On Monday, the heads of state and government of Bahrain, Canada, Finland, France, Norway and the Netherlands signed a joint declaration with ten AI companies, including Anthropic, Microsoft and OpenAI, and four civil-society organizations. The text pledges a common framework for evaluating frontier models, a shared repository of evaluation findings, disclosure of model capabilities and risks, fail-safe mechanisms that can shut down model operations within the relevant actor's control, and stronger systems for sharing information about serious incidents.
What decides how much the declaration is worth sits in two places: the signatory list, and the qualifying clauses.
The list is the limit
The United States, the United Kingdom and China are not among the initial signatories. Neither are Google, Meta, Amazon or Nvidia. The initial list leaves major governments and developers outside the declaration, limiting its participation rather than establishing refusal or noncompliance. Companies based in a non-signatory country can still sign, and Anthropic, Microsoft and OpenAI did.
The declaration itself leaves the door open. It says the statement "remains open to new endorsements and is being updated." So the accurate reading is not that Washington, London or Beijing refused. It is that, at signing, they had not joined a framework whose central claim is that no single state or company can manage these risks alone.
The qualifiers do the work
The declaration also commits signatories to strict compliance with rigorous safety standards established or recognized by competent national authorities, under their oversight and with credible, competent and independent third-party evaluation. That clause invokes national standards and oversight rather than establishing a new international enforcement system. It is the strongest-sounding commitment in the text, and it points back at rules the signatory states already run.
Read the operative sentences on the new pledges closely and a pattern forms. Signatories agree to publish evaluation findings, but only the "non-sensitive" ones. Developers commit to support independent evaluation, but "where appropriate and proportionate to the risk," and again "as appropriate and consistent with existing practices and arrangements." Fail-safe mechanisms apply to model operations "within their control." On information sharing, the text commits signatories to strengthen existing arrangements and to explore "whether a cooperation mechanism would help, and what its scope could be." Policymakers separately commit to secure and confidential channels for whistleblowers to raise concerns about systemic risks.
Each qualifier is defensible on its own. Together they mean the declaration leaves important implementation questions unresolved, including how these qualifications will be applied and reviewed. There is no deadline for the shared repository, no date for the common framework, and no penalty in the text for signing and then doing none of it.
This is not a reason to dismiss the document. Comparable evaluation methods and a standard format for publishing results are real infrastructure, and the networks it builds on, including the International Network for Advanced AI Measurement, Evaluation and Science, already exist. A shared vocabulary for incidents is useful groundwork for any binding rule that comes later.
What would change that assessment is specific and checkable: a US, UK or Chinese endorsement; a published repository with actual evaluation findings in the promised standard format; or a signatory government turning the disclosure and fail-safe pledges into a domestic rule with a date attached. Until one of those happens, the Paris Call is best read as a list of who is willing to be counted, and a draft of the terms on which they might later be bound.
Sources
- [1] Élysée, "Paris Call for AI safety and security" (November 9, 2026) — full text read first-hand by the desk; every quotation, signatory list and qualifier verified against itRead source
- [2] Prime Minister of Canada, "Joint declaration on Paris call for AI safety and security" (November 9, 2026)Read source
- [3] Office of the President of Finland, "Paris Call for AI safety and security" (November 9, 2026)Read source
- [4] Precise Impact AI, "AI Daily Brief: 10 November 2026" (November 10, 2026) — summary account of the declaration's commitmentsRead source