Yesterday, Dario Amodei published another essay, We Must Pace the Frontier, calling for a measured slowdown in the development of frontier AI capabilities to give safety research, testing, and regulation more time to catch up. Anthropic also announced that it would invite independent evaluators to be permanently embedded at the company to continuously review model training and safety measures. OpenAI CEO Sam Altman subsequently expressed support for a similar arrangement, saying it has also been a major topic of internal discussion at OpenAI in recent weeks. Elon Musk was supportive as well, responding simply on X: “Dario is right.”
Amodei devoted a substantial part of the essay to China. His argument is that how much U.S. companies can afford to slow down depends on how far ahead they are of China. If American companies slow down while China continues moving forward, the gap could narrow or even disappear. Therefore, while strengthening safety constraints at home, the United States also needs to preserve its technological lead over China in order to create room for slowing down frontier development.
He worries that if China were to take the lead in AI, it could gain significant military advantages, including through more powerful AI-enabled drones. He therefore argues that advanced AI chips and semiconductor manufacturing equipment should not be sold to China, while chip smuggling and Chinese access to computing power through data centers outside China should also be restricted. In his view, access to chips will be one of the main determinants of China’s future AI capabilities.
He also calls for action against unauthorized model distillation. Amodei argues that distillation could allow Chinese challengers to close the gap at a fraction of the cost of developing comparable capabilities independently. The concern is that while U.S. companies slow their own development for safety reasons, competitors could use American models to accelerate their own catch-up.
Amodei also argues that AI companies need stronger internal security to prevent model weights from being stolen. He believes that, if these measures are implemented effectively, they could significantly widen the U.S. lead over China over the next three to five years. A larger lead, he argues, would also give the United States and its allies greater negotiating leverage and make future agreements with China more achievable.
At the same time, Amodei advocates discussing some form of global coordination with China, starting with specific risks where agreement may be relatively easy. One area he considers more realistic is banning the use of AI to develop biological weapons and other clearly dangerous applications, since such risks threaten all sides. A further step could be an agreement requiring both sides to test frontier models before deployment for risks involving cyberattacks, biological threats and loss of control.
A more difficult form of cooperation would be to limit the speed at which AI systems can help develop the next generation of AI. Amodei draws an analogy with U.S.-Soviet strategic arms control: each side would retain its strategic capabilities while placing limits on forms of competition that could create enormous risks. As for a comprehensive slowdown or even a pause in frontier AI development, he believes it is worth discussing but unlikely to happen in the near term, because every side would worry that it might honor the agreement while others secretly continue developing more powerful systems.
Amodei also emphasizes that verification would be one of the biggest obstacles to international coordination. Model training can be harder to detect than missile facilities, while the computing infrastructure involved has many legitimate commercial uses. If some participants pause while others secretly continue, the balance of power could shift significantly. Any agreement would therefore need clear rules on when restrictions are triggered, when they can be lifted, and who is responsible for verifying compliance.
China’s state-run Global Times quickly responded, describing Amodei’s proposal as a “Cold War script” coming from America’s “tech right.”
The commentary argued that Amodei’s proposal may appear on the surface to be a rational call for greater global AI safety, but that a closer reading reveals a series of measures aimed specifically at constraining China. In its view, the underlying idea is to first widen the U.S. lead by restricting China and only then negotiate with Beijing over a mutual slowdown.
Notably, the commentary highlighted three measures in particular: restricting sales of advanced AI chips and semiconductor manufacturing equipment to China, cracking down on model “distillation,” and preventing the theft or leakage of frontier model weights.
According to the Global Times commentary, this approach departs from the idea of shared technology governance and instead politicizes technological issues. It characterized the proposal as a “quiet AI Cold War” that is both hypocritical and short-sighted. In its view, the real objective is to use technological barriers and control over rule-setting to constrain China’s AI development, preserve Washington’s dominance in frontier technologies, and exclude China from the global AI governance system. It argued that trying to contain China while simultaneously seeking cooperation with Beijing on global AI safety is fundamentally contradictory and unlikely to succeed.
