20 Nations Seek Global AI Body; US and China Excluded

Twenty countries and the European Commission signed a joint statement on Monday afternoon calling for mandatory safety testing prior to the launch of advanced artificial intelligence systems and paving the way for an international institution with verification authority. The United States, the United Kingdom, and China did not participate.

"AI must remain under human direction, supervision, and control," states the text, which adds that the technology "must be developed and used in accordance with international law." Furthermore, the statement urges UN member states to explore creating an international institution "capable of defining standards, enabling verification, and summoning states when capacity thresholds are exceeded."

Who Signed and What the Statement Asks For

The list includes Canadian Prime Minister Mark Carney, German Chancellor Friedrich Merz, Spanish Prime Minister Pedro Sánchez, Australian Prime Minister Anthony Albanese, and European Commission President Ursula von der Leyen, along with Norway, Finland, Ireland, the Netherlands, Singapore, South Africa, and Kenya. The statement was released hours before the opening of the high-level week of the UN General Assembly, from September 21 to 27.

There are three main requests. The first demands that companies conduct safety testing before public deployment and allow independent evaluators access to their internal development practices. The second asks that governments and regional blocs establish common evaluation standards, with scientific capacity to assess models. The third is the international institution, presented as something to be explored, rather than created immediately.

The distinction between the three is not rhetorical. Pre-deployment testing falls on the company that trains the model. Independent verification requires external intervention in the training process. Summoning states when a capacity threshold is crossed assumes an agreed-upon and measured threshold, which none of the signatories has published.

The Absence Carries More Weight Than Inclusion

The three countries that opted out host nearly all relevant frontier labs. OpenAI and Anthropic are based in the United States, Google DeepMind in the United Kingdom, while China focuses on the open models that currently exert price pressure in the global market. A verification regime without these three signatories impacts those who purchase the technology, not those who produce it.

The American stance did not originate on Monday. On September 17, reports indicated that Mark Zuckerberg, Elon Musk, and Jensen Huang spoke separately with President Donald Trump against the creation of an AI regulator funded by the industry, similar to Finra, a proposal associated with Demis Hassabis. The argument from the three was reportedly that such a structure would cement the positions of OpenAI, Anthropic, and Google behind a regulatory moat. Trump did not advance the idea.

The industry is also divided over the pace of developments. The paper "We Must Pace the Frontier," published by Dario Amodei on September 12, proposed granting third-party evaluators permanent and equivalent access to employees to ensure adherence to safety commitments during training. Sam Altman and Hassabis supported the idea of a coordinated slowdown. Zuckerberg and Huang rejected it, arguing that product competition and civil liability already penalize those who provide unsafe models.

What Changes for Buyers

For a German company, little changes in the short term, as the binding obligation already exists. The transparency rules of Article 50 of the European AI Act have been in effect since August 2, 2026, with fines up to 15 million euros or 3% of global revenue, whichever is greater. The statement from Monday adds a political layer above a regulation that already has enforcement mechanisms.

In Singapore, a signatory, the calculus is different. The city-state houses the Asian AI operations of global banks, and a common evaluation standard prior to deployment would affect models already in production for clients across the region.

Canada and Australia signed without hosting frontier labs, which shifts the equation. The Indian delivery centers of TCS, Infosys, and Accenture operate AI pipelines for European, Canadian, and Australian clients who have now endorsed the statement. The same applies to the Brazilian operations in shared services and BPO of multinationals: the pre-deployment testing requirement does not stem from local law, but rather from a client's contractual clause, shifting the cost of execution to the supplier, not to those who signed the statement.

The institutional dispute resolves faster than the document suggests. France holds the presidency of the UN Security Council in September and has convened an open meeting on AI and international security, chaired by Chancellor Jean-Noël Barrot, where Altman is expected to speak to the 15 members. The three countries that did not sign the statement all hold veto power in that discussion.