Human Control Over AI: What the UN Debate Really Changes

The Security Council held its 10,228th meeting on the topic of artificial intelligence and international security. Sam Altman, CEO of OpenAI; Dario Amodei, CEO of Anthropic; and representatives from Hugging Face and the UN’s International Scientific Panel presented their analyses to the member states.^[1][2]

OpenAI has advocated for cooperation among governments and for decisions guided by democratic processes. Anthropic emphasized the need to slow down deployment when safety has not been sufficiently demonstrated. These positions constitute public statements from companies directly involved in the race to develop advanced models. They shed light on the debate, but do not constitute an independent assessment or a legally binding commitment.^[2][3]

The meeting did not result in a Security Council resolution, a treaty, or the creation of a global regulatory body. It did, however, elevate AI governance to the level of an international security issue, moving beyond the usual debates on data protection, competition, or product liability.

The call for international governance of AI is not new. The OECD principles, the Council of Europe’s Framework Convention, the Global Digital Compact, and national or regional regulations have already established multiple layers of rules. What is new is the more explicit linking of three concerns: the capabilities of state-of-the-art models, the behavior of agents capable of acting in external environments, and the risk that industrial competition will reduce the time devoted to oversight.

A statement released on September 21 by the Finnish presidency—originally proposed by Finland and Norway—has received the support of leaders from more than twenty countries. It calls for transparent safety protocols, pre-deployment testing, independent assessments, shared reporting of serious incidents, and consideration of an international institution capable of intervening when certain capacity thresholds are exceeded.^[4]

The text also states that risk management should not widen the gap between countries capable of developing or evaluating models and those that depend on foreign technologies. This clarification is important: a governance framework that protected only the best-equipped countries would fall short of part of its objective.

MSc AI Governance, Risk & Compliance

The Bac+6 program that trains legal professionals to understand AI and Data technologies so they can master their governance, risks and compliance.

12 months — Bac+6 Entry with a Bac+5 Work-study internship, initial or continuing education Paris-Villejuif campus

Discover the program → RNCP Level 7 certification

Global governance does not necessarily imply a single authority with direct power over all models. The most realistic scenario instead resembles a network of complementary functions, with a common scientific foundation, compatible technical standards, and implementation carried out by national or regional authorities.

The UN already has two building blocks in place. In August 2025, the General Assembly established an Independent International Scientific Panel on AI and a Global Dialogue on AI Governance. The former brings together forty experts serving in their personal capacity, while the latter provides a forum for discussion among all states and stakeholders. Their mandate is to develop a common knowledge base and facilitate cooperation, not to issue authorizations or sanctions.^[5][6]

The disagreements are not limited to risk assessment. AI has become a matter of industrial, military, and economic power. Semiconductors, computing, data, the cloud, and the most powerful models are all integrated into sovereignty strategies. A state can therefore uphold security principles while rejecting an institution that might restrict its businesses or disclose strategic information.

The United States and China did not join the statement put forward by Finland and Norway. This absence shows that the support of the major technological powers cannot be taken for granted. However, it does not mean that they refuse to cooperate at all. Bilateral channels, technical standards, and national mechanisms can move forward without taking the form of a single global regulator.^[3][4]

Administrative capacities also vary widely. Not all countries have evaluation laboratories, specialized authorities, or sufficient access to models. An audit requirement designed by the most advanced economies could prove difficult to implement elsewhere, or even increase dependence on a few suppliers or private firms.

Finally, the companies calling for regulations are also the ones that would be subject to them. Their technical expertise is necessary, but it should not allow them to single-handedly define thresholds, evaluation methods, or the information made public. Credible governance must involve independent researchers, public authorities, civil society, users, and countries that are underrepresented in the AI value chain.

Three findings have been established. First, the Security Council now recognizes that certain developments in AI may affect international peace and security. Second, a coalition of leaders supports independent assessments, incident reporting, and the maintenance of human oversight. Third, the UN already has a scientific panel and a global dialogue that can serve as an institutional foundation.^[1][4][5]

These factors do not demonstrate that a regulatory consensus exists. The statement remains political and voluntary. No international capacity threshold has been defined, no verification procedure has been adopted, and no authority has been tasked with preventing the deployment of a model deemed too risky.

OECD data illustrate this gap between principles and implementation. In 2025, 30 of the 36 countries studied had at least one institution responsible for AI governance in the public sector. However, only 14 required a risk assessment prior to deployment, 12 had internal review committees, 11 conducted post-deployment audits, and 6 maintained an open algorithmic registry.^[8]

The situation therefore justifies neither the notion of a complete vacuum nor that of a global governance system that is already operational. The institutions exist, but their powers, methods, and coordination remain fragmented.

Executive MBA AI & Business
Transformation

The MBA reimagined for the age of AI. For experienced executives who want to lead their organization’s transformation. Paris, Nice & Dubai.

12 months — Part-time 10 years of experience min. Early bird €20,000 Paris · Nice · Dubai

Discover the program → 100% in English

The first challenge is defining risk. A threshold based on computing power is measurable but does not fully capture actual capabilities. A threshold based on cybersecurity, manipulation, or autonomy tests may be more relevant, but it depends on the chosen protocol. These measures must be subject to review and accompanied by explicit statements of uncertainty.

The second challenge concerns evaluators’ access. Verifying an advanced model may require access to the system, logs, training methods, or internal incidents. Legitimate trade secrets must be protected without turning confidentiality into a permanent obstacle to auditing. Accredited evaluators—who are independent of the provider and subject to security obligations—are one possible solution, but their funding and accountability must be clarified.

The third issue is a democratic one. Decisions regarding the acceptable uses of AI touch on fundamental rights, labor, information, and security. They cannot be left solely to technical experts. The Council of Europe’s Framework Convention provides a legally binding example centered on human rights, democracy, and the rule of law, but it specifically excludes national defense activities from its direct scope. This limitation serves as a reminder that the most sensitive areas are also those where cooperation is most difficult.^[9]

Finally, human oversight must be defined in concrete terms. It may involve validation prior to an action, the ability to shut down a system, assigning responsibility to a competent individual, or providing a recourse mechanism for those affected. A general provision without the authority to intervene, without an activity log, and without a means of appeal risks remaining merely symbolic.

The first issue will be the national implementation of the declaration. Will the signatories adopt common testing and reporting requirements, or will they stick to general commitments? The publication of a shared format for serious incidents would constitute verifiable progress.

It will also be important to monitor the work of the UN’s International Scientific Panel, particularly its ability to access data from companies and maintain its independence. Its usefulness will depend less on the number of reports published than on their influence on standards, regulatory bodies, and deployment decisions.^[6][7]

The participation of the United States and China will remain crucial. A universal agreement seems unlikely in the short term, but more targeted mechanisms are possible: reporting of major incidents, common assessment methods, protection of critical infrastructure, or coordination among public security laboratories.

Global governance of AI could thus emerge in successive stages rather than through a single treaty. The September 23 meeting brought these differences into sharper focus. Above all, it shifted the focus of the question: it is no longer enough to ask whether international cooperation is desirable, but rather what evidence, what institutions, and what responsibilities can make it effective without handing over control of the technology to those who develop it on their own.