Anthropic CEO Dario Amodei’s calls to slow the pace of artificial intelligence development over potentially existential safety risks have thrust the AI rivalry between the US and China into the spotlight days before the two superpowers’ leaders meet in Washington.
China, which already imposes strict internet controls, has also sounded the alarm on AI safety. However, Beijing’s approach to mitigating such risks is through government regulation, rather than hitting the brakes on the dazzling sector.
At the same time, Amodei’s appeals for US regulators to help maintain American developers’ lead over China by tightening restrictions on sales of AI chips and chipmaking equipment to China have irked Beijing. Chinese officials and state media have criticized his warnings as an effort to curb their country’s rise, casting a shadow over the chances of significant cooperation.
AI safety is expected to come up during Chinese President Xi Jinping’s visit to the White House for talks with US counterpart Donald Trump, who has downplayed AI fears. Here are five things to know.
How has China responded to calls for an AI slowdown?
Chinese officials have criticized the warnings from Amodei and other American AI leaders as an effort to “monopolize” development. A Foreign Ministry spokesperson warned that the executives are “spreading narratives of threat, engaging in confrontation and resorting to malicious competition.”
The Chinese Communist Party-affiliated Global Times wrote that Amodei’s latest essay on risks—in which he said that a swarm of AI agents “could be capable of taking over the entire internet” in six to 12 months—is “geared toward safeguarding the US lead in AI relative to China.”
China has also pushed back against allegations by Amodei and US government agencies that Chinese labs are using “industrial-scale” distillation, a technique used to train AI models using input from larger models, to soak up the capabilities of frontier American models. A commentary published by the state-run People’s Daily described distillation as common practice, and accused Washington of a “blatant double standard” by allowing US companies to distill Chinese models while labeling the same activity by Chinese companies an “attack.”
The US “continues to adhere to a hegemonic logic where ‘only the US wins,’ rather than ‘shared success for humanity,'” the article argued.
How does Beijing view AI risk?
Although Chinese officials dispute the US warnings and characterization of Chinese AI as a threat, they have become increasingly vocal about security and the need to avoid a “loss of control.”
In an article published on September 13, Chen Yixin, China’s minister of state security, warned that the latest Anthropic and OpenAI models can “exploit cyber vulnerabilities” and develop malware. He also wrote that the open-source AI agent tool OpenClaw, which went viral in China earlier this year, can remotely hijack devices and leak sensitive user information.
His worries echoed a speech Xi gave at the World Artificial Intelligence Conference in Shanghai in July, where the president called for stronger oversight and governance to “ensure that AI remains under human control at all times.”
Over the past year, China’s AI governance has expanded beyond content control to address broader risks, most visibly those posed by increasingly autonomous agents, according to a July report by Beijing-based AI safety consultancy Concordia AI. The company also found that frontier AI safety research in China has intensified: Agent safety accounted for more than a quarter of new papers in early 2026, up from 8% in early 2025.
“I think their biggest concern is that frontier AI systems, especially when available to the US government, could destabilize China’s cyberspace in ways that affect the physical world, hence threatening regime security,” said Rebecca Arcesati, head of the science, technology, and innovation program at the Mercator Institute for China Studies (Merics).
Jacob Cooke, co-founder and CEO of WPIC Marketing and Technologies, an e-commerce and technology consultancy, said that while the Chinese government seems to be taking general AI risks more seriously, it is unlikely to embrace a global research slowdown or self-regulation because it deems such proposals “vehicles for containing China’s AI development, similar to US chip restrictions.”
How have Chinese companies responded?
Discussion of AI’s long-term threat to humanity appears muted among China’s tech giants. They are focused instead on immediate operational risks.
Since Anthropic’s Claude Mythos preview shook the cybersecurity world earlier this year—the company deemed the model too dangerous to release publicly—Chinese companies have been racing to catch up. Several leading players have formed dedicated teams to develop products comparable to Mythos, though they are still lagging, according to a source close to the efforts.
Zhipu, also known as Z.ai, has been one of the more outspoken Chinese companies on AI safety. When it released its GLM-5.3 model last month, it also launched a “Shield of Open Source” initiative, offering free security audits to help users fix software vulnerabilities, and automated code-auditing tools, through its ZCode platform.
However, during a media roundtable at an annual Huawei event in Shanghai on September 17, rotating chairman Eric Xu said Chinese companies should move faster so that they can feel the same risks that American AI leaders already sense.
Xu argued that China’s largely open-source AI models are relatively transparent, while US companies with far greater computing power may be the only ones who know how sophisticated their models have become and what dangers they pose.
Xu added that the industry must balance rapid development with effective safeguards. “The bottom line is we need to ensure AI for good,” he said, “not AI for bad or evil.”
What is the current state of US-China AI competition?
Industry watchers say the gap between US closed-weight models, which are run through their vendors, and Chinese open-weight models, which can be downloaded and run on one’s own hardware, is rapidly closing.
And while the US is still considered ahead in AI performance, China leads in adoption. In one survey published by Morgan Stanley on September 14, 80% of respondents in China used AI at least once a week, compared to 54% in the US.
“The proliferation of Chinese open-weight models as a cheaper alternative is not only shaping global tech markets seeking to distance themselves from dependence on US tech, but also having impacts within the United States, as many startups have begun using open-weight models to build and train their own models more cheaply,” said a report by the Center for Strategic and International Studies published last week.
What can be expected during the Trump-Xi summit?
US treasury secretary Scott Bessent and Chinese vice premier He Lifeng met in New York ahead of Xi’s visit to the White House. In the run-up to the visit, the US was thought to have discussed “shared risks” in AI with China, Bessent told US media outlet Axios.
“Given the strategic utility of AI superiority is undeniable, striking the right balance and establishing an appropriate framework for addressing these issues may be the most important topic of the Trump-Xi meeting,” said Andy Mok, a senior research fellow at the Beijing-based Center for China and Globalization.
Sun Chenghao, a fellow with the Center for International Security and Strategy at Tsinghua University, said the two sides may find common ground in AI safety. “Both countries increasingly recognize that some AI risks are transnational and will not stop at national borders simply because the two countries are strategic competitors,” Sun wrote on Thursday. “Addressing these issues does not require prior agreement on chip controls or technological leadership.”
But other analysts expect mutual distrust to limit progress. The two sides agreed to hold a formal AI dialogue after the presidents met in May, according to Beijing, but no talks have been held officially.
Gavekal analyst Laila Khawaja said in a note that bilateral talks on AI are unlikely to produce binding agreements or major policy shifts. “The prisoner’s dilemma of AI competition among US AI labs and the competitive nature of the US-China relations means the call to slow down frontier AI development is unlikely to materialize,” she said, adding that the most realistic outcomes are continued dialogue, establishment of crisis communication mechanisms and incremental cooperation in narrowly defined areas of shared concern.
Arcesati at Merics cautioned, “Trust between the two countries is extremely low, and no one side seems to have the incentives to put safeguards on its own industry’s development so long as the other doesn’t move first.”
This article first appeared on Nikkei Asia. It has been republished here as part of 36Kr’s ongoing partnership with Nikkei.