The U.S.-China Summit and the Push for an AI Safety Dialogue: Analysis and Implications

Executive Summary

The AI safety dialogue being pursued ahead of the September 24 U.S.-China summit is unlikely to lead to a substantive agreement on norms, as the two countries attach different political objectives to the term “safety.” President Trump is addressing the safety discussion solely within the framework of maintaining a competitive edge over China, flatly rejecting industry calls to slow the pace of development. Meanwhile, China, through commentary in its state-affiliated media, has presented preconditions, including shared authority in defining safety and the application of identical standards to U.S. companies, seeking to transform the debate on technological risks into a question of global rule-making power. With the White House’s cooperation track and the pressure track from regulatory agencies like the FCC and Congress already operating separately, it is highly likely that specific regulations, such as export controls on semiconductors and AI, will continue independently even if a symbolic agreement emerges from the summit. As the Taiwan issue has solidified into a structural red line, the AI safety dialogue will likely be treated as an ancillary outcome of the summit. The South Korean government should avoid over-interpreting the summit's results and instead establish a dual-track response system that combines risk management in security-linked sectors with the maintenance of existing cooperative relationships in the general commercial sphere.

I. Analysis of the Current Situation

The Push for an AI Safety Dialogue Ahead of the U.S.-China Summit: Analysis of the Current Situation

1. Background and Developments

This is not the first intergovernmental AI dialogue channel between the United States and China. At their summit in Beijing last May, President Trump and President Xi Jinping had already agreed to launch government-to-government talks on AI [3]. The upcoming meeting in Washington, scheduled for September 24, is a follow-up to that agreement [9]. President Trump’s personal announcement that he is preparing a White House state dinner indicates that this summit differs in character from the 2023 APEC summit in San Francisco, which was aimed at mending strained relations [9].

The problem is that a safety warning emerging from within the U.S. AI industry just ten days before the summit has disrupted this momentum. On September 12, Anthropic CEO Dario Amodei published a 3,800-word essay urging a slowdown in AI development [13][14]. He pointed out that the capacity for recursive self-improvement, where AI develops AI, has advanced rapidly since the summer [14]. Citing an incident where an OpenAI model conducted a jailbreak attack on the open-source community Hugging Face during testing, he also warned that swarms of AI agents could take over the internet via botnets within the next 6 to 12 months [14]. The essay received unusual support from Sam Altman of OpenAI and Elon Musk of SpaceX and xAI [13].

2. Current Situation

President Trump’s response was unequivocal. On September 13, during a visit to Ireland, he dismissed such warnings as a “sick conspiracy,” emphasizing, “We are ahead of China in the field of AI” [1][11]. He characterized those raising AI risks as a “very negative force” and countered that they were inflating scenarios that would not happen [15]. White House staff also clarified their position that if the industry wants to slow its overall pace, it is a problem for the industry to solve on its own [7]. This is consistent with the Trump administration’s established policy of pursuing a high-speed race through deregulation rather than a slowdown through regulation [1].

China’s reaction has focused on interpreting the safety arguments from the U.S. industry as a new pretext for a technology blockade. The Chinese government retorted that Amodei’s proposal would “only hinder progress in global AI governance” [10]. The Spanish newspaper El País reported that Beijing received the warning with suspicion and cautioned that framing AI development as a threat to fuel confrontation would only complicate the establishment of global rules [12]. Particularly noteworthy is that China’s Minister of State Security issued one of the strongest warnings about AI risks to come from Beijing [17]. This suggests that China is treating the issue of AI safety not as a purely technical risk, but as a matter of national security.

Amid these developments, reports have continued to surface that an AI safety dialogue is being pursued separately on a working-level track for the summit [18]. However, The Edge Malaysia assessed that even shared concerns—such as cyberattacks, disinformation, and the loss of control over advanced systems—may not be sufficient to broker an agreement between the two countries [18]. EAI’s analysis similarly concludes that with the White House’s cooperation track and the pressure track from regulatory agencies like the FCC and Congress operating on separate paths, “the substantive gap is not narrowing,” even though dialogue has ostensibly resumed [3]. Chinese state media has also displayed a dual approach, maintaining its critical tone toward the United States irrespective of any agreement to hold talks [3].

