
What the Declaration Says
A joint declaration on the Paris Call for AI safety and security says it was signed on Monday, November 9, 2026, by the governments, companies, and civil society organizations it lists. It sets three commitments: a common framework for evaluating and managing frontier-model risk, accountability by role, and information sharing, reporting, and early warning.
Signatory lists differ by copy. The Élysée page lists 18 governments, ten companies, and 11 civil society organizations. Both the Canadian and Finnish copies list six governments, the same ten companies, and four civil society organizations. A Dutch copy dated November 10 lists nine governments, nine companies without Sakana AI, and no civil society organizations. Three copies say the statement remains open to new endorsements and is being updated, and these counts are as retrieved on November 16. Google DeepMind, Meta, and xAI appear on none of the four copies.
What an Evaluator Could Gain
The first commitment says policy makers, researchers, and developers are committed to establishing a common framework for evaluating and managing frontier-model risk, with a stated focus on mitigating dangerous misuse, including in the cyber and CBRN domains, and risks from growing autonomy or model misalignment, including loss of control. That framework "would be supported by" common technical standards and evaluation tools. Signatories also agree to "publish the non-sensitive findings of our evaluations in a standardized format" and to create a publicly accessible shared repository.
On outside evaluation, one sentence commits signatories to "strict compliance" with safety standards that national authorities establish or recognize, with "credible, competent and independent third-party evaluation". The developer pledge is narrower. Developers commit to disclose model capabilities, limitations, and known risks and, "where appropriate and proportionate to the risk", to support evaluation by qualified independent evaluators, "as appropriate and consistent with existing practices and arrangements". They also commit to fail-safe mechanisms that shut down model operations "within their control", including during training and internal use.
Limits and Open Questions
The text sets no deadline or enforcement mechanism, and it names no operator or address for the repository. It does not define "non-sensitive" or state whether the commitments are legally binding. On incidents, it leaves open whether a cooperation mechanism would help and what its scope could be, and the whistleblower commitment belongs to policy makers.
Developer support for independent evaluators applies only where appropriate and proportionate to the risk, and the text does not say who decides or how evaluators would obtain model access. CeSIA, which the Élysée page lists as a signatory, says the text sets no common risk thresholds, no clear guarantee of model access for evaluators and public authorities, and no suspension mechanism controlled by public authorities. We found no evidence in these sources that any signatory has implemented a commitment. Our editorial reading: an auditor can cite these commitments when requesting standardized findings or independent evaluation, but the text creates no access right and sets no deadline.
Follow Audit Commons
Keep up with new reporting, practical guides, and resources in your feed reader.