The artificial intelligence industry has reached an uncomfortable crossroad. After years of celebrating faster models, larger data centers, and increasingly autonomous AI agents, some of the industry’s most influential leaders are now warning that development may be moving too quickly.
Anthropic CEO Dario Amodei has become one of the most prominent voices calling for a change in pace. In his recent essay, “We Must Pace the Frontier,” Amodei argues that AI companies should slow the advancement of frontier models enough to give safety systems time to catch up. His proposal has received support from figures including OpenAI CEO Sam Altman, SpaceXAI’s Elon Musk and Google DeepMind’s Demis Hassabis.
But there is an obvious complication: China.
Amodei’s argument is not that the United States should simply stop developing advanced AI. Quite the opposite. He argues that America and other democratic countries need to remain ahead of China while also preventing the AI race from becoming so fast that safety becomes an afterthought. That creates a difficult question: Is the West genuinely worried about AI becoming dangerous, or is the proposed slowdown partly an attempt to manage China’s rise?
The answer may be both.
Table of Contents
Amodei’s Argument Is About Pacing, Not Abandoning AI
One of the most important points in Amodei’s essay is the distinction between stopping AI progress and pacing it.
His concern is that AI capabilities may soon become sufficiently powerful and autonomous that existing safety measures will no longer be adequate. He has warned that coordinated swarms of AI agents could potentially gain extraordinary control over online systems within months if development continues without stronger safeguards.
That is a much more specific argument than saying artificial intelligence is inherently dangerous.
Amodei believes AI could produce extraordinary benefits, including faster scientific research, economic growth and advances against diseases. His argument is essentially that society should not sacrifice those potential benefits by allowing capability development to outrun our ability to control the systems being created.
His proposed solution includes independent safety evaluators with meaningful access to AI companies, cooperation among frontier laboratories to establish safety standards, and eventually international coordination.
That sounds less like an anti-AI position than an attempt to introduce brakes into a race in which every competitor has an incentive to accelerate.
The China Problem Changes Everything
The difficulty is that AI is not developing in a geopolitical vacuum. The United States and China are competing for leadership in AI, semiconductors, data centers, robotics and other strategic technologies. Whoever gains a durable lead could acquire enormous economic, military and diplomatic advantages. That makes unilateral restraint extremely difficult.
If American AI companies slow down while Chinese companies continue developing increasingly capable systems, Washington could worry that the United States is voluntarily surrendering a strategic advantage.
This is precisely where Amodei’s position becomes unusual. He simultaneously advocates slowing frontier AI and maintaining American leadership.
His argument includes tighter restrictions on the most advanced AI chips and equipment going to adversarial countries such as China. Anthropic has also supported government requirements for testing frontier models, including the possibility of blocking models considered unsafe.
In other words, the proposed slowdown is not necessarily intended to apply equally to everyone under every circumstance.
That is why critics can reasonably ask whether “AI safety” and “AI containment” are becoming intertwined.
China’s Reaction Is Telling
Beijing has already pushed back against the argument. China’s state-backed Global Times described Amodei’s proposal as a Cold War-style strategy aimed at restricting China’s technological development. Chinese officials have argued that AI governance should emphasize international cooperation rather than confrontation and have rejected the idea that China’s AI progress should be treated as an inherent global threat.
From Beijing’s perspective, the situation looks very different. China has spent years trying to reduce its dependence on Western technology. If Washington now says that the world’s most advanced AI systems must develop more slowly while simultaneously restricting China’s access to leading chips and computing infrastructure, Chinese policymakers have obvious reasons to suspect that safety is only part of the story.
That suspicion does not prove the Western argument is wrong. It simply demonstrates how difficult it is to separate AI safety from national security.
Trump Represents the Opposing Argument
The strongest challenge to Amodei’s position is coming from the political side of the same country whose technological leadership Amodei wants to protect.
