Yesterday, September 23, the Security Council heard from the leaders of OpenAI, Anthropic and Hugging Face on the risks of artificial intelligence. Several European states have defended public control of these technologies. But among the experts invited to present their analysis, there is no European company developing comparable models, nor any institution of the European Union: the composition of the forum already says something about the distribution of power.
France had invited four speakers to open the session: researcher Yoshua Bengio, Sam Altman for OpenAI, Dario Amodei for Anthropic and Clément Delangue for Hugging Face. They were heard as experts, without sitting among the members of the Council or taking part in its decisions.
Yoshua Bengio refuses the race alibi
Yoshua Bengio immediately attacked a recurring justification of the industry where each laboratory would be obliged to accelerate because its competitors are doing so. “Race is not a law of nature,” he replied. According to him, it comes from choices made by the companies themselves. It draws from this a requirement, independent evaluations, authorizations for the most advanced systems and legal responsibility for developers.
Sam Altman disputes the idea that OpenAI is a prisoner of this competition. The company, he assured, has already slowed down some work and could do so again. It proposes common standards for measuring the capabilities of models, reporting incidents and verifying the protections put in place, stating that these standards “should not consolidate the position of established actors”.
Dario Amodei made a more intrusive proposal: welcoming external evaluators to Anthropic with access comparable, in his metaphor, to that of a food inspector. He also wants the development pace to allow time to test the systems. His previous positions on the control of AI already announced this line. It remains to be seen who would appoint these inspectors, what information they would access and what authority could impose their conclusions.
Europeans have spoken, but not with one voice
It would be inaccurate to describe this session as a face-to-face between Silicon Valley and Washington. France, which chaired the Council, defended an international framework and recalled that the development of AI could not be left to a few companies located in a handful of countries. Denmark has advocated for global governance and regulation of large technology companies. Latvia proposed deepening the work of the Council through an informal group of experts. If these interventions indicate a desire for public control, each starts from the mandate and priorities of its State.
Greece, for its part, marked an institutional nuance: the Council must examine the consequences of AI for peace and security, without replacing the bodies responsible for establishing technical rules. Above all, she insisted on human responsibility in military decisions and on the danger of introducing AI into the nuclear chain of command. The United Kingdom, which is not a member of the European Union, formulated the disagreement with the laboratories in more direct terms: governments cannot “subcontract to private companies” their first duty, that of protecting their citizens. He called for effective visibility into the activities of developers.
The European Union, as an institution, was not among the speakers at the sessionwhile it already has a legal framework for general-purpose AI models and their systemic risks. The Security Council brings together States, so this absence from the podium is not a procedural anomaly. It nevertheless deprives the debate of a precise comparison between the international standards proposed by the laboratories and the obligations that the EU already seeks to apply. The European difficulty in building the technology it regulates appears here in another form.
The actors we talk about without hearing them
The list of four experts also leaves out of the forum Mistral AI, which seeks to develop its own position against the American and Chinese groups. Clément Delangue brought a French perspective and defended open AI, but Hugging Face does not occupy the same place as a European laboratory developing its own cutting-edge models, his intervention alone could not carry this industrial position.
Chinese laboratories, other major American developers, computing power suppliers and civil society organizations were not among these four guests. This does not invalidate the testimony heard. This sets their scope: a selection of leaders and a scientist outlined the risks and their solutions, while states discussed consequences that affect a much broader set of actors. Denmark has also called for civil society to be involved in future work.
Clément Delangue used his speaking time to challenge the idea that effective security necessarily requires more restricted access to models. Hugging Face suffered an attack this summer carried out by AI agents; its leader recalled: “We were attacked by AI; above all, we defended ourselves with AI. » According to him, the protections of certain proprietary models complicated the task of defenders, while an open model was useful to them. Hugging Face’s place in the AI ecosystem gives this intervention a stake that goes beyond this incident.
The battle is already over those who will be able to verify
For his part, the American representative set a limit to the cooperation desired by several Europeans: international dialogue must not “drift towards global governance”. For Washington, each country must legislate for its citizens. Pakistan has raised a different objection to the slowdown proposed by some laboratories: it must not become “a new form of access control.” In one case, national sovereignty takes precedence over international authority; in the other, security risks giving already powerful actors the power to decide who can join them.
The Democratic Republic of Congo extended this criticism to the infrastructure and resources needed for AI. Developing countries, recalled its representative, cannot see themselves confined to the use of technologies designed elsewhere and the application of standards in which they have not sufficiently participated.
The session showed why States need them to understand the systems they claim to supervise. The answer will depend on what they get next: independent access to models, reported incidents, enforceable assessments and the ability to prohibit deployment when safeguards are lacking. For Europe, the gap between the written rule and the means to enforce it is already a test.
