Today, when China and the US list the risks posed by artificial intelligence, they often mean different things. The cognitive frameworks through which the two sides understand risk are not the same.
This is hardly a new phenomenon. It is a recurring feature of US–China communication. For example, when discussing “crisis management,” the US side hopes to establish hotlines so that intense competition can continue “safely.” China asks the US to stop conducting close-in reconnaissance, thereby eliminating what it sees as the crisis's root cause. Similar conceptual differences have repeatedly emerged in the two countries’ definitions of “terrorism,” “sovereignty,” and other terms. They are now resurfacing in discussions of “AI risk.”
I’m not saying that the two sides have no common ground. During the Biden administration, China and the US both supported the Bletchley Declaration, which pointed out that frontier AI could pose serious risks, including cybersecurity risk, malicious use, and insufficient human control over AI systems. This shared understanding has at least provided some basis for cooperation.
But still, the two countries think of different kinds of risks first when they use the term.
My insufficient observation is that, partly because AI company CEOs shape the public debate, the most obvious concern among US frontier AI companies and in some American policy circles is whether humanity can remain in control of AI. And the NIMBY effects of data centers can also be partly attributed to locals’ lack of confidence in controlling it; once it’s built, it’s out of their control. (Also recommend to read Jasmine Sun’s piece on that )
While China seems more concerned with how AI services may affect existing social structures, for example, how “AI family members” might change young people’s understanding of kinship and family relationships. The central question is whether society can absorb the disruption AI causes.
The former worries that the rapid growth of AI capabilities could outstrip humanity’s ability to control them. The latter focuses more on AI’s effects on interpersonal relationships and social interaction. These examples do not capture the full range of views in either country, of course, but they do offer an interesting contrast.
In an article published in September 2026, Dario Amodei argued that the development of frontier AI models should be slowed. His central concern was that the cycle of capability improvement could exceed humanity’s ability to understand and control these systems. He called on the US government to impose tougher restrictions on China’s ability to develop AI, including limits on technology transfers and China’s access to advanced chips, as part of this effort to slow development.
His argument quickly provoked heated discussion, or I think it’s more accurate to say unified opposition on the Chinese internet. Most believed Amodei was trying to capture the regulatory to consolidate his company’s position in advanced AI models. The Global Times characterized his proposal as an attempt to constrain China’s AI development through technological barriers and a monopoly over rule-setting, thereby preserving America’s dominance in frontier technologies.
Some Chinese regulatory and economic policy debates reflect different concerns.
On the regulatory side, on April 10, 2026, the Cyberspace Administration of China and several other government agencies published the Interim Measures for the Administration of Anthropomorphic Interactive Artificial Intelligence Services. The measures prohibit providers from offering minors services that simulate intimate relationships, like virtual family members and romantic partners. They also require providers to guard against emotional manipulation, excessive dependence, and harm to real-world interpersonal relationships.
In my view, these measures reflect an understanding of AI risk centered on how AI shapes human interactions and people’s perception of reality. Policymakers seem particularly concerned that if people become accustomed to AI family members that constantly accommodate their preferences. Forming emotional attachments to them may also make youth struggle to understand the disagreement and mutual compromise inherent in real family relationships. AI could become like the memory parasites in Rick and Morty, which reproduce by creating only pleasant memories of themselves.
A similar way of concern appears among some of China’s senior economic policy advisers. Economist Luo Zhiheng, who attended economic symposiums hosted by Premier Li Qiang in July 2023 and October 2024, focused in an article published earlier this month on the “K-shaped divergence” between AI and the rest of the Chinese economy.
Luo argued that the strong performance of AI-related industries and investment, combined with the weakness of traditional sectors such as real estate, is widening income disparities across industries and social groups. Because AI industries create relatively few jobs, the economic gains they generate are less likely to be shared across different sectors of the economy.
His policy recommendations emphasize reforms to China’s income-distribution systems. The point is to ensure that gains from the economic transition are shared and to prevent current disparities from hardening into long-term structural imbalances.
Huang Yiping, dean of Peking University’s National School of Development, has a similar argument. He warned that the AI era could reproduce the “Engels’ pause”, which is a period when labor productivity rises rapidly while workers’ real wages remain low, causing social inequality to grow sharply. Huang has called for an AI dividend-sharing fund that would channel some of the gains generated by AI back to the public.
Taken together, these discussions have a similar concern framework. Compared with concerns about “AI escaping human control,” which have become influential in the US, Chinese regulatory and economic policy debates focus more directly on the social consequences people may have to endure.
This also helps explain why any study of China’s understanding of AI risk cannot rely on the voice of a single government agency.
The latest article by Minister of State Security Chen Yixin is definitely a policy text worth reading. But an article's importance and the range of views it can represent are not the same.
If the director of the FBI published an article about AI’s effects on US national security, the Chinese US watchers should certainly read it. But they won’t treat it as the basis for understanding how the US perceives AI risk and its competition with China. Chen’s article should be approached in the same way.
Under the background of US–China competition in AI, the national security perspective undoubtedly carries more weight. But there’s a reason that Chinese policy language routinely calls for “coordinating development and security,”统筹发展与安全, with “development” and “security” appearing side by side. Concerns about AI’s social effects on the economy, employment, income distribution, and youth development remain relevant.
Chen’s article discusses AI cyberattacks, espionage, imbalances in global technological development, the failure of governance mechanisms to keep pace with technological change, and how it might change the war. These issues also reflect a concern with how governments and societies should respond to the social consequences of AI.
Such social concerns overlap with national security objectives, but it would be overly simplistic to reduce it to political security. Summarizing China’s view as “China cares only about political security” would obscure the range of concerns in Chinese policy debates.
Returning to the differences between Chinese and American understandings of AI risk, I would interpret them as different starting points rooted in each country’s social realities, which should not be treated as two mutually exclusive national positions.
It is neither realistic nor necessary to require the two sides to reach the same conception of risk before they can engage with each other. Instead, each side needs to understand what the other is trying to save in the name of “safety,” and what outcomes it would regard as reducing risk.
Otherwise, even if they use the same terminology and endorse similar principles, China and the United States may still fail to reduce the risks posed by AI.
Very helpful framing. Different definitions of the same word explain a lot of failed dialogue. Agreeing on a shared vocabulary for AI incidents could be a modest but meaningful first step.