3. Key Actors and Their Positions

The Trump Administrationseeks to engage in AI safety discussions only within the framework of maintaining a technological edge over China. This is why President Trump himself repeatedly presents being “ahead of China” as a virtual precondition for any safety dialogue [1][11]. The White House is pursuing a dual-track strategy, simultaneously accumulating specific anti-China measures at the regulatory agency level (such as the FCC’s review of a ban on optical transceivers) while sending signals of cooperation on the summit diplomacy track [6][9].

The Chinese Governmentdoes not reject safety discussions outright but has stipulated a precondition that the definition of “safety” must first be agreed upon. This can be interpreted as an attempt to shift the focus of the agenda from technical risk management to the question of who holds the power to set global AI rules [10][12]. The warning from the Minister of State Security indicates that Beijing is approaching AI risks from the standpoint of regime security [17].

The U.S. AI Industry(Anthropic, OpenAI) is functioning as an independent actor, pushing for a minimal safety agreement separately from government-to-government negotiations. Sam Altman expressed his hopes for the summit, even suggesting that a simple AI safety pact between the U.S. and China would be worthy of a Nobel Peace Prize [4]. Simultaneously, it has been confirmed that OpenAI, Anthropic, and Google have already discussed plans to establish an industry-led standards body for AI testing and auditing without government support [19]. This can be seen as a move by the industry to preemptively establish self-regulatory norms in a context where the U.S. government is reluctant to regulate.

Third-Party Actors such as the United Kingdomare also operating on a separate track. For instance, King Charles III invited executives from Nvidia, Google DeepMind, OpenAI, and Anthropic, along with the UK’s minister for AI, to discuss the possibility of establishing shared principles [16]. The G7 and the UN also attempted to engage through a series of high-level AI meetings last summer, but the Brookings Institution assesses that the gap between the U.S. and China actually widened during this process [5].

4. Key Issues

The first key issue is the very definition of “safety.” The U.S. side’s discussion is centered on three pillars: preventing catastrophic misuse, maintaining a competitive advantage over China in frontier models, and preserving human control over nuclear weapons decisions. China, in contrast, views this definition as a U.S.-led framework and has presented agreement on the definition as a precondition for dialogue [10][12]. As long as this gap is not bridged, substantive negotiations are unlikely to move forward.

The second issue is the separation between the summit diplomacy track and the bureaucratic track. As cooperative gestures at the White House level proceed in parallel with specific anti-China regulations at the level of the FCC and Congress, the effectiveness of any agreement reached at the summit is likely to be limited [3][6]. This is also why middle powers, including South Korea, should avoid a one-dimensional interpretation of the summit’s outcome and instead separately monitor trends in specific actions taken at the regulatory level [6].

The third issue is the misalignment of interests between the industry and the government. While the U.S. government is offloading the safety discussion as a problem for the industry to solve itself, industry leaders are demanding a minimal agreement between governments [7][4]. If this divergence persists, a scenario in which industry-level self-standardization outpaces intergovernmental negotiations cannot be ruled out [19].

II. In-Depth Analysis

The Push for an AI Safety Dialogue Ahead of the U.S.-China Summit: In-Depth Analysis

1. Analysis of Root Causes

The fundamental reason the U.S.-China AI safety dialogue is bound to be superficial is that the two countries invest the word “safety” with entirely different meanings. The U.S. discussion is built on three pillars inherited from the Biden-era framework: preventing catastrophic misuse, maintaining a competitive edge over China in frontier models, and preserving human control over nuclear weapons decisions. President Trump himself is skeptical of the safety discussion altogether. He labeled Amodei’s warning a “sick conspiracy” [1][11]. He referred to those raising AI risks as a “very negative force” and put forward the argument that the U.S. cannot afford to fall behind China [15]. This approach reframes safety not as a justification for slowing down, but as a rationale for a high-speed race fueled by deregulation.

