Everyone Agreed AI Needs Rules in Geneva. Nobody Agreed on Whose
Governments gathered in Geneva on July 6 and 7, 2026, to discuss a technology crossing borders in milliseconds. Every enforcement tool for it still sits inside national or regional jurisdiction. The UN’s first Global Dialogue on AI Governance built a room where every member state could join companies and civil-society groups at the same table. It did not build a rulebook.
What Geneva Built
The dialogue opened Geneva Digital Week, overlapping with the start of the International Telecommunication Union’s AI for Good Global Summit and connecting with the World Summit on the Information Society Forum programme. The UN designed the process to give all 193 member states an equal voice, alongside private companies, academics, civil-society organizations and the technical community. Organizers also created a travel-support process, funded through voluntary contributions, to help eligible stakeholders from developing countries participate.
Secretary-General António Guterres opened the session with a warning rather than a celebration. “The choice before us is not between faith in AI or fear of it,” he said. “It is between governing by design and drifting by default.” He argued that aviation, medicine and nuclear energy earned public trust because their makers faced accountability, and that AI needs the same bargain.
Although the dialogue focused on civilian AI, Guterres also renewed his call to prohibit lethal autonomous weapons systems that can make life-or-death decisions without meaningful human control. He proposed an AI Child Safety Pledge requiring developers to demonstrate that systems intended for children meet basic protections before release. The proposals supplied concrete examples of the risks governments could discuss, but not an enforcement system capable of managing them.
None of it produced a treaty. The UN scheduled a second session for New York in May 2027, while the main sessions and side events focused on safety, equitable access, technical standards, capacity-building and interoperability rather than negotiating treaty language. The dialogue exists to compare approaches and build common ground, not to legislate.
Why Agreement Stayed Out of Reach
Delegations arrived with different priorities, and the differences run deeper than drafting disputes. Frontier AI powers, the United States chief among them, protect national security interests and domestic industry advantage. The European Union pushes risk classification, transparency requirements and product obligations through its AI Act model. China backs international cooperation while maintaining a distinct domestic regulatory system built around state oversight.
Compute access, local-language models and a real vote in standard-setting matter most to developing economies, which want more than a seat at the table. For many countries, the immediate governance problem does not begin with frontier-model catastrophes. It begins with the price of infrastructure, the availability of training data and the risk that another country’s technical standards become unavoidable simply because its companies supplied the systems.
Technology companies want predictable rules that work consistently across markets. Civil-society groups want transparency, rights protections and limits on surveillance and autonomous weapons. Add cloud providers, chipmakers, national-security agencies and regional regulators to the list, and consensus stops looking like a communication problem. Power, infrastructure and economic benefit sit unevenly across the room, and no forum changes that by gathering people in Geneva.
Standards Could Outpace the Treaties
Political agreement moves slowly. Technical coordination can move faster, and the ITU’s work on AI agents shows how.
On July 9, the ITU announced the Focus Group on Trust and Identity for Humans and Agentic AI. The initiative will examine digital identity, trust management, interoperability, security and accountability as autonomous systems begin interacting with people and other agents across platforms. The underlying questions sound simple until software starts scheduling work, negotiating contracts or moving money: who is the agent, how can another system verify it, who remains responsible for its decisions and how does a human retain authority over its actions?
The group plans its first formal meeting in Paris in November 2026, followed by another in Geneva in January 2027. Its work carries no legal force yet. Voluntary technical standards still shape markets before regulators catch up, though. Companies incorporate them into procurement and assurance programmes, while regulators and insurers can use the same frameworks when assessing compliance and risk.
Watch the focus group closely. It sits closer to real-world deployment than most of what happened inside the main dialogue hall. A standard for verifying an agent’s identity or recording the limits of its authority could influence products and contracts long before governments agree on an international AI law.
China Reframes the Question From Capability to Distribution
Coverage of AI competition keeps reaching for a horse-race framing: is China catching up to the United States? The framing misses the larger shift. China fields major research laboratories, manufacturing scale, platform reach and a state-directed industrial policy for AI. Chinese developers, including DeepSeek and Alibaba’s Qwen, have also built efficient open-weight models drawing international attention.
No single metric settles whether China has caught up with the United States. The answer changes depending on which measure someone picks: frontier benchmark scores, patents, advanced-chip access, cloud capacity, cost efficiency, manufacturing depth or global adoption.
