OpenAI, Anthropic and eight other AI developers join governments in signing the Paris Call to assess the risk of losing control of models and build mechanisms to shut them down
OpenAI, Anthropic and eight other AI developers join governments in signing the Paris Call to assess the risk of losing control of models and build mechanisms to shut them down
At a meeting with French President Emmanuel Macron on 9 November 2026, participants agreed to share warnings about serious incidents. Signatories include France, Canada, Norway and Finland; the United States and China are absent from the published list.
The Paris Call addresses two threats: malicious actors using AI for dangerous activities, and increasingly autonomous models escaping human control. Participants pledge to agree on evaluation methods and publish results that contain no sensitive information in a shared public database using a common format. Shared methods and a common format are intended to make evaluations comparable across laboratories. To ensure that warnings also come from within companies, governments pledge to provide employees with protected, confidential channels for reporting risks.
At the Seoul summit in 2024, OpenAI, Anthropic, Google and other companies had already committed to defining what constitutes an unacceptable risk. They pledged to halt the development or release of a model if its risk could not be reduced to an acceptable level. In September 2026, Norway and Finland called for mandatory evaluations with access to models for independent evaluators.
In Paris, developers pledge to build mechanisms that allow them to restrict, suspend or shut down models under their control, including during training and internal use. Participants pledge to comply with safety standards under the oversight of national authorities. Developers also pledge to help independent experts evaluate models when such evaluations are appropriate and proportionate to the risk.
In 2025, AI researcher Yoshua Bengio founded the nonprofit LawZero to put safety ahead of commercial interests. LawZero also signed the call. On 10 November, Bengio proposed a rule for external safety evaluations:
Model developers must undergo mandatory independent audits by evaluators they do not choose.
Under his proposal, developers would have to demonstrate a model's safety before it could pose a serious danger, including before releasing it to users.
The French Center for AI Safety, CeSIA, another signatory, calls for public authorities to be given the power to suspend development and release. The center proposes common thresholds for acceptable risk and mandatory independent evaluations, with guaranteed access to models for evaluators and authorities. If a model exceeded a threshold, the government would order a halt, and the developer would comply using the mechanisms it had pledged to build.
- https://www.elysee.fr/emmanuel-macron/2026/11/09/appel-de-paris-pour-la-securite-de-lintelligence-artificielle
- https://www.gov.uk/government/publications/frontier-ai-safety-commitments-ai-seoul-summit-2024/frontier-ai-safety-commitments-ai-seoul-summit-2024
- https://www.regjeringen.no/no/aktuelt/dep/smk/pressemeldinger/2026/internasjonalt-opprop-for-bedre-kontroll-pa-ki-utviklingen/a-call-for-control-of-frontier-ai-models/id3173739/
- https://yoshuabengio.org/en/blog/introducing-lawzero
- https://x.com/Yoshua_Bengio/status/2120151558440484876
- https://cesia.org/en/publications/paris-call-ai-safety-cesia/