As Dario Amodei and Sam Altman call for slower development of frontier AI, parts of the English-language press sound increasingly apocalyptic. Amodei, a persistent advocate of restrictions on China, now wants Washington to talk to Beijing about preventing catastrophe. Yet Chinese perspectives remain strangely peripheral to a debate that supposedly cannot succeed without China.
I do not take either executive’s assurances at face value. But distrust of American AI companies does not automatically create trust in Chinese ones. Americans have legitimate questions: Do Chinese companies harbor equally expansive ambitions? And, more importantly, can China be trusted on AI safety?
Start with how the laboratories describe their work. When dangerous capabilities emerge, Anthropic and OpenAI emphasize the possibility of systems escaping human control. Chinese executives tend to frame progress differently. In reported investor discussions, DeepSeek’s Liang Wenfeng described his team as ordinary people. Z.ai co-founder Jie Tang has emphasized open models’ auditability and controllability. Zhilin Yang, whose company Moonshot AI develops Kimi, focuses on the technical work of scaling model capabilities.
The distinction is not simply investors versus scientists. Amodei has a research background, while Liang made his fortune in quantitative finance. Rather, I see different public postures: American executives increasingly present themselves as stewards of civilization; their Chinese counterparts more often sound like researchers building a technology.
DeepSeek illustrates the difference. Despite Liang’s financial background, its early development resembled a personally funded research undertaking more than a conventional venture-backed startup. Outside fundraising was not initially its organizing principle.
Who Gets to Define the Risk? #
Before comparing the two approaches, we need to define the primary actor responsible for AI risk management: the institution or individual expected to identify AI safety risks, organize the response and answer for how those risks are handled. The question is not simply who writes the rules, but who takes responsibility for managing the danger.
In the US debate, frontier-lab CEOs increasingly cast themselves in that role. They define the risks, propose safeguards and call for third-party oversight. In China, the government occupies that position: AI safety is part of technology policy, with public authorities setting priorities and coordinating the response, while companies operate within—and help implement—that framework.
This distinction does not absolve Chinese companies of their own safety obligations, nor does it mean American CEOs possess formal regulatory authority. It identifies the central difference in the two approaches: who assumes the lead responsibility for managing AI risk.
Anthropic helped popularize Constitutional AI, an approach that trains models using explicit principles. Amodei has also argued that authoritarian governments’ access to advanced AI creates particular dangers. His latest proposal combines safety coordination with tougher restrictions on China’s access to chips, semiconductor equipment and overseas computing.
The logic is explicit: preserving and expanding the American lead would give US labs room to slow down, while increasing Washington’s leverage in negotiations with Beijing.
Alongside that geopolitical argument sits a familiar warning: sufficiently capable AI could escape human control and undermine existing social institutions. Other countries might continue developing it without adequate restraint, exposing everyone to danger.
To a Chinese reader, this combination can look both insulting and incoherent. Amodei is entitled to advocate policy. But he has no mandate to decide what constitutes an acceptable risk for another country.
If the concern is genuinely humanity’s collective safety, why must cooperation begin with preserving America’s technological superiority? If export restrictions are intended to widen the gap, why should China understand the resulting negotiations as a partnership between equals? Even on competitive grounds, the starting positions differ. Chinese developers cannot freely purchase Nvidia’s Blackwell GPUs and face constraints in assembling frontier-scale training infrastructure. Huawei’s Ascend processors offer an alternative, but replacing a chip supply also requires software, networking and manufacturing capacity. Access to competitive compute remains a practical bottleneck.
Against that backdrop, pressure to restrain China’s computing expansion does not sound like a neutral safety measure. It sounds like the stronger competitor asking the weaker one to accept a permanent handicap.
Imagine a school organizing a charitable collection, then letting its biggest bully beat donations out of the other children. The worthy cause does not legitimize the coercion. That is how this mixture of safety rhetoric and technological containment can land in China.
