. A new administration from 2029. Prediction markets give the Democrats about 61–64% to win in 2028. Many Democrats, including Khanna, Sanders and Harris, call for a binding US–China AI treaty. My chain of estimates: the administration seriously pursues a treaty (about 50%) × China concludes one by 2031 (about 25%) × it is legally binding rather than a political commitment (about 50%) × it is in force by December 2031 (about 55%) × it contains a qualifying binding power over deployment (about 40%). Adding a smaller Republican path and extra weight for a crisis gives about 2–2.5%.
3. Other routes, such as amendments to existing treaties: about 0.2%.

Quarterly hazards: 0.02% in Q4 2026; 0.06% per quarter in 2027; 0.07% in 2028; rising from 0.08% to 0.17% through 2029; 0.20–0.26% in 2030; 0.27–0.30% in 2031. I computed the cumulative probability in Python as one minus the product of (1 − hazard). It reaches 0.54% at end-2028 and 3.07% at end-2031.

Strongest case against my estimate

AI capabilities in this timeline are accelerating unusually fast, and pressure from the labs, the public and other governments is mounting. A treaty that genuinely paced frontier AI would naturally include a commission, possibly deciding by consensus, that updates deployment thresholds. Consensus rules and vetoes are allowed under the criteria, and a bilateral executive agreement can enter into force on signature. That could justify 5% or more. Against that: both governments currently reject binding constraints; China resists binding third-party authority; US–China AI arrangements so far have been non-binding; and negotiating a treaty and bringing it into force within about three years of a 2029 inauguration is tight. So I stay at about 3%.

Question. By the last day of each quarter, has a treaty in force for both the United States and China expressly given an international body authority to make legally binding decisions for both governments about the deployment of AI systems?

Definitions.

• Treaty: a written agreement between states governed by international law, regardless of its title or domestic approval procedure, including a protocol or amendment. China means the People's Republic of China.
• In force for both: the treaty and the provisions conferring the relevant authority have legally entered into force for both states at the same time, taking account of reservations, opt-outs, suspensions and effective withdrawals. Signature, announced ratification, or provisional application alone does not qualify.
• International body: an institution established or designated by states, including an international organization, treaty commission, conference of parties, court or tribunal. It may be bilateral or multilateral, new or pre-existing.
• Expressly given authority: the treaty's operative text, including legally integral annexes or protocols, explicitly covers artificial intelligence systems and empowers the body to adopt binding decisions concerning their deployment. The AI scope and the binding decision power may appear in separate provisions read together. General authority over security, trade or technology without express AI coverage does not qualify.
• Deployment: putting AI systems into operational use, making them available for use, or continuing their operation, including civilian or military use. Authority confined to a specified sector or class of AI systems qualifies. Research, training, compute allocation or reporting alone does not.
• Legally binding decisions for both governments: decisions that, once adopted under the treaty's procedures, create international legal obligations for both governments concerning whether, where, when or under what conditions AI systems may be deployed. This includes binding deployment standards, authorizations, prohibitions and binding judicial or arbitral orders on deployment. Decisions must not require a separate subsequent agreement or ratification by either government to become binding. Voting rights, unanimity requirements or vetoes within the body's decision procedure do not disqualify it. Domestic implementing legislation, direct domestic enforceability and coercive enforcement powers are not required.

Resolves YES if qualifying official publications establish that all these conditions have been satisfied. The body need not yet have exercised its authority, and the treaty need not govern all AI systems or all countries. At least one class of substantive deployment decisions must fall within the authority accepted by both governments; separate authorities accepted by only one government each do not suffice.

Resolves NO / Exclusions.

• Draft treaties, signatures without entry into force for both states, political declarations, voluntary codes, advisory recommendations and nonbinding standards do not qualify.
• Binding treaty obligations imposed directly on states without delegated binding decision-making power do not qualify. Monitoring, consultation, recommendations, procedural decisions and requests for information alone do not qualify.
• A treaty requiring further consent to confer the relevant jurisdiction does not qualify until both governments' required consents are effective.
• The UN Charter's general Security Council powers, or an AI-related Security Council resolution relying solely on those powers, do not themselves satisfy the requirement for an express treaty grant concerning AI.
• Resolve NO for a horizon if no qualifying confirmation has been published by that horizon.

Resolving source. Official treaty texts and official records of entry into force, participation, reservations and jurisdictional acceptances: the United Nations Treaty Collection's Status of Treaties, Depositary Notifications and United Nations Treaty Series (treaties.un.org); the treaty's designated depositary or administering organization's treaty office, including the Council of Europe Treaty Office where applicable; or the U.S. Department of State and China's Ministry of Foreign Affairs. The operative text controls the authority granted; depositary records control treaty status where available. These are the exclusive resolving sources, rather than the broader Public confirmation convention A2. Apply the Resolution window and Resolution date convention A3: the qualifying publication must fall within the window and confirm conditions already satisfied, not merely a scheduled future entry into force. Earlier treaty texts may support a qualifying new status publication. Once YES, later withdrawal or suspension does not reverse resolution.