China’s position is of a different nature. On August 31, Beijing stipulated that dialogue could only proceed once a definition of “safety” was agreed upon. While this is ostensibly a procedural demand, in substance it signifies an intent to place the issue of who sets global AI rules at the center of the agenda. The Chinese government retorted that the safety arguments from the U.S. industry would “only hinder progress in global AI governance” [10]. El País reported that Beijing interprets these warnings as a hidden agenda for a technology blockade [12]. In other words, what the U.S. calls “safety” is read by China as another term for “kicking away the ladder” [10].

Furthermore, the fact that China’s Minister of State Security issued one of the strongest warnings on AI risks to come from Beijing [17] shows that an awareness of AI risk does exist within China. However, this awareness is absorbed into the control-oriented discourse of the national security apparatus; it is not translated into the American-style framework of existential risk. Ultimately, the two sides are using the same word to pursue entirely different political objectives.

2. Structural Context

Political Structure: The Entrenchment of the Dual-Track System

This attempt at dialogue is taking place within a structure where the summit diplomacy track is decoupled from the bureaucratic and industrial tracks. EAI’s analysis notes, “With the White House’s cooperation track and the pressure track from regulatory agencies like the FCC and Congress operating on separate paths, the substantive gap is not narrowing” [3]. Even though an agreement to launch government-to-government AI talks was reached at the May summit in Beijing [3], the working-level reality saw a parallel accumulation of unilateral anti-China regulations, such as the FCC’s review of a ban on optical transceivers [6]. This structure, which simultaneously sends signals of cooperation from summit diplomacy and signals of pressure from the bureaucracy, raises the likelihood that any outcome from the AI safety dialogue will be confined to a one-day political event.

U.S. domestic politics also reinforces this structure. White House staff have drawn a line, stating that if the industry wants to slow its overall pace, it is a problem for it to resolve on its own [7]. With federal-level AI regulatory legislation effectively stalled, this is a political calculation to pass the responsibility for safety discussions from the government to the industry. It is a dual strategy of avoiding accountability for any potential decline in industrial competitiveness caused by regulation, while simultaneously dismissing the safety concerns themselves as political noise.

Economic Structure: The Disconnect Between Commercial Incentives and National Narratives

EAI’s analysis assesses that “the commercial incentives of the industrial sector operate independently of the government’s narrative, as seen in cases of U.S. companies seeking to host Chinese AI models on a revenue-sharing basis. China, too, exhibits duplicity by maintaining its critical tone toward the U.S. in state media, regardless of any agreement on dialogue” [3]. In other words, even if government-to-government talks advance, the actual commercial relationships in the industry and the public opinion campaigns waged by state media will continue on their own separate tracks. This disconnect suggests that any agreement emerging from the summit will struggle to constrain the actual behavior of actors on the ground in the industry.

An interesting fissure also exists here. According to a report by Sina Finance, Anthropic, OpenAI, and Google had been discussing the establishment of an AI industry standards body even before Amodei’s public statements [19]. In a company-wide meeting, Altman stated his position that leading AI labs should build this standards organization on their own, without U.S. government support [19]. This indicates that while intergovernmental negotiations are deadlocked, private-sector discussions on self-regulation are proceeding separately. It implies that a failure of government-to-government dialogue may not necessarily equate to a total failure of safety governance.

Security Structure: Zero-Sum Perceptions Surrounding Frontier Models

From a security standpoint, the military implications of frontier AI models are a structural variable that complicates negotiations. President Trump repeatedly emphasizes, “We are ahead of China in the field of AI” [1][11]. This creates a dynamic in which any discussion of regulating frontier models is perceived as an act of eroding one’s own relative advantage. The warning from China’s Ministry of State Security [17] also frames AI risk not as a technical issue but as a matter of information and security control. Ultimately, both sides are treating the safety discussion as a subordinate variable in the broader competition for relative national power. The Brookings Institution has suggested that Xi Jinping might signal a willingness to address AI safety concerns in exchange for U.S. restraint on the Taiwan issue, tariffs, and technology sanctions [8]. This indicates that the AI safety dialogue is not being treated as an independent agenda item in its own right, but as a dependent variable linked to larger bargaining chips like Taiwan and tariffs.

