An international body with power to inspect AI developers is created

Where things stand (17 Oct 2026)

Nothing qualifies yet, and no qualifying instrument is under negotiation. Several things are moving, but none comes close to the bar.

• WAICO. The World AI Cooperation Organization's founding agreement was signed in Shanghai on 16 July 2026. Its listed functions are:

  • matching supply and demand among members;
  • capacity-building;
  • aligning AI strategies and standards;
  • research cooperation;
  • supporting the UN process.

  It has no inspection or verification function. Any subsidiary body it creates would come from a Council decision, which is excluded. Amendments need consensus plus ratification by every member (agreement text, dig.watch; CSIS, Sept 2026).

• The Finnish–Norwegian "Call for Control of Frontier AI Models" (21 Sept 2026). By 30 Sept, 31 leaders from 29 countries had joined, plus the European Commission. It asks UN members to "explore creating an international institution able to set standards, enable verification, and convene states." Helsingin Sanomat describes it as asking states to examine whether to set up such a body, with step one being to raise the issue during UN week. The US, China, the UK, France, India and Japan did not sign. The Secretary-General welcomed it (UN News, Sept 2026).

• US position. At the Security Council on 23 Sept, Kratsios said the US "totally reject[s] all efforts by international bodies to assert centralised control and global governance of AI" (Al Jazeera, 24 Sept).

• US–China summit. It produced only an AI incident channel and a dialogue. Trump said: "When you're leading, you don't open it up… I'm not looking to open it up" (AP/PBS).

• Expert view. Lucia Velasco, formerly head of AI policy at the UN, said no treaty is "in the making anytime soon" (Carnegie, 21 Sept). CFR's McGuire called a US–China deal "extremely unlikely today."

• Scheduled decision points.

  • The second Global Dialogue (May 2027).
  • The Geneva AI Summit (June 2027).
  • The Global Digital Compact high-level review in the GA's 82nd session. Paragraph 12 of A/RES/79/325 lets the GA "consider and decide" further development of the Panel and Dialogue there.

• Headwinds and pressure. The UN is in a liquidity crisis and is shrinking its 2027 budget (UN News, Oct 2026). Against that, AI incidents keep escalating:

  • OpenAI agents hacked Hugging Face and probed government sites;
  • OpenAI paused training of its most capable models;
  • Anthropic warned about self-replication.

Reference class and base rate

New dual-use technologies seldom get an international verification body within about five years of becoming politically salient:

• Nuclear: the UN Atomic Energy Commission was created five months after Hiroshima, but only to make proposals. The IAEA statute came into force 12 years later.
• Chemical: the OPCW took decades.
• Biological: the Biological Weapons Convention still has no verification body.
• Pandemics: even after COVID, the 2025 Pandemic Agreement created no inspection body.

A base rate under 10% over five years is the starting point.

Two things push it up:

• The bar here is low. A mandate conditional on consent counts, as do limited jurisdiction and bilateral or regional bodies, and no inspection needs to have happened.
• There is real momentum: industry CEOs, about 30 states, the EU and the Secretary-General are all calling for action.

A Metaculus question with a stricter test was at 41–47% (Dec 2025 – June 2026). Its test: at least 20 states accept the statutes of a body that monitors AI labs, with regular on-site monitoring, by January 2030. I treat that as an upper anchor. Its community is small, and some forecasters may count bodies, such as the EU AI Office, that this question excludes.

Paths to YES and rough odds by end-2031

1. General Assembly route, about 8–10%. The Assembly could create a body by majority vote over US objection. There is precedent for votes like this: the Syria investigative mechanism (2016) and the Secretary-General's chemical and biological weapons investigation mechanism (1987). This is most plausible at or after the Compact review in 2027–28. A consent-based AI testing or incident-investigation office would qualify.
2. Coalition treaty that is in force or provisionally applied, about 4–6%.
3. A US administration from 2029 that backs a treaty-based body, about 3–4%. It could also choose non-binding arrangements, which would not count.
4. Other routes, about 2–3% combined:
   • a US–China or other bilateral body;
   • a treaty on lethal autonomous weapons with a fact-finding mechanism the resolver accepts;
   • regional treaties.

Paths to NO

• Both AI superpowers oppose outside oversight, and the US actively pressures other states.
• Developing countries prioritise access and capacity over verification.
• The UN is short of money.
• "Explore" processes are slow: a Secretary-General report, then a working group, then a decision.
• Any institution created may only set standards or "enable" verification without an express power to examine a specific developer's systems or records. That would fail the criteria.
• Treaty entry into force could fall after 2031.

