GENEVA – For two days in a Swiss convention city, the governments of every United Nations member state are being asked to agree on something the technology industry cannot itself resolve: who is responsible when an artificial intelligence system causes serious harm, and how the world finds out when it does.
The first UN General Assembly Global Dialogue on AI Governance opened Monday in Geneva, drawing delegations from all 193 member states to confront a problem that the technology’s own architects acknowledge they have not solved. The diplomatic ambition is significant. So is the obstacle at the center of it: the computing infrastructure that makes advanced AI possible is concentrated almost entirely in one country whose government arrived in Geneva having already rejected the meeting’s central premise.
An independent scientific panel convened by the General Assembly released its initial findings concurrent with the summit’s opening, and the conclusions were deliberately stark. The 40-member International Scientific Panel on AI, co-chaired by Turing Award winner Yoshua Bengio and Nobel Peace Prize laureate Maria Ressa, concluded that science currently cannot guarantee that as AI capabilities continue to increase, those systems will not cause catastrophic harm. The risks run from autonomous misbehavior to deliberate misuse by malicious actors. Bengio told delegates that “science currently cannot guarantee that AI will not cause catastrophic harm, either on its own or due to malicious users.”
Translated out of diplomatic register, that finding carries a clear meaning for the 193 delegations assembled in Geneva. No technical guarantee exists that advanced AI systems will reliably follow the instructions they are given. The summit is, in part, an attempt to build governance scaffolding around a category of technology whose builders cannot yet promise it will stay inside that scaffolding.
Negotiators are working through three interlocking proposals: a cross-border incident reporting mechanism that would require countries and companies to share information when an AI system causes significant harm; a set of international safety evaluation standards against which frontier models could be measured; and a liability framework for autonomous AI systems, meaning who pays when something built to act without direct human supervision goes wrong. None of these is a simple negotiation under the best circumstances. They become considerably harder when the entity most required to implement them has already signaled it does not intend to.
The United States explicitly rejected what its delegation described as “centralized global governance” of artificial intelligence before the summit opened. The position is not without logic: Washington’s concern is that a UN-administered framework would give authoritarian governments meaningful oversight of AI systems they currently cannot access on their own terms. What makes the stance awkward is the arithmetic. According to figures presented at the summit, US-based companies host approximately 75 percent of the world’s top-500 AI supercomputing capacity. The nations filing into the Palais des Nations on Monday govern 193 members of the international system. They do not, collectively, govern that 75 percent.
That gap between where AI power sits and where the demand for rules originates is precisely the kind of structural asymmetry that UN multilateral processes exist to address, and precisely the kind that they most frequently fail to resolve. Washington has been pursuing its own path: earlier on Monday, the White House and leading AI laboratories moved closer to a voluntary framework for frontier models, a domestic approach that implicitly bypasses the binding international accountability mechanism Geneva is trying to build.
Ressa, the Filipina journalist whose Nobel Prize was awarded for documenting the consequences of information manipulation, came to Geneva with a specific warning about where the current governance vacuum is already extracting costs. AI-generated content, she argued, spreads with unusual speed when paired with emotional manipulation, and the infrastructure to deploy it at scale exists today, running on the same computing clusters whose concentration in American data centers has made the Geneva talks structurally complicated from the outset.
The summit’s conveners, co-chaired by ambassadors Rein Tammsaar of Estonia and Egriselda López of El Salvador, structured the two days around a premise that has driven UN technology diplomacy since its earliest multilateral work on telecommunications: that universal participation is both harder and more durable than agreement among the willing few. The presence of all 193 member states is unusual for a technology-specific gathering of this kind. The General Assembly’s AI governance commission, announced just days ago with Nvidia’s Jensen Huang and Amazon’s Andy Jassy on its roster, had already illustrated how difficult it is to keep both industry and member states aligned on a single governance architecture. A communiqué signed by 193 governments is a different instrument from one signed by the major AI powers and their closest allies. Whether it is a better instrument depends on what it actually requires.
That question remains conspicuously open as the summit enters its second and final day. The White House’s willingness to participate in any binding outcome has not been publicly signaled. The scientific panel’s full report, released in conjunction with the summit’s opening, has not yet been fully absorbed by negotiators working through the governance text in parallel. And the liability framework, which would require legal systems across 193 countries to recognize a category of harm that most of those systems were not designed to address, is the provision with the fewest precedents and the most resistance.
What the Geneva dialogue has already accomplished is narrower but not trivial: it has produced the first UN-level scientific consensus that the systems currently being deployed at global scale carry no technical guarantee of human control. That finding will outlast whatever communiqué emerges on Tuesday. It is the ground on which every subsequent governance argument will need to be built, whether or not the country that built most of those systems agrees to be part of building what comes next.
