The Resumption of U.S.-China AI Safety Dialogue and the Governance Gap: A Dual-Track Structure and Middle Power Responses

Executive Summary

Following the Trump-Xi Jinping summit, the United States and China have restarted their intergovernmental dialogue channel on AI. However, the substantive gap between them is not narrowing, as the White House's cooperation track operates separately from the pressure track led by regulatory agencies like the FCC and by Congress. The commercial incentives of the private sector are moving independently of the government's narrative, as seen in cases where U.S. companies seek to host Chinese AI models on a revenue-sharing basis. China also exhibits a dual approach, maintaining its critical tone toward the U.S. in state media regardless of the agreement to hold dialogue. While the U.S. has an overwhelming lead in infrastructure and investment, this is most likely to result in selective controls limited to security-related items rather than a full decoupling. The most probable path forward is the institutionalization of this dual-track approach, where the dialogue channel remains limited to confidence-building information exchange. Middle powers like South Korea will need to pursue a parallel response: managing risks in security-related areas while maintaining existing relationships in the general commercial sphere. In this process, establishing a system to constantly monitor the individual actions of U.S. regulatory agencies emerges as a practical priority.

I. Situational Analysis

U.S.-China AI Safety Dialogue Resumes, but Fundamental Gaps Remain

1. Background and Developments

At the Trump-Xi Jinping summit in Beijing last May, the two leaders agreed to launch an intergovernmental dialogue on AI [1]. This agreement was interpreted as a signal that both countries would operate practical cooperation channels in parallel with their AI competition [1]. It was also confirmed that ahead of the summit, Senator Steve Daines of Montana, a special envoy for President Trump, visited Chinese AI and space companies in Hangzhou and Guangzhou [10]. This was clearly a preparatory move to ensure the summit's success [10].

At the same time, within the United States, technological containment measures against China were accumulating on a separate track at the regulatory agency level. The Federal Communications Commission (FCC) was building its own set of regulations targeting China, including a review of a ban on optical transceivers, which constituted bureaucratic pressure distinct from the White House's diplomatic track [3]. EAI had previously pointed out that the flashpoint of this conflict was not the White House but the regulatory agencies, and that localized friction was likely to continue regardless of the summit's outcome [3]. This dual-track structure has persisted even after the agreement on AI dialogue.

2. Current Situation

In the six weeks following the summit, G7 leaders brought executives from frontier AI companies to the table in Evian, France, and related discussions continued at the United Nations [2]. The Brookings Institution's assessment of this "summer summit season" is that the governance gap between the U.S. and China is widening, despite the agreement to cooperate [2]. The U.S.-China Forum on Cooperation in AI (FCAI) dialogue held on July 23 was conducted under the Chatham House Rule, which indicates that both sides prefer private exploration over public statements [2].

During the same period, the state-run Global Times has been running stories highlighting the gap between Washington's narrative of controlling Chinese AI and the reality on the ground in the industry. It cited a Reuters report that U.S. cloud companies like Microsoft, Amazon, and Google are in negotiations to host Kimi K3, a model from China's Moonshot AI, on a revenue-sharing basis, assessing this as revealing a fissure between the U.S. narrative of control and industrial reality [14]. The Global Times also commented that the pressure on U.S. companies to sever ties with Chinese AI labs is itself evidence of American anxiety [16]. It also cited a New York Post report that four U.S. firms supplying data to Anthropic, Google, Meta, OpenAI, and the Department of Defense have been secretly continuing their cooperation with Chinese AI labs [16].

Conversely, opposite concerns are being raised in the United States. The South China Morning Post reported that at the same time that calls were being made for the Trump-Xi AI safety dialogue, Chinese state media intensified their criticism of U.S. AI governance [4]. Dean Ball, Head of Strategic Futures at OpenAI, had previously viewed U.S.-China AI safety cooperation as desirable but unlikely; however, he recently assessed that the situation has changed in recent months and a window for cooperation has opened [4]. He warned, however, that this window is not permanent and could widen and then close within the next few months [4]. Meanwhile, at the G20 Technology Ministers' Meeting in North Carolina in early September, the United States demanded a "hands-off" approach to AI regulation [12]. This position aligns with the interests of most major U.S. AI companies and is read as an attempt to push for minimal regulation or the establishment of U.S.-led rules [12].

