On November 9, six governments and ten AI companies signed a joint declaration on AI safety and security in Paris, pledging shared evaluation standards, independent oversight and fail-safe controls for the most advanced models. The declaration was published by the Élysée Palace and the Canadian Prime Minister's Office, and the Dutch government posted the full text on its website on November 10. It is the second time this year, after the Seoul summit, that governments and leading model developers have been brought onto the same safety commitment.
The governments signing are Bahrain, Canada, Finland, France, Norway and the Netherlands; on the company side are Anthropic, Microsoft, OpenAI, Mistral AI, Hugging Face, Cohere, Black Forest Labs, Domyn, Sakana AI and Synthesia, joined by four civil society organizations, and the declaration remains open to new signatories. The core commitments are fourfold: share non-sensitive evaluation findings in a common format; build incident reporting and early warning systems; disclose known risks; and equip models with fail-safe mechanisms, meaning the ability to shut down model operations within one's own control when necessary. Governments, in turn, pledged to protect whistleblowers in AI safety.
The declaration does not come out of nowhere. It explicitly builds on the Hiroshima Process, the Frontier Safety Commitments of the Seoul AI Summit, the G7 Evian AI meeting and the United Nations Security Council meeting on AI held on September 23. The reasoning is stated plainly: the impact of frontier models crosses borders, and no single country, company or international organization can manage the risks alone. A useful comparison is OpenAI's recently disclosed internal system Control Room, which can halt all research model training within 35 seconds of a decision to intervene — the kind of 'ability to stop' the declaration asks every developer to have.
But the declaration's binding force remains to be seen. It is a political commitment, not a treaty, and it says nothing about consequences for non-compliance; implementation will depend on what each company and government does next. National legislation is moving in the same direction: Australia recently proposed frontier AI rules requiring companies to report incidents quickly and setting up a dedicated regulator. The declaration's value lies more in setting a shared direction: evaluation standards, incident reporting and emergency shutdown are becoming the default configuration of frontier model governance.
