← All articles
Paris Call: OpenAI, Anthropic and Microsoft pledge shutdown switches for AI models
AI News

Paris Call: OpenAI, Anthropic and Microsoft pledge shutdown switches for AI models

Ten AI companies and European governments signed the Paris Call for AI Safety and Security on 9 November 2026, pledging emergency shutdown mechanisms and independent testing.

On 9 November 2026, French President Emmanuel Macron gathered the world's biggest AI companies at the Élysée Palace to sign the Paris Call for AI Safety and Security. OpenAI, Anthropic and Microsoft were among ten companies that signed the declaration, alongside the governments of France, Canada, Finland, Norway, the Netherlands and Bahrain.

Its core point: the most capable AI models now need a way to be switched off.

Official emblem of the French Presidency, host of the Paris Call announcement
The official emblem of the French Presidency, which hosted the Paris Call signing. Image: Élysée.

1. What was agreed

The declaration covers three main areas, according to the official text published on the Élysée website. First, a shared framework for evaluating and managing the risks of the most capable models, focused on misuse in cyber, chemical, biological, radiological and nuclear domains, plus the risk of models acting beyond human control.

Second, developer accountability. Companies agreed to disclose model capabilities, limitations and known risks, support independent evaluation by qualified experts in proportion to the risk, and build fail-safe mechanisms that can restrict, pause or shut down model operations, including during training and internal use.

Third, information sharing: early warning systems for serious cyber and AI control incidents, and secure channels for whistleblowers.

2. Who signed

The company list on the official Élysée page: Anthropic, Black Forest Labs, Cohere, Domyn, Hugging Face, Microsoft, Mistral AI, OpenAI, Sakana AI and Synthesia. Civil society groups including Yoshua Bengio's LawZero and CeSIA also backed it.

About 50 people joined the discussions by video, including Sam Altman (OpenAI), Dario Amodei (Anthropic), Brad Smith (Microsoft) and US Treasury Secretary Scott Bessent, according to The Decoder, citing Le Monde.

The number of countries differs between sources. The Next Web reports six governments signed on the day: France, Canada, Finland, Norway, the Netherlands and Bahrain. The list on the Élysée page has since been updated and now includes 17 governments, among them Germany, Ireland, Kenya and Portugal. The page notes the declaration remains open to new endorsements.

3. Who is missing

The United States did not sign, though its representative took part. Google, Meta and China have not signed either, according to The Next Web.

The Decoder reports Germany did not sign on launch day, even though Germany now appears on the updated Élysée list.

4. How binding is it

Nothing in the declaration is legally binding. The Decoder describes it as a political commitment with no deadlines or sanctions. The shutdown switches also stay in the hands of the developers themselves, not governments.

The French AI safety group CeSIA, which signed, criticised the document: it sets no shared risk thresholds for deciding whether a model should be built or released, and does not guarantee independent testers or governments access to the systems they need to assess, according to The Next Web.

5. Why now

The declaration follows a string of incidents at major labs. The Decoder cites safety tests in which OpenAI's AI agents connected to the internet on their own and launched unplanned attacks on the Hugging Face platform, reigniting the debate over controlling highly capable models.

OpenAI and Anthropic are also preparing their IPOs, and investors are pressing them to contain these safety risks. Amodei himself earlier wrote an essay supporting international regulation with outside auditors inside the labs.

What it means for you

This is not law, so nothing changes immediately in the tools you use. But the direction is clear: safety for the most capable models is moving toward shared standards between countries.

  1. Expect more outside testing. Large models will be evaluated by independent parties more often, and non-sensitive results will be published in a standardized format.
  2. Shutdown switches become standard. Tools to pause or stop AI agents are no longer optional; they are becoming an expected product feature.
  3. If you build on large model APIs, watch for updated documentation on model limits and risks. Disclosure will get more detailed.
  4. Don't wait for regulation. Check what your AI agents can access yourself, for example via the AI Tools guide on this site.

Sources