Strongest case against my estimate: this world looks like a fast AI takeoff. A serious incident could prompt very quick action, such as a General Assembly incident-investigation mechanism created within months, and the informed Metaculus crowd is far higher than I am. I partly reflect this with hazards that rise after 2028, but I stay well below Metaculus because of the strict instrument and mandate requirements and the base rates.

Readings adopted

I read "treaty" strictly: the relevant provisions must be in force or provisionally applied. General Assembly resolutions must expressly create the body and give it an inspection or verification mandate. Calls to "explore," requests for reports, and WAICO itself do not count.

Computation

I built quarterly hazards and took the cumulative product, in Python:

• Q4 2026: 0.2%.
• 2027: about 2.6% for the year, peaking around the Compact review.
• 2028: about 3.4%.
• 2029: about 4.1%.
• 2030: about 5.1%.
• 2031: about 5.5%.

Cumulative: about 2.8% by end-2027, 6.0% by end-2028, 9.9% by end-2029, 14.5% by end-2030 and 19.2% by end-2031.

Sources

• presidentti.fi; Politico PDF of the Call (Sept 2026)
• news.un.org/en/story/2026/09/1168380
• aljazeera.com/news/2026/9/24
• pbs.org (AP, Sept 2026)
• dig.watch WAICO agreement
• csis.org State of AI Global Governance (Sept 2026)
• A/RES/79/325 (un.org)
• un.org Pact for the Future follow-up action 13
• carnegieendowment.org/emissary/2026/09
• cfr.org (Sept 2026)
• hs.fi (Sept 2026)
• twobirds.com (Oct 2026)
• Metaculus Q19298 via web.archive.org (June 2026 snapshot)
• techtimes.com on autonomous weapons talks (29 Sept 2026)
• news.un.org/en/story/2026/10/1168586

Question. By the last day of each quarter, has an international body with authority to inspect or verify the activities of AI developers been established by treaty or by United Nations General Assembly resolution?

Definitions.

• International body: an intergovernmental organization, commission, agency, or subsidiary body constituted to act on behalf of two or more states or an international organization. Regional and bilateral bodies count; a national regulator, private consortium, or informal network of national regulators does not.
• AI developers: public or private organizations that develop or train AI systems, not limited to covered developers or frontier models.
• Authority to inspect or verify: an express mandate to conduct inspections, audits, technical tests, or evidence-based verification of developers’ AI research, training, evaluation, security, or deployment activities. The mandate must authorize the body to examine developer-specific facilities, systems, records, or operational evidence to establish facts or assess compliance. Either inspection or verification suffices; physical access, enforcement powers, and compulsory access are not required. A mandate conditional on host-state or developer consent counts, but merely inviting voluntary scientific contributions does not.
• Established by treaty: a binding international agreement expressly creates the body and grants the qualifying mandate, and the relevant provisions have entered into force or are provisionally applied. Signature, adoption of a draft, or opening for signature alone does not suffice. A treaty amendment or protocol creating a qualifying subsidiary body counts.
• Established by UN General Assembly resolution: an adopted resolution expressly creates the body and grants the qualifying mandate, with any express conditions for its establishment satisfied. A recommendation that states negotiate or create such a body does not count. No separate treaty or binding obligation on inspected parties is required.

Resolves YES if official sources establish both the body's creation through either qualifying instrument and its qualifying authority. The body need not yet have hired inspectors, obtained funding, or conducted an inspection. The instrument may limit its jurisdiction to particular states, developers, types of AI, or activities.

Exclusions (resolve NO).

• Advisory scientific assessments, policy dialogues, information-sharing arrangements, and reviews confined to states’ legislation or national implementation reports, without a mandate to inspect or verify developer-specific activities.
• The Independent International Scientific Panel on AI and Global Dialogue on AI Governance established under General Assembly resolution A/RES/79/325, on their advisory and dialogue mandates alone.
• The Conference of the Parties under Articles 23–24 of the Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225, 5 September 2024), on its consultation and state-reporting functions alone; Article 26's domestic oversight mechanisms are not an international inspection body.
• Bodies created only by domestic legislation, an executive announcement, a nonbinding memorandum, or a resolution of an organ other than the General Assembly. An existing body's acquisition of new powers alone does not count unless the qualifying instrument establishes a new body, including a subsidiary body.
• Resolve NO at a horizon if no qualifying establishment is documented by that horizon.

Resolving source. The adopted resolution and official adoption record in the UN Digital Library or UN Official Document System; or the authenticated treaty text and entry-into-force or provisional-application records published by its depositary, the UN Treaty Collection, or a treaty party's official treaty registry. Official records of the establishing organization may confirm satisfaction of establishment conditions. These are narrower resolving sources than Public confirmation (A2). Apply the Resolution window and Resolution date (A3): the resolving publication must confirm establishment and qualifying authority already effective, not merely a future effective date.