The "Paris Call for AI Safety and Security," announced by French President Emmanuel Macron on November 9, 2026, has been signed by 16 countries, 10 companies including OpenAI and Anthropic, and seven civil society organizations. Those figures come from the signatory list published by the French presidency (Élysée). In the declaration, AI developers pledge to implement fail-safe mechanisms that allow them to restrict or halt operations when necessary, including during model training and internal use. However, the pledge covers only operations under the developer's own control, and the declaration does not give governments or third parties any authority to stop a model. Although a broad group of countries and companies has aligned on the direction of safety measures, how to make them effective is left to future institutional development.
What the 16 countries and 10 companies signed
The signatory list published by the Élysée includes 16 national governments, among them Austria, Bahrain, Canada, Finland, France, Ireland, Kenya, the Netherlands, Norway, and Portugal. Companies taking part include Anthropic, OpenAI, Microsoft, Mistral AI, Hugging Face, Cohere, and Sakana AI, for a total of 10. The governments of Germany, the United States, the United Kingdom, and Japan do not appear on the list. The list states that it continues to accept new endorsements and will be updated, so these figures reflect the situation as of November 11, 2026.
The number of signatory countries differs from some media reports. Both Le Monde and The Decoder reported that eight countries took part, but the Élysée's published list names 16. It cannot be confirmed whether more countries signed after the announcement or whether the counts were based on different criteria. This article uses 16 countries, based on the official list as of November 11.
The declaration rests on three pillars. The first is a shared framework for evaluating and managing the risks of large frontier AI models, with the aim of publishing non-sensitive evaluation results in a standard format and collecting them in a shared repository. The second reaffirms the principle that developers, policymakers, and others bear responsibility according to their roles. The third is building systems to classify, share information on, and provide early warning of serious cyber incidents and incidents involving loss of control over AI. The preamble positions the declaration as building on earlier international efforts, including the Hiroshima AI Process, the Frontier AI Safety Commitments from Seoul, and the G7 Evian summit in June 2026.
The declaration contains no specific implementation deadlines, penalties for violations, or legally binding provisions. In relation to Japan, Sakana AI, a Japanese AI developer, is a signatory, while the Japanese government is not a participant. The declaration also explicitly names the Japan-led Hiroshima AI Process as one of its foundations, carrying forward existing international discussions on safety measures.
Will AI shutdown mechanisms be required?
The key to judging the declaration's effectiveness lies in five provisions aimed at developers. The first clearly commits them to disclosing information on model capabilities, limitations, and known risks. The second addresses independent third-party evaluation, but support for assessment by qualified external evaluators applies only where it is "appropriate and proportionate to risk," and it must also be consistent with existing practices. The third commits developers to implementing fail-safe mechanisms that enable restriction, pause, and full shutdown of model operations, including during training and internal use. However, the scope is explicitly limited to "model operations under the developer's control."
The remaining two provisions focus on affirming principles rather than specific obligations. The fourth acknowledges the risk that safety measures may be removed or circumvented, while stating a policy of preserving the benefits of developing open-weight models. The fifth recognizes the importance of internal governance and commits to establishing governance structures with clear lines of accountability. In other words, of the five items, the only one pledged without limits on conditions or scope is disclosure of information on model capabilities, limitations, and known risks. The shutdown mechanism is constrained by what the developer controls, and independent evaluation comes with conditions for implementation.
Understanding the scope of the shutdown mechanism requires considering how models are delivered. When a company runs a model on its own servers and offers it to users via an API, it can pause or fully stop operations. With the open-weight approach, in which model weights are published, third parties can run copies in their own environments, so operations after release are not necessarily under the original developer's control. The declaration acknowledges the risk that safety measures may be stripped out or bypassed, but says the benefits of open-weight development will be maintained. A pledge to implement fail-safe mechanisms therefore does not mean copies of already released models can be stopped.
The provisions for policymakers likewise set out no specific authority to trigger a shutdown from outside. On responsibility when large-scale harm occurs, the declaration recognizes the need to clarify it within legal frameworks. By contrast, it explicitly commits to establishing confidential channels through which whistleblowers can report concerns without fear of retaliation. What the current declaration shows, then, is a policy of asking developers to implement shutdown mechanisms, not a system in which governments or independent bodies can directly invoke them.
Could it develop into regulatory oversight?
That said, shutdown mechanisms will not necessarily remain a matter of voluntary management by developers. The declaration's second pillar includes a provision in which all actors commit to complying with rigorous safety standards, set or accredited by competent national authorities, under authority supervision and independent third-party evaluation. If regulators in signatory countries later build shutdown mechanisms into safety standards, they could evolve from voluntary corporate commitments into requirements subject to external oversight. The declaration leaves room for institutional development of that kind.
At present, however, no specific safety standards making shutdown mechanisms mandatory have been presented. The technical standards meant to support the common evaluation framework are described only in terms of future development, and cooperation mechanisms for incident response are at the stage of examining their usefulness. It is also unclear what powers national regulators would have or what they could require of developers who fail to meet the standards. The declaration therefore cannot be said to have established a system of external stop orders or forced shutdowns for AI models.
Two developments will matter in judging future effectiveness. The first is whether regulators in signatory countries build shutdown mechanisms into accredited safety standards and put in place a system to verify implementation and demand corrections when deficiencies are found. The second is whether evaluation results, including tests of shutdown mechanisms, are registered in a standard format in the public repository the declaration envisions, so that third parties can verify them. The former is a measure of regulatory effectiveness, the latter of transparency about implementation. Even if evaluation results are published, however, that alone does not give regulators the authority to stop a model.
Because the declaration sets no implementation deadlines, it is unclear when such systems will be in place. Attention will turn to how signatory countries put safety standards and oversight systems into concrete form, and to whether Japan, Germany, and others join the updated signatory list. That 16 countries and 10 companies have endorsed common safety principles is a step toward international cooperation, but developers having shutdown mechanisms is a separate matter from outside bodies being able to verify and supervise their implementation and, if necessary, force a shutdown. Whether the declaration develops into effective international safety measures depends on national institution-building and the creation of a verifiable evaluation system.