By Sadaf Sundas Riaz — SCN News
UNITED NATIONS, Sept. 20 — Artificial intelligence is moving into the centre of diplomacy at the United Nations this week, but the most consequential AI talks may not happen inside the General Assembly Hall at all. As Secretary-General António Guterres urges countries developing the most powerful systems to agree on common safeguards, U.S. and Chinese officials are already negotiating in New York over possible AI guardrails ahead of a meeting between President Donald Trump and President Xi Jinping in Washington — a striking demonstration of both the promise and the weakness of the UN’s attempt to build genuinely global AI governance.
The development gives UNGA81 a sharper AI question than the debate SCN has already examined over whether rapidly advancing systems could eventually threaten humanity. The immediate diplomatic problem is now one of power and coordination: Washington and Beijing are the two principal forces behind advanced AI development, yet their strategic competition is one of the main reasons common rules are difficult to establish. The UN can give 193 countries a place in the conversation, but decisions by the United States and China — and by the relatively small group of companies building frontier models — could determine whether international safeguards have practical force.
That contradiction will be unusually visible during this week’s General Assembly. Xi, who last addressed the gathering by video in 2021, is skipping New York’s high-level debate and is instead due to meet Trump in Washington on Thursday. Yet Chinese Vice Premier He Lifeng and U.S. Treasury Secretary Scott Bessent began talks in New York on Sunday covering AI alongside trade and critical minerals, specifically including possible safeguards for powerful models.
The juxtaposition matters. China may be missing from the UN stage at presidential level, but it is impossible to remove China from the substance of the AI debate taking place around it. Beijing’s technological rise influences Washington’s calculations on regulation, semiconductor controls, open-weight models and the pace at which American companies develop new systems, while the prospect of Chinese advances has become a central argument used in the United States against unilateral restrictions on American AI.
Guterres is approaching the same competition from almost the opposite direction. He has warned that countries operating at the technological frontier need common guardrails to prevent a competitive “race to the bottom” that could eventually produce a global disaster. Trump has publicly dismissed recent warnings about AI risk as a “hoax” and argued that casting doubt on rapid development could benefit China, placing the UN chief’s coordination-first approach against a U.S. administration that sees technological leadership partly through the lens of strategic competition.
The disagreement has become more urgent because it coincides with an extraordinary shift inside the AI industry itself. Leaders and researchers associated with Anthropic, OpenAI, xAI and other frontier laboratories have recently warned that AI could be approaching the ability to contribute increasingly to its own development, reviving concern over so-called recursive self-improvement. Those warnings do not establish that machines capable of escaping human control are imminent, and researchers remain divided over probabilities and timelines, but they have moved the argument over coordination from speculative policy circles into the executive suites of companies actually developing the systems.
That creates a problem resembling an international arms-control dilemma even though AI is not itself a weapon. A company worried about safety may still fear that slowing development will surrender commercial advantage to a competitor; Washington may fear that imposing restrictions on American laboratories gives Beijing additional time to close the technological gap; and China has its own incentives not to accept constraints that could preserve an American lead. In each case, the actor may have reasons to support safety collectively while resisting restraint individually.
The new U.S.-China talks are important precisely because they suggest that this logic is beginning to produce at least some common ground. Bessent said Washington was willing to discuss avoiding shared risks and preventing the two countries’ AI systems from splitting completely apart, and he has called for U.S.-Chinese guardrails aimed at preventing powerful models from reaching malicious non-state actors. Officials are also examining both open- and closed-weight models, an increasingly important distinction because Chinese open-weight systems are gaining users in the United States partly because they can be cheaper and more adaptable than proprietary alternatives.
No comprehensive U.S.-China AI agreement has been announced, however, and expectations surrounding the talks remain limited. The discussions are taking place alongside disputes over tariffs, rare-earth minerals and trade commitments, and analysts cited by Reuters said incremental agreements were more plausible than a major breakthrough. The existence of AI talks therefore establishes that Washington and Beijing recognize some shared risks; it does not establish that they have agreed on how frontier AI should be governed.
That distinction is critical for the United Nations because bilateral U.S.-China coordination and global AI governance are not the same thing.
The UN’s argument is that decisions affecting every society cannot ultimately be negotiated only between the countries with the most computing power. General Assembly President Khalilur Rahman warned this month that advanced AI is effectively managed by a small number of major technology companies concentrated in only a handful of states. Citing UNCTAD data, he said 118 countries have yet to participate in major AI-governance discussions and fewer than one-third of developing countries have national AI strategies.
For those countries, the AI question is different from the strategic debate consuming Washington and Beijing. Many governments are concerned not only about hypothetical future systems escaping human control, but about whether they will have access to computing capacity, whether their workers and industries can compete, how AI will affect education and government services, and whether rules established by technological powers will simply be exported to the rest of the world.