3. Key Actors and Positions

Positions within the U.S. administration are not monolithic. While the White House agreed to open a dialogue channel as part of its summit diplomacy, regulatory agencies like the FCC have independently accumulated pressure tactics against China [3]. A draft executive order within the Trump administration discussing the establishment of a public-private self-regulatory body to oversee AI model releases has also stalled [15]. This suggests a lack of consensus on the direction of regulation even within the United States.

China has adopted a strategy of directly refuting the U.S. narrative on AI control through its state-run media [14][16]. While Beijing has so far refrained from direct retaliation against U.S. regulatory measures, there are assessments that this patience is being tested [3]. At the same time, the rising competitiveness of Chinese AI models since DeepSeek is having real-world effects, such as putting downward price pressure on Silicon Valley [17][7].

Silicon Valley companies, perceiving stronger regulations on China as a constraint on their business opportunities, are working with the U.S. government at venues like the G20 to push for a minimal regulation approach [12]. In contrast, suppliers providing data to companies like Anthropic are maintaining practical cooperation with Chinese labs despite political pressure [16].

4. Key Issues

The first key issue is the disconnect between top-level agreements and the actions of bureaucratic organizations. The White House's agreement to launch an AI dialogue is proceeding simultaneously with independent pressure from regulatory agencies like the FCC, creating a situation where localized friction persists regardless of the summit's outcome [3][15]. The second issue is the fundamental difference in governance approaches. The U.S. is demanding minimal regulation at the G20 [12], while China is pursuing a state-led industrial policy in parallel with a narrative competition via state media [14][16]. As Brookings pointed out, this gap has widened rather than narrowed over the summer summit season [2]. The third issue is the time-limited nature of the window for cooperation. As the OpenAI official assessed, the current conditions for cooperation are favorable but not permanent, and the scope of information sharing and the potential for finding common ground remain uncertain [4][17]. This dual structure has significant implications for middle powers, including South Korea. As U.S. pressure to choose sides on AI and the infrastructure gap both grow, a dual-track response that distinguishes between strategic technologies and commercial domains, rather than prematurely aligning with a specific bloc, emerges as a realistic option [9].

II. In-Depth Analysis

U.S.-China AI Safety Dialogue Resumes, but Fundamental Gaps Remain

3. In-Depth Analysis

Root Cause: A Structure Where Cooperation and Competition Are Inseparable

The fundamental reason for the slow progress in the U.S.-China AI dialogue, despite its launch through a top-level agreement, lies in the fact that the dialogue and competition channels are operated by different groups of actors. The White House opened a window for cooperation on the diplomatic track [1]. However, regulatory agencies like the FCC have been accumulating technological pressure on China based on their own bureaucratic judgments [3]. EAI has previously noted that the flashpoint of this conflict was not the White House but these regulatory agencies, assessing that "localized friction was likely to continue regardless of whether the talks succeeded" [3]. This structure is repeating itself after the AI dialogue agreement. The agreement between leaders does not control the actions of the entire bureaucracy.

The actions of the private sector also diverge from the government's narrative. This is demonstrated by the fact that Microsoft, Amazon, and Google are in negotiations to host Kimi K3, a model from China's Moonshot AI, on a revenue-sharing basis [14]. Reports that U.S. firms supplying data to Anthropic, Google, Meta, OpenAI, and the Department of Defense are secretly continuing their cooperation with Chinese AI labs tell a similar story [16]. Washington's discourse of control is in direct conflict with the profit motives of Silicon Valley and Wall Street. The Global Times has defined this as a "fissure between Washington's narrative and industrial reality" [14]. This fissure creates a paradox where the more the U.S. government tightens control, the greater the incentive for companies to find workarounds.

A similar duality is found on the Chinese side. Even immediately after the agreement on intergovernmental dialogue, state media intensified their criticism of U.S. AI governance [4]. This is an approach of keeping cooperation channels open while maintaining a competitive frame in external propaganda. This mirrors the U.S. dual-track approach, in that the narrative for domestic politics and the practicalities of foreign negotiations operate separately.

Structural Context: Asymmetry on Three Levels

The structure underlying this issue can be analyzed on three levels.

The first is the asymmetry at the infrastructure level. The United States operates 5,427 data centers, more than 10 times the number of any other country [9]. This gap stems from the fact that from 2013 to 2023, cumulative private AI investment in the U.S. reached $335.2 billion, while China's was significantly lower [9]. An EAI analysis judged the probability of this infrastructure gap leading to "selective decoupling limited to security-related items rather than a full hardening of blocs" to be the highest, at 55-60% [9]. This implies that because its infrastructure advantage is not absolute, the U.S. is also opting for selective controls over a full blockade.

