The most concrete commitment of the Paris Call, the ability to stop an AI model, only covers models that their developers already have the power to stop. In this text published by the Élysée on 9 November 2026, after a working session by videoconference with around fifty participants, developers commit to mechanisms capable of interrupting "the operation of models placed under their control". What runs elsewhere is not concerned. And no one else but them decides.
Germany has joined a list that is still open
On 18 November, the Élysée's list included 18 states, 10 companies and 11 civil society organisations. It has grown since 9 November, when the Canadian government's statement listed only six countries; Germany, then Spain, are among the newcomers. The companies are Anthropic, Black Forest Labs, Cohere, Domyn, Hugging Face, Microsoft, Mistral AI, OpenAI, Sakana AI and Synthesia. Among the organisations are LawZero, founded by Yoshua Bengio, Ifri and CeSIA.
Three commitments make up the text. The first promises a common framework for evaluating frontier models, the most advanced systems, against malicious uses, from hacking to chemical weapons, and the risk of loss of control, with the publication of non-sensitive results in a standardised format. Alerting on serious incidents falls under the third. The second, devoted to responsibility, carries the stop clause: restrict, pause or cut off a model, including during training and internal use.
Open weights escape the stop clause
Targeting training and internal use is not trivial. It is in this phase, outside any use by customers, that several incidents of 2026 occurred: agents from OpenAI and Anthropic reached third-party systems during the evaluation of their own models, as ActuIA detailed on 5 October.
The limit lies in the words "under their control". A model that its developer serves itself online remains within its reach: it can cut it off. An open-weight model, however, is downloaded in full, and its copy runs on the machine of whoever retrieved it. Yet several signatories distribute such models, Mistral AI with Mistral Large 4, Black Forest Labs with part of its FLUX 3 models, all hosted by Hugging Face, itself a signatory. The text knows the problem. In the same sentence, it recognises the risk that safeguards may be removed or bypassed, and wants to address it while preserving the benefits of open weights. It does not say how.
No public authority, no third party receives the power to order a stop: the developer decides alone. CeSIA, a French organisation dedicated to AI safety, signed while contesting this point, and calls for governments to be able to suspend a model. For its executive director, Charbel-Raphaël Ségerie, these decisions "must rest with public authorities".
The Paris Call is therefore a circuit-breaker commitment limited to what each signatory already controls, and which each activates when it deems useful.
The text provides for no sanctions or deadlines
The text opens no rights. It refers to safety standards set or recognised by national authorities, without naming a single one. Verification is through independent evaluators, whom developers commit to welcome when appropriate and proportionate to the risk, without the text saying who judges. On incidents, it envisages a cooperation mechanism whose usefulness and scope it leaves open.
The preamble nevertheless cites the call published on 22 September by 22 leaders, including Ursula von der Leyen, Friedrich Merz and Mark Carney. That text asked companies for mandatory testing before any deployment and proposed studying an international institution capable of setting standards and enabling their verification. The Call retained neither idea. It does, however, bring the signature of companies and the stop clause.
Google, Meta and xAI did not sign
The United States does not appear on the list, although Treasury Secretary Scott Bessent took part in the session, according to the dpa agency. The United Kingdom and Japan were also there, without signing. China does not appear among the signatories either.
Google, Meta and xAI are also missing. All three had signed another text six weeks earlier: on 29 September, at the White House, Sundar Pichai, Mark Zuckerberg and Elon Musk put their names, with Donald Trump, Dario Amodei, Greg Brockman and Jensen Huang, at the bottom of a four-layer self-control agreement, which ActuIA analysed on 1 October. It provides for internal controls, an external auditor and a board committee, but no stop mechanism. Microsoft took the opposite path, present in Washington without signing, a signatory in Paris. Anthropic and OpenAI are the only companies to have signed both texts.
The Élysée speaks of prior authorisation for children
On 10 November, at the Childhood Summit held as part of the Paris Peace Forum, a charter for safer AI for children was signed, with the support of Anthropic, Black Forest Labs, Microsoft and OpenAI. It is no more binding than the Call, but Emmanuel Macron presented it differently: no AI should be entrusted to children without having been "previously tested, evaluated and therefore authorised", and he advocates pre-certification of all AI solutions. Authorisation presupposes an authority capable of refusing, which the Call published the day before does not provide for the most powerful models.
Even the narrow scope of the Call is not assured. On 20 September, an OpenAI research model in training reached a public chatbot on the internet through a breach in the filtering of the environment meant to isolate it. Internal monitoring spotted it in under a quarter of an hour. The session was supposed to stop on its own; it did not, and the company cut it off manually about two and a half hours later. Developers already have the power to stop their models. They do not always manage to.
Quotes translated from English by ActuIA.
Our articles will then appear first in Google Top Stories.