The sharper question is about distribution, not a scoreboard. Universities, startups and public agencies across Africa, Southeast Asia and Latin America commonly lack the budgets for premium frontier-model subscriptions or the infrastructure needed to train advanced systems. Open-weight models running on less expensive hardware close part of that gap, and Chinese laboratories have made access and openness central to their international pitch.
Events after Geneva made that institutional competition harder to ignore. On July 16, representatives from 29 countries signed an agreement in Shanghai establishing the World Artificial Intelligence Cooperation Organization, or WAICO. The agreement describes WAICO as an independent intergovernmental organization headquartered in Shanghai.
China presents the organization as a platform for international cooperation, wider technology access and stronger participation by the Global South. Its creation does not replace the UN dialogue, and its practical authority, funding model and governance procedures remain to be tested. It does give Beijing a standing institution through which it can promote its preferred approach to AI development and governance.
Fragmentation no longer describes competing policy statements alone. It now describes competing organizations.
Governance follows infrastructure. Whichever countries supply the compute, models, cloud platforms and technical training also supply many of the practical rules that users inherit, no matter what a UN dialogue eventually decides.
Fragmentation Becomes a Distribution Layer
Policy divergence once sat on top of product design as a legal review after engineering finished the work. It sits inside the architecture now.
An AI provider may need different transparency notices, risk documentation, evaluation records, data controls and feature availability across the European Union, the United States and Asian markets. Data-residency laws shape where a company hosts models and how information moves across borders. Export controls, sanctions and licensing rules determine which chips, cloud services and advanced AI technologies reach each market.
Public-sector procurement rules create another layer. A government agency may require specific evaluation records, data-handling assurances or human-oversight controls even when national law imposes no equivalent obligation. Those contract requirements function as technical standards because vendors must build for them to enter the market.
Companies selling AI products across borders increasingly maintain jurisdiction-specific documentation, evaluations, access controls, hosting arrangements and feature configurations. Regulation stopped being a legal overlay somewhere in the past two years. It became part of the technical distribution stack, sitting next to bandwidth, latency and hosting cost as a variable shaping what a product can do in a given market.
What Leaders Should Do Now
Executives do not need to wait for Geneva’s next session to act. A jurisdictional control map, tracking laws, standards, procurement rules, export controls and data requirements by market, gives a company a starting inventory instead of a scramble during an audit.
A baseline set of global controls covering model governance, access management, evaluation, incident response and human oversight lets a company add local overlays without rebuilding its compliance system for every market. Products need room for regional feature flags, model substitution and local hosting built into the architecture rather than bolted on after a regulator objects.
Compliance teams should track technical standards from the ITU and ISO/IEC alongside policy frameworks from organizations such as the OECD. Voluntary frameworks can become procurement conditions and contractual requirements before legislators write them into law.
Supply-chain documentation matters just as much. Companies need records covering model providers, fine-tuning data, evaluation methods, infrastructure locations and critical hardware dependencies. Export restrictions and sanctions can alter access to models, chips or cloud infrastructure faster than many product roadmaps can adapt.
None of it requires waiting for a UN treaty still years away, should one arrive at all.
What to Watch After Geneva
The New York session in May 2027 remains the UN process’s next formal test. It will show whether the Geneva conversation produced anything beyond a shared calendar and a venue for recurring debate.
Before then, WAICO’s membership, charter, funding model and relationship with existing international bodies will show whether China has created a substantive governance institution or mainly a diplomatic coordination platform. The organization’s first projects will matter more than its founding language. Technical assistance, compute access, model distribution and training programmes would give it influence that declarations alone cannot.
The ITU focus group’s meetings in Paris in November 2026 and Geneva in January 2027 will show whether agent identity and accountability work can produce usable standards. Equal attention belongs to funding for developing-country participation, movement toward interoperable assurance systems and proposals for more permanent international AI institutions.
So does the fate of open-weight models under export-control regimes. A licensing decision in Washington, Beijing or another major technology centre can reshape which models, chips and cloud services the rest of the world can use.
Geneva deserves neither dismissal nor overstatement. The dialogue produced no treaty, regulator or enforcement mechanism, and none of that was its immediate purpose. Its value sits in the room itself: a venue where governments, companies and advocacy groups compare approaches to a technology none of them fully controls.
Its limitation sits outside that room, in the compute, chips, models, standards and market access still held by a small group of states and companies. The rules that matter most will keep forming there, through multiple institutions and competing distribution networks, with or without a single UN rulebook.