The second is the asymmetry in governance philosophy. At the G20 Technology Ministers' Meeting, the U.S. strongly advocated for a hands-off principle regarding AI regulation [12]. This aligns perfectly with the interests of Big Tech, the majority of which are American companies [12]. In contrast, China is attempting to build its own bloc through its own frameworks, such as the World AI Cooperation Organization (WAICO) [9]. However, this framework is still incomplete, and Europe also remains skeptical, making it difficult for the Chinese alternative to gain immediate traction [9]. Ultimately, the U.S. is aiming for minimal regulation and market leadership, while China is pursuing the creation of a state-led alternative order, meaning the very axes of normative competition are being formed differently.

The third is domestic political constraints. Within the U.S., a plan to establish a public-private self-regulatory body to oversee AI model releases remains stalled at the draft executive order stage [15]. The enforcement of export controls is also inconsistent. A contradiction exists where measures are implemented to prevent foreigners, including Anthropic employees, from using certain models [8], while cooperation with China by data suppliers is effectively overlooked [16]. This suggests that positions have not been reconciled between proponents of industrial promotion and security hardliners within the Trump administration.

Historical Precedents: Similarities and Differences with Cold War Arms Control Talks

The structure of the U.S.-China AI dialogue is partially similar to the initial phase of the U.S.-Soviet Strategic Arms Limitation Talks (SALT) during the Cold War. A common feature is the maintenance of minimal communication channels to avoid catastrophic outcomes even amid extreme mutual distrust. Foreign Policy justified the current dialogue with the logic that "AI safety can be built despite mutual distrust" [1], which is in line with the logic used during the SALT negotiations. This is a method of building procedural trust to prevent catastrophic miscalculations without halting competition.

However, there are also crucial differences. SALT dealt with verifiable quantities of weapon systems, whereas with AI, the boundaries of model capabilities are ambiguous and verification tools themselves are lacking. In a survey of 350 experts conducted by the Council on Foreign Relations (CFR), expert opinions diverged on almost every point regarding how the distribution of frontier AI capabilities and governance structures would turn out by 2035 [11]. However, there was a consensus that "no one is prepared to govern upcoming AI capabilities" [11]. This contrasts with the SALT era, when concepts of nuclear arms control were relatively well-established. Given that the very object of verification is indeterminate, it is highly likely that the AI safety dialogue will remain at a much lower level of substantive binding power than SALT.

Another precedent is the U.S.-Japan semiconductor friction of the 1980s and 1990s. At that time, the U.S. also used a combination of export controls on Japan and pressure for market access to maintain its technological superiority. However, Japan was an ally, and their security interests were not in fundamental conflict. In contrast, the current U.S.-China relationship is qualitatively different in that technological competition is directly linked to military and security competition. This is also the background for why semiconductors and AI are treated as key controlled items within the broader domain of U.S.-China strategic competition.

Key Variables

The variables that will shape the development of this issue can be narrowed down to three.

First is how long the window for cooperation remains open. Dean Ball warned that while the scope for cooperation could widen in the coming months, this window is not permanent [4]. If hardline sentiment against China re-emerges in the U.S. domestic political calendar, particularly during the 2026 midterm election cycle, the momentum for cooperation could rapidly diminish.

Second is the degree to which independent actions by regulatory agencies erode the diplomatic agreement. If bureaucratic measures against China, such as those from the FCC, continue to accumulate [3], the effectiveness of the top-level agreement could be hollowed out. The key question is where Beijing's threshold for retaliatory measures, which it has so far restrained, lies [3].

Third is the speed of commercial penetration by Chinese AI models. Whether the negotiations for U.S. cloud hosting of Kimi K3 succeed [14] and whether low-cost, high-efficiency models continue to emerge after DeepSeek [6] will determine how much further the U.S. narrative of control diverges from industrial reality. If the interests of the private sector continue to outpace the government's control narrative, the possibility that the U.S. export control regime itself could be adjusted due to internal pressure cannot be ruled out.

These three variables are not independent. They form a cyclical structure where wavering trust on the diplomatic track accelerates hardening on the bureaucratic track, which in turn increases the incentive for the private sector to find workarounds. Middle powers, including South Korea, need to distinguish at which phase of this cycle signals of tighter regulation appear and at which phase practical relaxation occurs, and coordinate the timing of their responses accordingly.