An international agreement names an AI as party or guarantor

Where things stand (as of 15 Oct 2026)

No international agreement has ever named an AI system, or a body made up of AI systems, as a party or a guarantor. I found nothing in negotiation that would do so. AI is now high on the diplomatic agenda: since summer 2026 there have been agent-containment failures, including the OpenAI–Hugging Face swarm incident. Amodei's 12 Sept essay "We Must Pace the Frontier" was echoed by other lab leaders. Even so, every concrete instrument and proposal treats AI as the thing being regulated, or at most as a tool:

• US–China summit (24–26 Sept 2026). It produced a "Super Intelligence Dialogue" due to meet by November and a channel for reporting AI incidents. There was no treaty. Trump said the US is "not going to be putting on brakes" and told the UN General Assembly he rejects any "globalist scheme to control" AI. (White House fact sheet, 25 Sept: whitehouse.gov/fact-sheets/2026/09/fact-sheet-president-donald-j-trump-advances-a-fair-and-reciprocal-relationship-with-china-while-hosting-historic-state-visit/; AP, 26 Sept: apnews.com/article/china-us-agreement-xi-trump-visit-e8f858ed9094b99bc8d3d339f9899f31; Carnegie, Prescott, 30 Sept: carnegieendowment.org/emissary/2026/09/trump-xi-ai-safety-no-deal)
• The 30 Sept White House accord with AI CEOs. This is domestic and voluntary, and it relies on human auditors (npr.org/2026/09/30/nx-s1-5985699).
• Proposed verification schemes. Amodei's four levels of agreement rely on embedded human evaluators and a standards body (darioamodei.com/post/we-must-pace-the-frontier). The AI Futures Project's "Plan A" verification plan relies on hardware and human inspection (ai-2040.com/supplements/verification-plan). CFR's McGuire calls a US–China AI deal "exceedingly unlikely under current strategic conditions" (cfr.org/articles/how-to-reach-an-ai-arms-control-deal-with-china, 24 Sept). Under the resolution criteria, supplying verification or enforcement tools to human guarantors does not count.
• Novel legal moves at home, not abroad. Albania's AI "minister" Diella is mired in a lawsuit and a corruption probe of the agency that built it. Argentina has proposed "non-human corporations" run by AI. Neither involves an international agreement.

Reference class and base rate

The reference class is radical additions to who can be a treaty party or guarantor. Treaty parties have long been states, later international organisations; guarantors have been states, sometimes the UN. In the historical record the base rate for an AI in either role is zero. On nearby questions, Metaculus as reported on 25 Sept 2026 (verity.news/daily-newsletter?date=2026-09-25) gave:

• about 25% that the US and China are parties to an AI-regulation treaty by 2030;
• about 5% for a formal agreement limiting frontier AI before 2029.

Naming an AI as guarantor is a much narrower event than either.

Paths to YES

1. (Main path) AI that is better than humans arrives in 2027–29, which lab insiders now expect. A later US–China or multilateral deal then assigns enforcement to a jointly built AI, as the fictional AI 2027 scenario imagines. My rough estimate: about 30% chance of fast takeoff, times about 30% chance of a deal given that, times about 25% chance the published text names the AI as responsible for ensuring the agreement is observed. That comes to roughly 2–3%.
2. AI-enthusiast governments (Albania, the UAE, Argentina) stage a stunt, naming an AI as guarantor in a published memorandum of understanding: about 0.5–1%.
3. Other agreements with AI enforcement roles, or deals in which an AI system or AI collective is itself a party: about 1%.

Paths to NO

• Treaty lawyers are conservative and will usually name states or human-led commissions as responsible, with AI as an instrument.
• Misalignment incidents make governments reluctant to give AI formal authority.
• The current US administration is hostile to binding AI deals until January 2029.
• The resolution requires an official published text.
• Catastrophe or a secret deal would also block resolution.