The Risks China Is Already Governing #
China’s government-led approach starts more directly with AI’s effects on society. The premise is that deployment can gradually alter social relationships, institutions and the distribution of power.
A recent essay by Minister of State Security Chen Yixin illustrates that perspective. He discusses AI-enabled cyberattacks and espionage, unequal technological development, and governance mechanisms struggling to keep up. His concerns also explicitly include political and ideological security—an important reminder that Beijing’s definition of safety reflects its own political system.
These are not identical to a lab CEO’s warning that a model could become uncontrollable. Both belong in an AI-safety discussion, but treating them as interchangeable obscures what each side wants protected.
Consider China’s Interim Measures for the Management of Anthropomorphic AI Interactive Services, which took effect on July 15. They prohibit providers from offering minors simulated intimate relationships, including virtual relatives and romantic partners. They also prohibit emotional manipulation and practices that encourage dependency or damage real-world relationships.
The concern is understandable. A child accustomed to an artificial family member that endlessly accommodates their preferences may struggle with the disagreement and compromise that real relationships require. This is not proof that every AI companion causes harm. It is a reason to examine the incentives of services designed to keep users emotionally attached.
The logic recalls Black Mirror: shielding a child from every distressing sight does not necessarily prepare them to deal with distress. A world without visible injuries is not the same thing as learning what to do when someone is bleeding.
Employment presents another version of the same problem. In an interview with X.PIN, demographer Huang Wenzheng argued that work cannot be the sole measure of human worth. As AI reshapes employment, society needs distribution mechanisms that reduce the fear of being left without an income.
He calls his proposal “starting-point income”: citizens, understood as shareholders in their country, should receive a dividend. This is an individual policy proposal, not an adopted government program. But it addresses an AI-related risk through institutions and distribution rather than through limits on model intelligence.
The emphasis matters. Chinese regulatory and economic debates are often directed at consequences people may actually have to live with: damaged relationships, lost earnings, concentrated power and compromised security.
That does not mean Chinese researchers cannot understand loss-of-control risks, or that nobody studies them. It means the American framing does not automatically organize the Chinese debate.
A call to slow AI, presented through a story in which China is the designated villain, therefore faces a problem before negotiations even begin. China is already developing under American-imposed hardware constraints. Being asked to accept another brake is unlikely to sound reassuring.
A Trust Problem Before a Safety Pact #
There is a less abstract obstacle: distrust toward Anthropic.
Some Chinese users associate Claude with expensive access, suspended accounts and a company that treats China primarily as a security threat. Anthropic’s allegations of unauthorized model distillation have deepened that antagonism.
Those issues should not be collapsed into one accusation. Anthropic says it does not offer commercial access in China, and its distillation allegations concern organized campaigns to extract model capabilities—not evidence that every Chinese person with a suspended account was doing so.
Nevertheless, the distinction can disappear in the customer’s experience. People who feel rejected or treated with suspicion by a service provider are unlikely to regard its CEO as an impartial guardian of their future. Anger at Amodei can spill over into dismissal of his entire safety argument.
That is unfortunate. A messenger’s credibility and the validity of a technical risk are separate questions. But anyone seeking international cooperation must take both seriously.
From outside the US, the slowdown debate also suggests another struggle: technology executives are contesting who gets to govern AI. The Information reports that Anthropic, OpenAI and Google have discussed creating an industry standards body. Altman reportedly told employees that major labs might have to establish it without US government support.
An industry body is not inherently illegitimate. Technical standards often require industry expertise. The concern is whether the companies being supervised also get to determine who supervises them, which risks count and how competitors gain approval.
Trump has pushed back. During his visit to Ireland, he emphasized America’s lead over China and rejected the idea that alarming predictions should dictate the pace of development. He has repeatedly promoted data centers as engines of local prosperity.
His response exposes a commercial fault line as well as a political one. Chipmakers and infrastructure providers benefit from continued expansion. Frontier-model companies increasingly argue that the pace of capability development needs supervision. Their interests overlap, but they are not identical.