The commentary:
Targeting China’s AI: America’s “tech right” rolls out a Cold War script
Anthropic CEO Dario Amodei published an essay calling for the pace of frontier AI development to be controlled in order to reduce the risks of technological loss of control. The proposal quickly received support from prominent figures in the U.S. AI industry, including Elon Musk and Sam Altman.
On the surface, this appears to be a rational call focused on global AI safety. But a closer reading reveals numerous provisions aimed at containing China, effectively amounting to a “Cold War script” for AI. Amodei argues that democratic countries “cannot afford to be naive” in dealing with China. To prevent China from overtaking the United States while American companies slow their AI development, he proposes first widening the gap by constraining China, and only then seeking negotiations with China over a mutual slowdown. The proposed measures include banning sales of high-performance AI chips and semiconductor manufacturing equipment to China, cracking down on so-called model “distillation,” and preventing the theft or leakage of core model weights through hacking.
The rapid development of AI has undoubtedly created safety risks, ethical concerns and the possibility of technological loss of control—global challenges facing all humanity. But the core of Amodei’s proposal runs counter to the principle of shared technology governance by politicizing technological issues. This “quiet AI Cold War” is hypocritical and short-sighted. Its real purpose is to use technological barriers and control over rule-setting to constrain China’s AI development, preserve Washington’s dominance in frontier technology, and exclude China from the global AI governance system. It follows the same logic as Washington’s so-called “Silicon Peace” initiative, which also seeks to exclude China.
Trying to talk about global AI safety while simultaneously “containing China” is both a fundamentally mistaken judgment and an unrealistic proposition. China is already one of the world’s largest and most open suppliers within the global AI ecosystem, and its AI capabilities can no longer simply be treated as a variable that can be excluded. Artificially fragmenting global AI industrial and innovation chains and excluding China and Chinese companies would only increase the costs of experimentation and the risks of losing control. Under such circumstances, the idea of “controlling the pace” would itself lose much of its meaning.
Another Chinese critique was even more direct. An article republished by Phoenix New Media was titled Anthropic Hits the Brakes for Safety: Slowing Silicon Valley While Blocking Chinese Companies. The author puts chip restrictions, anti-distillation measures and the broader safety initiative into the same framework, arguing that the proposal also contains a clear competitive objective: making it harder for Chinese companies to catch up.
The article acknowledges that AI safety incidents and regulatory pressure are real, but argues that readers should also pay attention to the interests of those writing the rules:
But if we interpret this collective effort to hit the brakes simply as a sudden awakening of conscience, we are giving Silicon Valley too much credit. There are at least two other sides to this “braking” proposal.
One is competition. The speed-limit proposal comes from the companies currently leading the race—Anthropic and OpenAI. The more complicated the rules become, the higher the thresholds are set, and the greater the emphasis on “checkpoints,” the more expensive it becomes for those behind them to catch up.
Dario’s plan even explicitly calls for measures such as chip controls to preserve a three-to-five-year technological gap over China. Embedded within a proposal calling for a global slowdown is therefore a clear competitive provision: hit the brakes on yourself, but block the road for your competitors.
Two days earlier, after repeatedly accusing Chinese large-model developers of “distilling” American models, Anthropic once again went on the offensive, publishing a lengthy report accusing several Chinese AI companies of improperly using user data.
Anthropic alleged that Chinese AI companies fed conversations between users and Chinese large models into Claude, and then used Claude’s responses to train their own models, thereby “distilling” Claude’s capabilities.
Anthropic has repeatedly taken the lead in calling on the AI industry to put on its safety belt. But this means not only slowing Silicon Valley down; it also means blocking the road for Chinese companies by cutting off distillation as a shortcut for rapid catch-up.
The other side is reality. So far, not a single company has actually hit the brakes.