Hazard model

I computed this in Python, taking the cumulative product of quarterly hazards:

• Q4 2026: 0.05% (2.5 months left, nothing in the pipeline);
• 2027: 0.06–0.12% per quarter;
• 2028: 0.14–0.23%;
• 2029: 0.26–0.31% (new US administration, possibly superhuman AI);
• 2030–31: flat at 0.32%.

That gives about 0.4% by the end of 2027, 1.1% by the end of 2028, 2.3% by the end of 2029, and 4.75% by the end of 2031. A rough breakdown by scenario gave about 3.75%, and the series sits between that and about 5.5%.

Strongest case against my number: AI 2027 has so far tracked this timeline closely, and in both of its endings an AI enforces the US–China treaty by about 2028. If the published text of such a deal names the AI as enforcer, the probability could be several times higher. I give this partial weight because real drafting practice and the strict wording of the criteria push the other way.

Question. By the last day of each quarter, has an international agreement been concluded that names an AI system, or a body made up of AI systems, as a party or as its guarantor?

Definitions.

• International agreement: a written agreement involving at least two national governments as participants, including a treaty, executive agreement, peace settlement, or expressly nonbinding intergovernmental agreement. It must set out agreed commitments, not merely shared aspirations or a proposal for future negotiations. Commercial contracts, private cross-border agreements, and agreements solely between subnational governments do not count.
• Concluded: the final text has been signed by authorized representatives of at least two participating national governments, or those governments have formally accepted it by an exchange of instruments or an equivalent recorded procedure. Signature subject to ratification counts; ratification, entry into force, UN registration, and implementation are not required. Adoption or initialling of a negotiating text alone, or opening a treaty for signature without the required signatures, does not count.
• AI system: an identified computational system that uses AI models to generate decisions or actions, from any developer. Identification by a proper name, unique designation, or an unambiguous description in the agreement or an incorporated annex counts. The designation may specify a system to be created under the agreement.
• Body made up of AI systems: an identified collective whose members exercising its decision-making authority are all AI systems. Human support staff, hosting, or external oversight do not disqualify it; a mixed human–AI decision-making body does not count.
• Party: the agreement expressly identifies the AI system or AI body itself as a party or contracting participant in its own right, with rights or commitments attributed to it rather than solely to its developer, operator, or a government it represents. A human may execute the instrument on the AI's behalf if the text makes this capacity explicit. General recognition of AI legal personality is not required.
• Guarantor: the agreement expressly designates the AI system or AI body itself as a guarantor, or uses an equivalent designation assigning it responsibility for ensuring the agreement's observance. Merely supplying monitoring, verification, analysis, arbitration, or enforcement tools to human guarantors does not count. A guarantor need not also be a party or sign the agreement.

Resolves YES if a qualifying official publication establishes both conclusion of the agreement and either qualifying AI role. The designation must appear in the final agreement, an incorporated annex, or a formally concluded amendment or protocol; an amendment adding the role to an older agreement counts. Subsequent withdrawal, termination, or failure to implement does not reverse YES.

Exclusions (resolve NO). Drafts, negotiation announcements, simulations, fictional scenarios such as AI 2027, unilateral declarations, and descriptions of an AI as a guarantor only in commentary do not count. Nor do agreements merely regulating AI, agreements drafted or negotiated using AI, or agreements naming a company or government that uses AI rather than the AI itself. Resolve NO at a horizon if no qualifying publication exists by that horizon.

Resolving source. Official agreement texts and signature records, or depositary notices that establish the required designation and conclusion, published by a participating government's foreign ministry or official gazette, the agreement's designated depositary, the United Nations Treaty Collection (treaties.un.org), or an intergovernmental organization's official treaty repository. A depositary notice may establish signature status while an official text establishes the AI role. This is a narrower source requirement than Public confirmation (A2); developer publications and media reports alone cannot resolve the question. Apply the Resolution window and Resolution date (A3): resolution is on the publication date of the qualifying official evidence, not the signature or entry-into-force date; where several documents are necessary, use the date the last necessary document is published.