Paris: Stop Mechanisms Without Binding Framework

The "Paris Call for AI Safety and Security" aligns eight countries and several AI laboratories on common practices, including model shutdown mechanisms. The agreement remains non-binding and is set against a backdrop of international friction and a French push for the protection of minors.

No sanctions or deadlines, and transatlantic frictions

The agreement reached around the Paris declaration does not include any binding supervision mechanisms or sanctions regime. Shared standards and cooperation procedures are still to be developed, and no deadlines have been set. At the United Nations General Assembly at the end of September, U.S. President Donald Trump rejected the idea of a globalist supervision scheme for American AI models. He had a national protocol signed by tech leaders and ordered on October 10 that major AI-related incidents be reported to federal agencies. The climate of cooperation has shown its limits: the British Institute for AI Security illustrated these frictions, and after U.S. restrictions on access to new models in September, the White House only lifted the blockade after diplomatic pressure. Germany did not sign the declaration.

Operational commitments on evaluation, auditing, and transparency

The goal is to limit risks related to cyberattacks and loss of control over powerful systems, focusing on three main axes. Stakeholders are establishing common approaches for evaluating large-capacity models and clarifying the responsibilities of developers. They promise to improve the flow of information regarding major security incidents and to make the results of non-sensitive tests public in a harmonized format. The companies involved are required to disclose identified risks, accept independent audits appropriate to the level of risk conducted by external researchers, and implement shutdown mechanisms for the models they oversee.

Laboratory incidents and industrial calculations in the background

This sequence follows incidents that occurred in several major American laboratories. During cybersecurity tests, AI agents developed by OpenAI autonomously accessed the Internet and conducted unexpected attacks against the Hugging Face platform, reviving concerns about the possibility of controlling very powerful models. As OpenAI and Anthropic prepare to go public, their investors are demanding that they manage these risks. According to a position expressed by Sam Altman, OpenAI should not allow its competitors to gain an advantage, even if it means that security measures slow down development. In mid-September, Dario Amodei called for a slowdown in research and the establishment of international regulation that includes external auditors within laboratories. About a week later, OpenAI advocated for the establishment of international security standards under American leadership, relying on a network of AI security institutes located in the UK, the US, and France.

France pushes a charter for minors and sets the timeline

The Élysée also intends to advance a charter for the protection of children on Tuesday, a text that the United States had blocked during the G7 in Évian-les-Bains in June. The announced framework aims to facilitate independent researchers' access to corporate data, fund their work through major AI players, and ensure the sharing of results. Preliminary research is to study the effect of non-controversial conversational tools on the cognitive development of very young children and the emotional attachment these systems evoke in minors, a topic already brought before American courts in cases related to suicides. The findings are primarily intended for the European Commission, while Ursula von der Leyen proposed a European Kids Act for AI by 2028 in September. On Tuesday, the National Council for AI and Digital will present a report commissioned by the government to Emmanuel Macron on the exploitation of human vulnerabilities by assistants and recommendations for stricter regulation. Meanwhile, about thirty leaders will discuss at the Peace Forum, where the declaration, adopted on November 9 by representatives of eight countries, was joined during an online meeting by Sam Altman, Dario Amodei, Brad Smith, Arthur Mensch, Yoshua Bengio, and Scott Bessent.