A new public call for U.S.-China cooperation on AI safety is testing whether leaders can turn a shared concern about advanced systems into a workable diplomatic channel. Dean Ball, OpenAI’s head of strategic futures and a former White House adviser, said a window for collaboration may be opening ahead of the planned September 24 meeting between U.S. President Donald Trump and Chinese leader Xi Jinping. His observation came as Chinese state media intensified criticism of American frontier-AI governance. The juxtaposition, reported by SCMP, captures the core problem: both governments may have reasons to talk about guardrails, but they disagree sharply about who should set them and how they should work.
The current development should not be mistaken for a signed bilateral agreement. Ball’s comments are a professional assessment, not a government commitment. Treasury Secretary Scott Bessent indicated to Reuters that officials planned to continue discussions on AI guardrails, but neither side has announced a detailed negotiating mandate, timetable, or common framework.
Still, the prospect is worth watching because AI safety is one of the few technology topics where competition and mutual risk can coexist. Systems developed in one country can be deployed, copied, evaluated, or misused across borders. A failure of control, a major cyber incident, or an escalation involving autonomous systems would not necessarily stay within national boundaries.
A Narrow Diplomatic Opening
Ball said he had long regarded U.S.-China AI safety collaboration as desirable but unlikely. He now sees a period in which the opportunity could widen, while warning that it might not last. The change in tone reflects a wider recognition that leading AI systems are moving from a research topic into infrastructure, commercial services, and political competition.
For the United States, a safety dialogue could offer a way to discuss incident reporting, evaluation practices, cyber risk, or controls on particular high-risk uses without giving up its broader technology strategy. For China, it could provide a venue to argue that international rules should not be written solely by U.S. companies and policymakers. Those are overlapping interests, not identical ones.
The planned Trump-Xi meeting gives the subject a possible political venue. But diplomacy around AI is difficult precisely because technical issues are tied to trade restrictions, semiconductor controls, data access, military concerns, and corporate competition. A conversation about model safeguards can easily become a dispute about who is allowed to train the most capable systems.
EastFrontier has followed this tension before. In July, it examined why U.S.-China AI cooperation might succeed where trade talks have failed. The fresh September discussion does not erase the obstacles identified then. It shows that policymakers and industry figures continue to look for a limited issue on which dialogue remains possible.
China’s Critique of U.S. AI Governance
The diplomatic opening is accompanied by sharp rhetoric. Yuyuan Tantian, a social-media account affiliated with China Central Television, published a critique of Anthropic and U.S. AI governance ahead of the planned talks. According to SCMP, the commentary accused U.S. firms of using safety standards and lobbying to shape rules in ways that disadvantage international competitors.
Those are accusations from a state-media-linked account, not findings established by an independent investigation. They nevertheless matter because they illustrate the argument China is likely to make in international AI debates. Beijing has often supported global governance in principle while arguing that Western governments and companies should not use safety, security, or export-control rules to preserve a technology advantage.
The criticism of Anthropic also brings private companies into a government-level dispute. Frontier-model firms are not merely product suppliers. Their safety policies, access restrictions, public warnings, and advocacy can become evidence in a larger argument about fairness, control, and the global distribution of AI capability.
This dynamic has been visible in past disputes over model access. EastFrontier reported in July that China issued an official security alert over Anthropic’s Claude Code. The new state-media commentary fits into that broader pattern, though it should not be treated as a formal bilateral negotiating position unless officials say so.
From Rhetoric to Rules
If U.S. and Chinese officials do pursue a safety conversation, the most realistic early outcomes may be modest. They could establish a regular channel, exchange views on how to manage catastrophic risks, or discuss principles for avoiding certain military and cyber escalations. Such steps would not resolve export controls, platform access, or competing ideas about open models and governance.
A serious dialogue would also need to distinguish between shared safety concerns and politically contested definitions of risk. One side may emphasize frontier-model misuse, while the other may focus on data sovereignty, national development, or what it sees as discriminatory restrictions. The goal should not be to pretend those differences do not exist. It should be to identify areas where misunderstanding or accident would be costly for both sides.
The timing is favorable in one narrow sense: the September 24 summit concentrates attention and creates a deadline for officials to consider an agenda. It is unfavorable in another: public criticism, trade tensions, and disputes about models and chips can harden positions before the meeting begins.
The latest calls therefore represent possibility rather than progress. Ball believes the window is opening, Bessent says dialogue will continue, and Chinese state media are framing U.S. governance as a source of concern. Whether that combination produces even a limited safety channel will depend on decisions by governments, not on commentary from industry or media figures. The policy test is whether both sides can separate shared catastrophic-risk interests from the many competitive disputes that surround AI.