Paris Call for AI safety: 17 countries and 10 AI developers sign, without Google, Meta or the US

The Paris Call for AI safety and security, published by the Élysée, the French presidency, on 9 November 2026, had 17 state signatories, 10 companies and 10 civil society organisations on 14 November, according to the official English text. The United States, the United Kingdom, Japan, Google, Meta and xAI are not among them. The Call creates no legal obligation, and for models sold in the European Union, the AI Act already requires most of what it promises.

The text is short and its commitments are many: evaluate the most advanced models, publish results, be able to shut models down, report incidents, protect whistleblowers. To see what it really changes, three checks are needed: follow how the signatory list has evolved, compare it with who attended the meeting, and set each commitment against the EU rules that already apply. For the wider picture of how the EU regulates AI, see our Society & Law section.

A signatory list that nearly tripled in days

The Élysée page says the statement “remains open to new endorsements and is being updated”. It has been. In the copy of the French page saved by the Internet Archive on 10 November 2026 at 08:05 Paris time, six states had signed: Bahrain, Canada, Finland, France, the Netherlands and Norway.

On 14 November 2026, the current version lists 17. Austria, Croatia, Cyprus, Estonia, Germany, Ireland, Kenya, Latvia, Liechtenstein, Luxembourg and Portugal have joined. Twelve of the 17 are EU member states, fewer than half of the 27: Italy, Spain, Poland and Belgium, among others, are missing. The page still says the declaration was “signed on Monday 9 November 2026”, without dating each addition.

Civil society support has grown from four organisations (LawZero, SaferAI, AI Safety Connect and Ifri, the French Institute of International Relations) to ten, according to the Élysée: the Geopolitical Center for the Datasphere, GPAI Policy Lab, CeSIA (the French Center for AI Safety), the Oxford Martin AI Governance Initiative, The Future Society and FAR.AI have been added. The company list has not moved: Anthropic, Black Forest Labs, Cohere, Domyn, Hugging Face, Microsoft, Mistral AI, OpenAI, Sakana AI and Synthesia.

At the table, but not on the list

The same day, the Élysée’s statement on the “AI and security” meeting (in French) listed who took part: the heads of state or government of eight countries (Bahrain, Canada, Finland, Japan, the Netherlands, Norway, Sierra Leone and the United Kingdom), the US Treasury Secretary, the UN, the OECD, and the leaders of seven AI companies, including Sam Altman, Dario Amodei and Arthur Mensch.

Comparing that statement with the current list, four countries that attended or were represented had not signed by 14 November: Japan, Sierra Leone, the United Kingdom and the United States. Japan’s absence stands out, since the Call itself cites the G7 Hiroshima process, alongside the Frontier Safety Commitments made at the Seoul AI Summit. Neither source explains why these countries have not signed.

Conversely, three companies signed without being named among the participants: Cohere, Domyn and Synthesia.

The big names missing among the labs

Three of the leading frontier model developers are missing: Google, Meta and xAI. That is not a blanket refusal of European rules. According to the European Commission’s list of signatories, Google has signed the General-Purpose AI Code of Practice, the voluntary code that helps providers show they comply with the AI Act, and xAI has signed its Safety and Security chapter. Meta is on neither list.

Sources: Élysée and European Commission list, both checked on 14 November 2026.

Commitment by commitment: what the AI Act already requires

The Call sets no date, no sanction and no follow-up body. Its wording is cautious: the public repository “could encompass” red-teaming protocols and stress-test results, and the signatories will explore “whether a cooperation mechanism would help” on incidents (Paris Call). In the EU, providers of general-purpose AI models with systemic risk already fall under Article 55 of Regulation (EU) 2024/1689, which the Safety and Security chapter of the Code of Practice spells out. That applies to US, Canadian or Japanese labs too: under Article 2(1)(a), the regulation covers providers placing general-purpose models on the EU market “irrespective of whether those providers are established or located within the Union or in a third country”.

Sources: Paris Call, Regulation (EU) 2024/1689, Safety and Security chapter, checked on 14 November 2026.

Two commitments go beyond EU law. The first: according to the text of the Call, developers commit to “implement fail-safe mechanisms to shut down model operations within their control, including during training and internal use”. Neither the AI Act nor the Code requires such a device. The system OpenAI presented the same day gives a concrete idea of what that means, and of its limits: see what OpenAI’s Control Room kill switch really stops.

In the United States, a different route was tried three days later: on 12 November 2026, four nonprofits filed a complaint in the San Francisco Superior Court asking for a court-ordered halt on more capable models at Anthropic, Google, Meta, OpenAI and SpaceXSI. We compare it with EU law in what the lawsuit to freeze frontier AI asks, and what the AI Act allows instead.

The second: a public repository, shared across countries, of evaluation results. The EU Code only asks each company to publish summaries, with exceptions (Measure 10.2). A standardised format would make models comparable, if it ever materialises.

On independent evaluation, the Call is weaker than the Code. It makes evaluators’ access depend on what is “appropriate and proportionate to the risk” (Paris Call), whereas Measure 3.5 of the Code provides free access to the most capable model version, to its chains of thought where available, and rules out legal or technical retaliation against evaluators who publish their findings.

What it means for you

If you use ChatGPT, Claude or Le Chat in Europe, nothing changes. The Call gives you no new rights and places no new duties on the services you use. Your protections remain those of the AI Act, whose deadlines are set out in our EU AI Act timeline.

If you work in compliance or legal, your checklist does not change either: for a model with systemic risk, Article 55 and the Code are what count, under the supervision of the AI Office. The Call can serve as a signal: its signatories have publicly committed to being able to shut down their models, a point you can raise in a supplier audit.

If you work for a model developer in the EU and see a risk, protection already exists: Article 87 of the AI Act applies the EU Whistleblower Directive to infringements of the regulation. Outside the EU, the Call only commits signatory governments to offer secure channels, without saying how.

In French policy, the Call is part of a sequence: on 10 November 2026, the Élysée received a French expert commission’s 18 recommendations on generative AI and vulnerabilities, which focus on consumer chatbots rather than frontier models.

What we checked

We read the English text of the Call on the Élysée website on 14 November 2026, the French page as archived by the Internet Archive on 10 November, the Élysée statement on the meeting (available in French only), the Commission’s list of Code of Practice signatories, the Safety and Security chapter of the Code, and Articles 2, 55 and 87 of the AI Act in the Official Journal.

What we do not know: the page does not date each new signature, and we have no intermediate copy between 10 and 14 November. Neither the Élysée nor the absent governments and companies have publicly explained why they did not sign. The text does not say who will host the public evaluation repository, or when.