Dario Amodei wants to slow AI. China isn’t taking orders

Anthropic CEO Dario Amodei has tried to put the brakes on the AI race, arguing that the strongest models should be slowed while the industry works out the risks. That pitch has run into a wall in China, where officials, state media and engineers are treating it less like a safety plan and more like a bid to protect U.S. advantage.

Amodei’s proposal went beyond a general pause. He called for stopping sales of powerful AI chips and semiconductor-making equipment to China, cracking down on chip smuggling, and restricting remote access to data centers outside China. He said the point was to buy “breathing room” for safer progress. He also drew support from some of Silicon Valley’s biggest names, including OpenAI’s Sam Altman and xAI’s Elon Musk. Mark Zuckerberg and Donald Trump went the other way, backing rapid AI growth instead.

Beijing answered quickly. Global Times accused the U.S. “tech right” of pulling a Cold War stunt, while foreign ministry spokesperson Guo Jiakun called the proposal “fear-mongering, confrontation and vicious competition.” On the surface, that sounds like a straight rejection. But China has its own expanding idea of AI safety, and it is not just about sci-fi doom. Its national cybersecurity standards body warned on September 14 about systems that could self-optimize and improve faster than humans can control. Chinese policymakers are also focused on information control, labor disruption, chatbots, cybercrime and other nearer-term risks.

There is a deeper irony here. China is pushing back on Amodei’s terms, but not pretending AI is harmless. State security minister Chen Yixin warned on September 13 that advanced U.S. models could threaten critical infrastructure and influence operations, and said whoever improves sensing, decisions and strikes fastest gets the edge on the battlefield. Chinese engineers are being even blunter. One MiniMax Code researcher mocked the idea of shifting from AGI to slowing down overnight. A DeepSeek engineer argued for making powerful AI cheap and widely available instead. The real fight is not over whether AI needs guardrails. It is over who gets to write them.