The most revealing thing about this week’s AI debate is not US-China disagreement. Beneath it, both powers are beginning to recognize the same danger.
The US and China are racing for AI dominance because each believes it will shape military power, economic strength and technological supremacy. But the faster the race advances, the harder it becomes for either side to control what it is building. That paradox frames Trump’s meeting with Xi starting today.
The New York Times captures the dilemma: the US leads in advanced chips and frontier models; China has strengths in AI talent, electricity and open-source systems. Neither can comfortably slow down—which may be exactly why both need to.
Leading US AI executives now warn that increasingly autonomous systems are already behaving in ways their creators did not anticipate. At the UN yesterday, OpenAI’s Sam Altman and Anthropic’s Dario Amodei appeared before the Security Council as governments confront AI’s security implications. The UN Sec-General urged leading countries to share safety risks, cooperate on testing and develop common safeguards.
This is not abstract. A reported incident from earlier this year involved a US military intelligence briefing generated with a chatbot that misidentified cargo on a Chinese ship as nuclear-related, triggering preparations to intercept it before officials found the error. An AI hallucination can become a geopolitical event when it enters a military decision chain.
That is what makes the US-China AI race different from the nuclear arms race. Nuclear weapons are hard to build, physically identifiable and controlled through command chains. AI is software. Once a powerful model is released—especially open source—it can be copied, modified and redistributed across borders almost instantly. Its safeguards can be altered; its capabilities repurposed. You cannot impose a national border on code.
Washington frames the competition starkly: whoever wins AI wins. Beijing resists anything that might constrain China’s rise. The result is familiar logic applied to an unfamiliar technology: if we slow down, they win. But if the technology creates risks neither side can reliably control, moving faster does not produce security. It produces mutual exposure.
The NYT report makes the dilemma clear: US companies keep tight controls around leading models, while major Chinese companies increasingly release powerful systems as open source. Open source accelerates innovation, but it also makes capabilities easier to modify and misuse. Unilateral regulation cannot solve this. If Washington imposes strict controls while Chinese open-source models advance, US developers may migrate or fall behind. If China slows while US companies accelerate, Beijing fears losing strategic advantage. If neither slows, every safety measure becomes subordinate to the next breakthrough.
One encouraging development: Washington and Beijing have begun discussing an AI safety notification mechanism to alert each other when an AI issue reaches national-security level. US Treasury Secretary Bessent described the proposed US-China AI dialogue as a move from opacity toward transparency. That sounds modest, but modesty may be what is needed.
The Cold War nuclear hotline did not end US-Soviet rivalry. It created a basic channel for communication when events outpaced diplomacy. AI needs the equivalent. The first objective should not be a grand treaty regulating every algorithm. It should be layered. Major AI laboratories should agree on common thresholds for catastrophic-risk testing, with independent evaluation before deploying the most powerful systems. Washington and Beijing should establish a permanent AI incident channel: if an autonomous system escapes containment, compromises critical infrastructure, conducts significant cyber operations or produces a potentially catastrophic intelligence assessment, the other side should know. The UN should build a global reporting and standards framework—not an AI police force—setting common definitions, testing principles and transparency requirements.
The temptation is to treat this week’s summit as another contest over who is ahead. The more important question is whether the two leading AI powers can recognize that strategic competition and strategic restraint are not opposites. They already cooperate, however reluctantly, when the alternative is catastrophic miscalculation. The emerging AI hotline is a small acknowledgment of that.
This week exposed a contradiction. Those closest to frontier AI warn that the technology may outrun our ability to control it. Governments remain convinced they can control the geopolitical race around it. Both cannot remain true indefinitely. AI regulation should no longer be framed as innovation versus fear, or US leadership versus Chinese leadership. The real choice is whether humanity can build common rules for a technology that does not recognize national borders. The question is whether we can build the guardrails before the race outruns them—and all of us.
