The two countries that host nearly all frontier AI labs and produce most of the models people actually use have each, for different reasons, refused to slow down. The US frames AI safety concerns as conspiracy and globalist overreach; China frames them as a pretext for technological containment. The outcome is the same: acceleration without coordinated guardrails. As Trump and Xi prepared to meet in Washington for their third summit this year, the best-case scenario floated by analysts was a dialogue mechanism and a notification system for national security threats — not the binding capability limits or moratorium that the most alarmed researchers say is needed. Trump told the UN General Assembly that Washington would oppose any "globalist scheme" to control AI. China's third security framework iteration this month focused on cybersecurity, regime stability, and countering "hostile forces" using AI to "fabricate political rumours" — not recursive self-improvement or loss-of-control scenarios. The structural problem is a perception mismatch that makes cooperation nearly impossible. American AI safety discourse centers on catastrophic or existential risk from highly capable systems escaping human control. Chinese security discourse centers on the geopolitical danger that the US gains a decisive AI advantage and weaponizes it. As Henry Gao of Singapore Management University put it: "That creates pressure to accelerate rather than slow development." Each side's threat model makes the other side's proposed solution look like a trap. Sun Chenghao of Tsinghua University named the core obstacle plainly: China increasingly recognizes loss-of-control and misuse risks, but cannot separate "AI safety" from export controls and chip restrictions. "If every safety discussion is linked to export controls or slowing China's development, Beijing will naturally question the purpose of cooperation." The US, meanwhile, treats verification as functionally impossible. As Alvin Wang Graylin of the Asia Society noted: "Verification is hard for software you cannot see." The AI industry itself is split. Anthropic's Dario Amodei and OpenAI's Sam Altman call for a slowdown and greater regulation. David Krueger of the University of Montreal demands "an immediate, indefinite international moratorium on frontier AI development." Otto Barten of the Existential Risk Observatory insists AI must never reach recursive self-improvement capability. On the other side, Nvidia's Jensen Huang and Meta's Mark Zuckerberg play down external oversight and stress transformative benefits. No consensus exists on whether government-led supervision is even needed. The recursive self-improvement threshold — where AI models learn to improve themselves without human supervision — is identified by alarmed researchers as the critical tipping point. Amodei acknowledged that any US-China agreement to limit development rates is unlikely given that a rule-breaking party could shift the global balance of power. This is the classic arms-race trap: the potential payoff for defection is so large that cooperation collapses even when both parties would benefit from it. Proposed solutions remain thin. Wang Graylin suggested agreement on "unacceptable capabilities" and common testing standards administered by an independent international body funded by multiple countries. Sun recommended the US draw a clearer line between safety policies and technological-advantage policies. Rui Ma argued the focus should be on evolving safeguards rather than pausing development. Jiang Tianjiao of Fudan University expressed hope the superpowers could overcome rivalry to "prevent AI from spiralling out of control." None of these proposals have binding force, enforcement mechanisms, or any indication either government is interested.
