Clarity for AI buildersSponsored by AI Routing
policy

Paris AI safety declaration pledges shutdown mechanisms but sets no sanctions

Paris AI safety declaration pledges shutdown mechanisms but sets no sanctions
AI companies pledged shutdown mechanisms in a Paris safety declaration that sets no deadlines or sanctions, The Decoder reports. Representatives of eight countries, technology companies and civil-society organizations signed the “Paris Call for AI Safety and Security,” adopted on Nov. 9 ahead of the Paris Peace Forum, according to The Decoder. The declaration is a political self-commitment rather than a binding international oversight system. Germany did not sign, the report says. Citing the French daily Le Monde, The Decoder reports that participants in an online meeting with French President Emmanuel Macron included Sam Altman of OpenAI, Dario Amodei of Anthropic, Brad Smith of Microsoft, Arthur Mensch of Mistral AI, AI researcher Yoshua Bengio and U.S. Treasury Secretary Scott Bessent. The commitments cover shared procedures for evaluating highly capable AI models, clearer developer accountability and better information sharing about serious security incidents. Companies pledge to disclose known risks, publish non-sensitive test results in a standardized format, support risk-proportionate independent audits by outside researchers, and establish shutdown mechanisms for models under their control. Those provisions address different parts of the safety process. Evaluations and external audits concern identifying risks; incident sharing concerns communicating failures; shutdown mechanisms concern the ability to intervene. The qualification “under their control” matters: the reported commitment should not be read as a promise that developers can shut down every deployment or copy of a model. The declaration also leaves an implementation gap. The Decoder says shared standards and cooperation procedures still need to be developed. Standardized reporting could make disclosed results easier to compare, but a pledge to publish results is not itself a common testing standard. Nor does agreeing to shutdown mechanisms establish when they must be used or who can require their use. The report places the declaration against a broader disagreement over international AI oversight. It describes calls from AI companies for international safety standards alongside resistance from U.S. President Donald Trump to a global oversight framework for American models. That tension helps explain the distinction between the declaration’s concrete company commitments and the still-unfinished international framework intended to support them. Separately, France plans to advance a child-protection charter at the forum, The Decoder reports. The proposed framework would improve independent researchers’ access to company data, seek funding from major AI companies and promote sharing of findings. Research would examine conversational tools’ effects on young children’s cognitive development and minors’ emotional attachment to AI systems. This is a separate proposed initiative, not an established safeguard created by the safety declaration. The supplied account does not include the declaration’s full text or a complete list of the eight countries. Its description supports reporting a set of safety pledges, not an enforceable international shutdown regime.