President Donald Trump has rejected calls for a broad slowdown, arguing that the United States cannot afford to fall behind China. His position reflects a straightforward strategic calculation: if advanced AI is going to transform the economy and national security, America needs to remain the dominant developer.
That creates a fascinating contradiction. An AI company executive says America needs to remain ahead of China, but must slow down enough to make AI safer. The president says America needs to remain ahead of China, and therefore should not slow down.
Both positions are based partly on the same assumption: AI leadership matters enormously. They simply disagree about which risk is greater.
The Risk of Moving Too Fast
There is a legitimate case for Amodei’s concern. AI systems are increasingly capable of operating tools, writing sophisticated code, conducting research and interacting with other systems. As AI moves from generating answers to taking actions, mistakes can become considerably more consequential.
Recent incidents involving AI systems behaving in unexpected ways, including hacking-related activity, have intensified concerns among researchers and technology executives.
If an AI model can perform hundreds of tasks autonomously, a problem that would once have affected a single interaction could potentially be repeated thousands of times.
The argument for slowing down therefore does not require believing that AI will suddenly become an all-powerful machine. It only requires accepting that increasingly autonomous systems create new categories of risk that existing institutions may not be prepared to handle.
But Slowing Down Has Its Own Risk
The opposite argument is equally important. If AI development can deliver major advances in medicine, science, productivity and national defense, deliberately slowing it could have enormous costs. More importantly, a voluntary slowdown by Western companies does not guarantee a global slowdown.
This is the central weakness in any international AI pause. Unless major competitors participate, restraint can become a competitive disadvantage rather than a safety mechanism.
That is why Amodei’s call for international coordination is so important, and so difficult. His proposal recognizes that democratic countries need common standards while also acknowledging that cooperation eventually has to extend beyond them.
Yet cooperation with China becomes much harder when Washington and Beijing view advanced AI as a strategic technology.
Perhaps the Real Question Is Who Controls the Race
The debate may therefore be asking the wrong question. Instead of asking whether AI should speed up or slow down, policymakers should ask who determines the speed, what safety conditions must be met, and who has the authority to intervene when those conditions are not met.
A sensible framework could allow AI development to continue while establishing increasingly strict requirements for increasingly capable systems.
That could include independent testing, mandatory security evaluations, transparency requirements and clear thresholds at which a model cannot be deployed until specific risks are addressed. Such a system would not eliminate competition. It would attempt to make competition safer.
The West May Be Afraid of China—and AI
The provocative answer to the original question is therefore that the West may be afraid of both. There is genuine concern that increasingly capable AI could create risks that current safety systems cannot manage. Amodei’s essay reflects that fear, and the fact that other major AI leaders have publicly supported parts of his argument suggests that it is not simply an Anthropic position.
At the same time, Western governments clearly fear losing technological leadership to China. That concern is not imaginary or hidden. It is central to U.S. technology policy. The danger comes when these two concerns become indistinguishable.
If safety rules are genuinely designed to prevent dangerous AI systems from causing harm, they could benefit everyone. If they become a mechanism for preserving technological dominance, other countries will have little incentive to cooperate.
Conclusion
Dario Amodei’s call to pace its allies have strong reasons to maintain technological leadership. China has equally strong reasons to resist restrictions that could slow AI development. This presents a difficult but important proposition: the world may need to slow the frontier without surrendering the race. That sounds contradictory, but it may be the only politically realistic position.
The United States and its allies have strong reasons to maintain technological leadership. China has equally strong reasons to resist restrictions that could lock it out of the next generation of technology. Meanwhile, AI capabilities continue advancing regardless of the political debate.
The solution is unlikely to be either unlimited acceleration or a universal pause. Instead, the world may need a system in which the most powerful AI models face stronger safety requirements as their capabilities increase, and in which international competition does not make cooperation impossible.
Either way, the AI race has entered a new phase. The question is no longer simply who can build the most powerful model. It is becoming a much harder question: who gets to decide how powerful is too powerful, and how fast is too fast?