That leaves several possible interpretations of Anthropic’s position:
- Its models may genuinely be advancing beyond the safeguards it can reliably implement, while safety work struggles to receive sufficient priority.
- It may want time to adjust its competitive position.
- Describing its technology as dangerously powerful may reinforce the story it sells investors.
These are hypotheses, not established explanations of Amodei’s motives. They can also coexist. A company can identify a genuine danger while benefiting commercially from the way it describes that danger.
Warning About Risk, Writing the Rules #
There is another possibility worth considering: warning about danger can become a way of shifting responsibility.
If a serious incident occurs, an executive who repeatedly demanded stronger oversight can point to those warnings and blame government inaction. The warning may be sincere. It can still provide reputational cover. Meanwhile, if Anthropic helps establish an oversight framework that government later adopts, it gains influence over the rules used to judge everyone else. Foreign developers and open-model projects could find themselves assessed against assumptions favored by a leading American competitor.
That outcome is not inevitable, and it is not what an independent regulator is supposed to deliver. But the possibility of regulatory capture deserves scrutiny, particularly when safety proposals are bundled with measures designed to preserve one country’s lead.
This is the paternalism I object to: a private executive presenting his preferred hierarchy as the natural starting point for everyone else’s safety.
Could AI profoundly disrupt human society? Certainly. Coding and information generation are already consequential. My skepticism concerns the leap from those capabilities to claims of autonomous mastery over the physical economy.
LLMs have not demonstrated a general ability to conduct experiments, organize production, manufacture equipment and continually improve themselves across that entire chain without human support. Fluent reasoning about factories is not the same as running one.
Anthropomorphizing agents—as discussions around interviewers such as Dwarkesh Patel sometimes do—does not turn them into workers spending 18 hours a day on an assembly line building Terminators. The physical world imposes constraints that a compelling conversation can make easy to overlook.
None of this eliminates digital risks. It does mean that claims about AI’s power should distinguish demonstrated capabilities from extrapolation.
Should Washington and Beijing discuss AI safety? Absolutely. But the starting point must be understanding what the other side wants protected, not requiring both societies to adopt an identical definition of safety.
Different political systems create different priorities. Negotiation exists precisely because those priorities do not automatically align.
Before their leaders reach an agreement, private executives should not seize the microphone and present a universal slowdown as a settled obligation. In April, Treasury Secretary Scott Bessent put America’s AI lead over China at three to six months.
It is a difficult sales pitch: tell someone they are behind, restrict their access to equipment, then ask them to run more slowly.
What a DeepSeek Engineer Fears #
For a different account of AI risk, consider the following passage from intlsy, a DeepSeek engineer. It is a personal argument, not a statement of company policy. The rhetoric is deliberately incendiary, but the underlying concern is who will own advanced intelligence and who will be allowed to use it. As AI develops, society may move toward one of two extremes: communism or Cyberpunk 2077. In the former, productive capacity expands enormously and living standards improve substantially. (I will leave it there; otherwise, I worry this will not get past the censors.)
In the latter, a handful of technology companies control most resources. Only a tiny minority can access the most advanced AI and other technologies, achieving something approaching technological transcendence, while most people are left with weak AI.
Upward mobility becomes increasingly difficult: you need the strongest AI to move up, but moving up is how you gain access to it. The result is a vicious circle.
If Anthropic permanently controlled the world’s most advanced AI, which future do you think we would get—communism or 2077? Take a guess. That is why I still believe frontier intelligence should be available to everyone, openly and affordably. I do not trust Anthropic or OpenAI to deliver that. In particular, I do not want Anthropic to control the most advanced AI or AGI. To put it hyperbolically, that would be no less alarming than Hitler acquiring atomic-bomb technology before the Allies.
This is also why I chose DeepSeek and have stayed: we develop powerful, fast, broadly accessible AI and open-source it. Perhaps that could pull the world a little farther away from 2077.
We also did a full translation of intlsy’s article. Click here to read →