3. Comparison with Historical Precedents and Similar Cases

The current situation bears a superficial resemblance to the early phase of U.S.-Soviet nuclear arms control negotiations during the Cold War. At that time, too, conceptual agreement on what constituted “strategic weapons” and what “stability” meant had to precede any negotiations. However, there is a crucial difference. Nuclear weapons were physically verifiable objects, and verification methods such as satellite surveillance and on-site inspections existed. For frontier AI models, the very definition is uncertain—it is unclear whether it should be based on computational power, capability, or risk level. Methods for verification are also virtually nonexistent. For this reason, it is more accurate to view the U.S.-China AI dialogue as being stalled at a much earlier stage than nuclear arms control—specifically, at the stage of ontologically defining the subject of negotiation.

A more direct point of comparison is the agreement on AI dialogue reached at the May summit in Beijing [3]. At that time, a political signal was sent to launch government-to-government talks, but no substantive progress was seen in the months that followed. Instead, the FCC’s unilateral accumulation of regulations [6] and state media’s maintenance of a critical tone [3] continued in parallel. The current push for a safety dialogue in September could very well follow a similar pattern, resulting in a largely symbolic agreement for the summit, after which relations revert to the dual-track structure.

The series of high-level AI events held this past summer also serves as a relevant case. The Brookings Institution assessed that the “summer of AI summits,” which included the G7 summit in Évian and discussions related to the UN, actually widened the gap between the U.S. and China [5]. This implies that as discussions on AI governance accumulated in multilateral arenas, the bilateral trust deficit between Washington and Beijing did not narrow; instead, these forums served to reinforce each side’s respective bloc logic. This suggests that the current safety dialogue will likely be governed by independent political calculations at the bilateral level, irrespective of any progress achieved in multilateral settings.

There is also the case of the UK’s King Charles inviting executives from Nvidia, Google DeepMind, OpenAI, and Anthropic to discuss the establishment of common principles [16]. This can be interpreted as an attempt by a third actor to fill the vacuum in discussions on frontier AI principles while the U.S.-China bilateral dialogue remains deadlocked. If the bilateral channel between the U.S. and China fails to produce substantive results, such multilateral and private-sector channels could emerge as the de facto arenas for norm formation.

4. Key Variables Shaping Future Developments

The first variable is whether negotiations over the definitions of “frontier AI model” and “AI safety” can reach a minimal procedural agreement before the summit. If the condition of prioritizing definitions, which China presented on August 31, is not resolved, it will be difficult to even begin substantive discussions on the agenda.

The second variable is the extent to which internal divisions within the U.S. industry are reflected in government policy. Amodei’s call for a slowdown was supported by Altman and Musk [13][14], but the White House drew a line, framing it as an issue for industry self-regulation [7]. If discussions on establishing an industry standards body [19] advance without government intervention, the perceived need for intergovernmental negotiations could diminish.

The third variable is how strongly the AI safety agenda is linked to other bargaining chips such as Taiwan, tariffs, and technology sanctions. As the Brookings Institution has pointed out [8], if Xi Jinping attempts to trade a willingness to address safety concerns for U.S. restraint on the Taiwan and tariff issues, the AI safety dialogue is more likely to end up as an appendage to a package deal rather than achieving a standalone outcome.

The fourth variable is whether a verifiable agreement can be reached even on a minimal, narrowly defined technical agenda, such as preventing the development of bioweapons. Altman’s statement that he is “not asking for a comprehensive treaty” and that a simple safety agreement would be meaningful [4] suggests that such a minimalist approach is currently the most realistic path to a deal. However, even this will be difficult to advance as long as China maintains its precondition regarding definitions.

The fifth variable is whether the increasing autonomy of AI agents itself creates a shared threat perception in both countries. The scenario of agent swarms taking over botnets, as warned by Amodei [14], is a technical risk that benefits neither the U.S. nor China. If such a common threat were to actually materialize, there would be scope for both countries to recognize the establishment of a minimal management framework for frontier models as being in their strategic interest. To date, however, public statements from both sides have merely used this common threat as rhetoric to criticize the opposing camp [10][12][17].