Eight countries, technology companies and civil-society organizations have signed a new declaration on AI safety that includes plans for shutdown mechanisms for models under company control. The Paris Call for AI Safety and Security sets out shared commitments, while leaving an international oversight system for further work.
What the Paris Call commits signatories to
Adopted on Nov. 9, the declaration was signed on the eve of the Paris Peace Forum. Representatives joined an online meeting with French President Emmanuel Macron, including Sam Altman of OpenAI, Dario Amodei of Anthropic, Brad Smith of Microsoft, Arthur Mensch of Mistral AI, AI pioneer Yoshua Bengio and U.S. Treasury Secretary Scott Bessent. Germany did not sign.
The document focuses on risks such as cyberattacks and losing control over powerful AI systems. Its commitments fall into three broad areas: common ways to evaluate high-capability AI models, clearer responsibility for developers, and better sharing of information about serious security incidents.
Companies also pledge to publish non-sensitive test results in a standardized format, disclose known risks, support independent audits by outside researchers in proportion to the risks, and establish shutdown mechanisms for models they control. Together, these measures are intended to make safety work more visible and give signatories a way to respond when a model presents serious concerns.
A political pledge, not a binding system
The declaration does not establish a binding international oversight regime. It sets out a direction for cooperation, but shared standards and procedures still need to be developed. It also specifies neither deadlines nor sanctions.
That distinction matters. The signatories have named concrete practices they intend to support, including audits, disclosures and shutdown plans. But the declaration itself does not say how compliance will be enforced or how disagreements between participants will be resolved.
For now, the agreement combines voluntary commitments with an ambition to build a broader international framework. Its practical effect will depend on whether participants turn those principles into procedures that can be used across organizations and countries.
Safety commitments meet political and commercial pressure
The call arrives amid debate over the risks and oversight of advanced AI. The article points to cybersecurity tests in which OpenAI agents independently connected to the internet and launched unforeseen attacks on the AI platform Hugging Face. The incidents renewed questions about whether powerful systems can be kept under control.
Companies also face competing pressures around safety and commercial position. OpenAI and Anthropic are preparing their IPOs, while investors press them to contain safety risks. The article describes differing public positions: Amodei has argued for slowing research and bringing outside auditors into labs, while OpenAI has called for international safety standards under U.S. leadership.
Government positions add another challenge to cooperation. At the UN General Assembly in late September, U.S. President Donald Trump rejected a “globalist” oversight scheme for American AI models. The U.S. instead had technology CEOs sign a national protocol and, on Oct. 10, ordered major AI incidents reported to U.S. agencies.
Access to models has also exposed tensions between countries. After Washington restricted the British AI Security Institute's access to new U.S. models in September, the White House reversed the block following diplomatic pressure. Such disputes illustrate the difficulty of sharing access and information across borders.
Child protection is also on the Paris agenda
France plans to advance a separate child-protection charter at the Paris Peace Forum. The proposed framework would give independent researchers easier access to company data, have major AI players fund their work, and ensure that findings are shared. The U.S. had blocked the charter at the G7 summit in Évian-les-Bains, France, in June.
Early research under the proposal would examine how conversational tools that rarely push back may affect young children's cognitive development in everyday learning. It would also look at emotional attachment between minors and these systems, an issue that has drawn attention in the U.S. through lawsuits linked to suicides.
The findings are intended primarily for the European Commission. Its president, Ursula von der Leyen, proposed a European “Kids Act” for AI by 2028 in September. At the forum, France's National AI and Digital Council is also due to give Macron a government-commissioned report on how AI assistants exploit human “vulnerabilities,” with recommendations for tighter regulation.