OpenAI, Anthropic and Microsoft are among 10 AI companies that signed a declaration Monday committing frontier developers to build mechanisms for shutting down their models and to share warnings of serious incidents, according to the text published Monday night by Canada's Prime Minister's Office. The United States and the United Kingdom are not among the signing governments, leaving the labs that many U.S. companies and agencies buy from making safety promises alongside other countries instead.
What Happened
The "Joint Declaration on Paris Call for AI Safety and Security" lists Bahrain, Canada, Finland, France, Norway and the Netherlands as signing governments. The Élysée Palace's version lists the same signers. The companies listed are Anthropic, Black Forest Labs, Cohere, Domyn, Hugging Face, Microsoft, Mistral AI, OpenAI, Sakana AI and Synthesia. Google, Meta and xAI do not appear. Civil-society signers include LawZero, the group founded by AI researcher Yoshua Bengio.
French President Emmanuel Macron convened about 50 participants by video link, the German news agency dpa reported. According to dpa, the participant list included the U.K., Japan and U.S. Treasury Secretary Scott Bessent, as well as OpenAI's Sam Altman, Anthropic's Dario Amodei and Microsoft's Brad Smith.
Developers commit to "implement fail-safe mechanisms to shut down model operations within their control," including during training and internal use, to disclose model capabilities and known risks, and to address the stripping of safeguards from open-weight models, whose internal parameters are publicly released. Signers back a common framework for testing frontier models for cyber, chemical, biological, radiological, nuclear and loss-of-control risks, and agree to publish non-sensitive evaluation results in a standard format. Governments commit to secure channels for whistleblowers and to clarify legal liability for large-scale and cross-border harm.
Why It Matters
The declaration is voluntary: it sets no deadlines and no enforcement mechanism. Even so, it gives corporate buyers a written benchmark to hold suppliers to, such as whether a vendor can actually halt a model and whether it reports serious incidents to governments.
The U.S. absence widens the gap between Washington and allies on AI oversight. The Primary Wire reported in October, citing an Axios interview, that Bessent said AI leaders who want rules should slow down on their own. Washington has pursued its own track, including an AI incident channel with Beijing that The Primary Wire reported in September, citing Axios.
What to Watch
The text says it remains open to new endorsements. Watch whether Google, Meta, xAI, the U.S. or the U.K. sign, whether signers publish evaluation results in the promised format, and what the parties decide after they explore a cooperation mechanism for sharing incident reports.