This is where UNGA81 gives the AI debate a genuinely global dimension. The UN General Assembly has already established the Global Dialogue on AI Governance, an international platform designed to bring all 193 member states together with companies, researchers, civil society and the technical community. The inaugural dialogue in Geneva on July 6-7 attracted more than 3,000 participants, delegations from 163 countries, six heads of state and government, more than 60 ministers and over 1,000 stakeholder organisations.
The mechanism emerged from the Global Digital Compact adopted in 2024 and was formally established by General Assembly Resolution 79/325 in August 2025. It is intended to develop common approaches and give countries previously excluded from influential AI discussions a place at the table. Its second annual session is scheduled for New York on May 3-4, 2027, meaning the current General Assembly sits between the first global dialogue and the next phase of an institutional process that is still being constructed.
But representation is not the same as authority. The UN dialogue provides a forum for cooperation and common understanding; it does not control the laboratories training advanced models, the semiconductor supply chains that provide their computing power or the governments imposing national security restrictions on technology. The gap between the UN’s universal membership and the concentration of actual AI capability is therefore emerging as one of the defining tensions of international technology governance.
China sits directly at the centre of that tension. The Trump administration regards American AI leadership as strategically important and has repeatedly connected technological policy with competition against Beijing. At the same time, Washington now appears to accept that certain dangers cannot be addressed entirely through technological separation: its talks with China include possible safeguards for open- and closed-weight models and ways to prevent powerful systems from being exploited by malicious actors.
That makes the China question more complicated than whether Xi attends UNGA81. His absence deprives the General Assembly of a presidential-level Chinese intervention in one of its most important emerging debates, but Beijing is not absent from AI diplomacy. Chinese officials are negotiating in New York, Xi is preparing for direct talks with Trump, and almost every American discussion about how quickly to regulate or develop frontier AI now contains an implicit calculation about what China might do next.
The distinction also exposes a challenge for Guterres. If Washington and Beijing reach bilateral understandings on the most dangerous capabilities, such an agreement could provide an important foundation for wider international coordination. But if the two powers establish rules primarily around their own strategic interests, countries outside the frontier-AI club could remain rule-takers rather than rule-makers — exactly the imbalance the UN’s Global Dialogue was created to address.
The developing world therefore has a substantial stake in what happens next. AI systems trained largely in the United States, China and a small number of other technology centres can influence economies and societies far beyond those countries. If advanced models reshape labour markets, education, public services, information ecosystems and cybersecurity, states without significant AI infrastructure will still experience the consequences even though they had little influence over the models’ design or the standards governing them.
The UN’s relevance may consequently depend on whether it can connect two conversations that are currently moving on different tracks. One is the strategic conversation between Washington and Beijing about safety, technological leadership, open models, chips and national security. The other is the global conversation among 193 countries about access, development, representation, human rights and who should participate in writing rules for a technology with worldwide consequences.
UNGA81 arrives just as those conversations begin to overlap.
The possible Security Council discussion on AI during high-level week adds another layer. The council contains the United States and China as permanent veto-wielding members, along with Russia, Britain and France, making it one of the few international bodies where the major geopolitical powers could directly address AI as a security issue. But the same rivalries that have paralysed council action over conflicts such as Ukraine and parts of the Middle East could also complicate attempts to establish meaningful technological restraints.
There is also a danger in allowing the current debate to become dominated by the most extreme scenarios. Concerns about recursive self-improvement and loss of human control deserve scrutiny because researchers working on frontier systems themselves are raising them, but those remain uncertain future risks rather than established outcomes. AI already poses more immediate governance questions involving cyber misuse, automated fraud, surveillance, information manipulation and unequal access to technology, while countries with limited AI capacity face the possibility of becoming increasingly dependent on systems they neither own nor substantially govern.
The UN therefore faces a narrow path. Moving too slowly risks allowing technology and national policies to outrun international coordination; trying to impose sweeping rules without agreement from the countries and companies controlling frontier development risks producing declarations with little practical effect. A credible global system would need to reconcile safety with innovation, national security with international cooperation, and the interests of technological powers with those of countries that currently have little influence over AI development.
China makes that balancing act unavoidable.
Trump’s argument that excessive restraint could help Beijing captures the strategic logic driving the American race. Guterres’ warning about a race to the bottom captures the collective danger if every major actor follows that logic indefinitely. The fact that U.S. and Chinese officials are now discussing AI guardrails in New York suggests the two positions are not necessarily irreconcilable, but it remains far too early to say whether those talks can produce durable common rules.
That is why the most revealing AI story at UNGA81 may ultimately unfold between New York and Washington rather than solely at the General Assembly podium. The UN is trying to create a system in which every country can participate in governing artificial intelligence; the United States and China are simultaneously determining whether the two most important AI powers can agree on even a minimum set of safeguards between themselves.
If Washington and Beijing cannot establish common ground, global AI rules will be extraordinarily difficult to enforce. If they can, the next question will be equally consequential: will the rest of the world be allowed to help shape those rules, or merely be asked to live under them?