U.S.-China AI Confidence-Building: A Path Through Strategic Limits

In the months before Chinese General Secretary Xi Jinping’s September 2026 visit to the White House, where artificial intelligence was among the subjects to be discussed, U.S. Treasury Secretary Scott Bessent met with Chinese Vice Premier He Lifeng at JPMorgan Chase’s New York City offices. American officials proposed a United States-China dialogue that would serve as a mechanism for notifying each other about artificial intelligence-related national security incidents. Bessent said the objective was to move “from opaque to more transparency between the No. 1 and the No. 2 AI powers in the world,” describing that shift as important for cross-border activity.

The Chinese government’s readout of the subsequent meeting between Xi and U.S. President Donald Trump said the two countries “can continue their dialogue on AI, exchange views on its risks and benefits, and jointly prevent the misuse and abuse of AI,” and stated that “AI must be kept under human control.” The statement echoed the countries’ 2024 agreement on human control over nuclear weapons, in which the United States and China recognized the need to maintain human control over the decision to use nuclear weapons. The U.S. readout of the 2026 meeting said the two countries agreed to establish the “U.S.-China Super Intelligence (SI) Dialogue” to exchange views on risks and benefits related to super intelligence, with the next exchange expected by November 2026. The two sides also agreed to set up “a bilateral communication channel for SI incidents.”

The proposed dialogue and communication channels would not restrain AI innovation in the United States or China, nor would they change the underlying balance of AI power between them. They also do not include enforcement or verification mechanisms, meaning they are not a form of AI arms control. Instead, they are better understood as confidence-building measures: voluntary, non-binding steps intended to reduce opacity and facilitate communication between rivals, potentially lowering tensions and the risk of inadvertent or accidental escalation.

Evaluating such efforts through the lens of Cold War confidence-building measures offers some insight. The 1963 U.S.-Soviet hotline emerged after the Cuban Missile Crisis, and later arrangements helped manage nuclear risks. The 1972 Incidents at Sea Agreement and subsequent confidence-building documents were also part of efforts to reduce miscalculation. But AI has features that may limit how far historical lessons can be applied.

For military applications of AI, the United States and China perceive a central tradeoff. On one hand, they remain locked in a broader strategic rivalry, of which AI competition — or even an “arms race” — is an integral part. Trump said in the context of U.S.-Chinese relations that “whoever wins AI wins,” while China’s five-year plan called for seizing “the commanding heights of science and technology development.” AI could provide military advantages in intelligence, targeting, command and control, autonomous systems, and decision-support.

On the other hand, both countries share concerns about AI’s national security risks, particularly regarding AI in decisions about the use of nuclear weapons. China has said AI may contribute to “the deepening of strategic miscalculation, undermining of strategic mutual trust, escalation of conflicts, and damaging global strategic balance and stability.” Even without full delegation of launch decisions to machines, AI integration into nuclear command, control, early warning, and launch decisions creates new uncertainty and escalation risks.

For example, AI tools used to support decision-making around nuclear use or to offer courses of action may bias decisions toward more escalatory options. Research has shown that some large language models, including China’s DeepSeek model, were more inclined to recommend nuclear options during simulated crises. AI-enabled intelligence and early warning could produce skewed or hallucinated analysis. A false AI-generated intelligence report earlier this year nearly “started a war” between the United States and China during a U.S. military operation against Iran. Neither side wants an AI-triggered war, especially a nuclear one.

One way to manage these risks is through arms control agreements. There are growing calls among experts and practitioners for AI arms control, especially between the United States and China. But this is problematic, given global arms control trends and the nature of AI. With the expiration of New START and the emergence of a new tripolar nuclear age, there appears to be little appetite among great powers for new arms control agreements. As a dual-use, general-purpose technology, AI presents challenges for detection and disclosure. Deep uncertainty about future military implications and the rapid pace of innovation make powers reluctant to accept constraints.

But arms control is not the only option. Historically, confidence-building measures have emerged even between rivals when states perceived shared interests in communication and transparency. During the Cold War, the hotline and related nuclear risk reduction efforts followed the Cuban Missile Crisis. The 1972 Incidents at Sea Agreement and later confidence-building documents were also examples. Confidence-building measures are not a panacea, and they often lack verification and enforcement. Russia’s 2023 suspension of information-sharing under the Vienna Document after its full-scale invasion of Ukraine illustrated that they can be abandoned. Still, well-scoped efforts to facilitate communication are often better than the status quo.

Four factors unique to AI may limit how well Cold War lessons can be generalized. The first is that an AI crisis comparable to the Cuban Missile Crisis has not yet occurred. Without a vivid, shared sense of danger, it is unclear that both sides fully agree on the specific nuclear risks posed by AI beyond broad support for human control.

The second is little agreement between the United States and China about key terms and concepts related to AI escalation. Both have articulated different definitions of lethal autonomous weapons. China applies a high threshold, defining “unacceptable” lethal autonomous weapons as those that act across the kill chain with an “absence of human intervention and control during the entire process of executing a task.” The United States defines such a system as one that, once activated, can select and engage targets without further intervention by an operator. The two sides also differ on the role of human beings in decisions about the use of force.

The third issue is that, unlike nuclear weapons, AI is not a single weapon system. It is a general-purpose technology with civilian and varied military applications, so even narrowly scoped AI-nuclear stability efforts must address broad ground. Many AI applications beyond the specific concern about human control over the decision to launch nuclear weapons may also contribute to nuclear instability, while many of these capabilities remain important to military competition, limiting prospects for voluntary restraints.

Finally, the private sector is an integral actor in any AI confidence-building effort. AI power is concentrated in a handful of firms that control access to key AI inputs, and militaries are dependent on them. Yet those actors may have interests not aligned with government interests, especially around intrusive measures or information sharing.

The article concludes that policymakers should prioritize four specific lines of effort. First, broader sensitivities in the U.S.-Chinese relationship about AI may create space for progress on narrower, mutually shared concerns about nuclear escalation, provided more contentious issues are treated as off the table. The U.S.-Chinese AI dialogue should focus on building progress within the one clear area where both parties have already reached some agreement: human control over decisions to use nuclear weapons. The challenge will be avoiding U.S. efforts to “pace the frontier” that could reinforce Chinese concerns about U.S. restrictions on AI development.

Second, policymakers should build on the existing foundation of the 2024 agreement about human control over nuclear weapons rather than immediately expanding the scope of the dialogue. The initial rounds should focus on improving clarity and consistency of core concepts and definitions. Parties should work toward a confidence-building framework that clearly defines shared understandings about what constitutes “autonomous” capabilities and what human control means in practice, including for early warning and nuclear command, control, and communications.

Third, policymakers should focus on observable behavior. The emphasis should be less on restricting technology development itself and more on developing observable indicators about dangerous use of AI in nuclear systems that both sides agree would be dangerous. Both sides could benefit from shared understandings about escalation pathways and off-ramps, and from exchanging information about how each side would perceive different escalation scenarios.

Finally, how the U.S. government regulates frontier AI labs and firms controlling AI infrastructure inputs can be a form of confidence-building. Confidence-building measures do not need to be official agreements. Unilateral, voluntary declarations of restraint can also build confidence. How the government collaborates with and controls private actors around incident reporting or safeguards can buttress credibility, while failure to do so may undermine bilateral efforts.

Looking ahead, there may be opportunities to explore cross-domain confidence-building efforts in areas such as AI, cyberspace, outer space, and nuclear systems. The focus in the immediate term should be on the most exigent risks.

Source link