Beijing Rejects Pause, Slams Dario's 'Fearmongering, Cold War Playbook'

China's Foreign Ministry and state press had rejected the China provisions of Dario Amodei's essay calling for a 'pause' in AI development, while state security minister and President Xi Jinping laid out what Beijing wants instead.

On Saturday morning, Anthropic CEO Dario Amodei published We Must Pace the Frontier, an essay arguing that the industry must slow the rate at which it improves frontier models and asking Washington to help it do so. By Monday afternoon in Beijing, China's Foreign Ministry responded. 

The essay's three steps, are third-party evaluators with permanent, employee-level access inside the labs, which Anthropic committed to unilaterally; coordination among labs in democratic countries on safety standards and the pace of progress, which Amodei concedes needs a narrow antitrust waiver from the U.S. government; and, eventually, agreements with various governments, including China. It also asks Washington to keep the ban on advanced chips and chipmaking equipment, enforce it against smuggling and remote access to overseas data centers, crack down on distillation, prevent model-weight theft, and use the resulting three-to-five-year window to widen America's lead before any bargain is struck. Sam Altman said OpenAI would match the evaluator commitment. Elon Musk's reaction was three words: "Dario is right."

Amodei himself acknowledged the problem. On CBS on Sunday he called the possibility that China and other adversaries would not slow down the "toughest dilemma" in his proposal.

According to Xinhua, Beijing's Foreign Ministry spokesman Guo Jiakun said on Monday that the development of AI bears on the well-being of all humanity and that all parties should jointly promote its open and inclusive development for good and for all. He added: "Fearmongering, confrontation and vicious competition will only disrupt the process of global AI governance which serves no one's interest."

On Sunday evening Beijing time, a Global Times editorial admitted that sure - on the surface the essay appears to be a "rational statement" about global AI security. A closer reading, the paper said, showed it "packed with containment provisions targeting China" and, in essence, a "Cold War playbook" for the AI sector - and that excluding China from the global innovation system would increase, rather than reduce, the "trial-and-error costs and risks of loss of control" in global AI development. China's AI capabilities "have long ceased to be a variable that can be excluded," reads the editorial. 

Other Chinese notables chimed in as well. Xiang Ligang, a telecom and technology policy commentator, called Amodei's rhetoric inappropriate, groundless and hostile. After all, the 'pacing' would ultimately require China's cooperation while advocating restrictions on chips, computing power and models to slow China down. Xiao Qian, vice dean of Tsinghua's Institute for AI International Governance, attributed the China provisions to commercial pressure: Anthropic's closed-model approach is competing with Chinese open-source models on cost, performance and the developer ecosystem, and restrictions would protect its position. Liu Shaoshan of the Shenzhen Institute of Artificial Intelligence and Robotics for Society, who the paper says previously worked with Amodei, added that framing AI as decisive for national security turns an AI company into "strategic infrastructure," which raises barriers to entry and valuations at the same time.

On Wednesday, Beijing's commerce ministry rejected a joint FBI, NSA and CISA advisory that accused Chinese developers of "aggressive, malicious" efforts to distill capabilities from Claude and GPT, calling the accusation "groundless in fact and without basis in law," distillation a normal technical and commercial practice, and the advisory further proof that Washington is "seeking to monopolize computing power."

Amodei's essay, published three days later, asks the U.S. government to crack down on the same practice in the same document that asks the industry to slow down for the sake of humanity. 

Trump Responds

Asked on Thursday in Dallas whether he had any concern about existential risk from AI, Trump said, "No, I don't have any." On Sunday, speaking to reporters at his Doonbeg resort during the Irish Open, he said the United States is "leading China in AI" and intends to stay there because "whoever wins AI, wins." Guardrails were possible, he said; the dire warnings came from "negative forces" raising things he insists will not happen. And on Monday, Trump slammed "perfect little angel" Dario over his screed.

House Speaker Mike Johnson made the same argument in institutional form: an emergency session to regulate AI would cost the United States the race with China.

The American Split

The domestic disagreement was already on the record. Barack Obama, at a Manhattan fundraiser on Thursday whose transcript his office released to the New York Times on Sunday, told House Minority Leader Hakeem Jeffries to make AI a governing issue if Democrats take the House. The technology, he said, is "moving very fast in private hands," and he positioned himself as neither an accelerationist nor a doomer.

David Sacks, the former White House AI czar, told Amodei and Altman in a Saturday-night post to go ahead and slow down if their unreleased models warrant it, but without the antitrust waiver, the regulatory approval process or METR, which he called intertwined with Anthropic's investors and staff. Demanding a preferred framework as the price of restraint, he wrote, "will look like blackmail of the public and the political system." China, he added, was "very unlikely to join a global agreement, as you know." Beijing confirmed the point two days later.

Beijing's Own Concerns

The most revealing document out of China this weekend wasn't the response to Dario. State Security Minister Chen Yixin's article, published Sunday in China Cyberspace, the Cyberspace Administration's journal, lists six categories of AI risk and, per Bloomberg, makes no mention of the Anthropic and OpenAI calls to slow down.

The first risk is regime security: "hostile forces" using deepfakes and bot networks to wage "cognitive warfare." The second is cyber offense, and here Chen named Anthropic's Claude Mythos and OpenAI's GPT-5.5-Cyber as systems that sharply raise the efficiency of vulnerability discovery and malware development and threaten China's critical information infrastructure.

Chen describes AI as "a new arena for strategic rivalry among major powers," warns that countries with an AI advantage may invoke national security to impose technology controls and build closed ecosystems, and calls for powers to "resolutely resist hegemonism, technological barriers, and exclusive blocs." None of this means Beijing dismisses loss-of-control risk: its cyberspace regulator has carried an explicit loss-of-control scenario in its safety framework since 2024, and Xi said at the World AI Conference in July that AI should "always remain under human control."

Chen's focus is on who controls the systems, who can weaponize them, and who is denied the hardware to build them, not about whether the frontier should move more slowly.

Then There's Xi

Xi's own contribution came at the BRICS summit in New Delhi on Sunday. China will take the lead in establishing a BRICS AI Open Source Zone to promote cooperation on large language models, AI training and an open AI ecosystem, he said. The logic is to build with the Global South and resist a ruleset written in San Francisco and enforced through American export licenses.

Trump and Xi are due to meet in Washington on September 24, with AI governance expected on the agenda.

Reuters reported that officials were preparing a separate mid-September AI-safety dialogue led by Treasury Secretary Scott Bessent, covering AI-directed cyberattacks, distillation and the prospect of a Chinese model with Mythos-level cyber capabilities; a White House official said "there is currently no planned AI-related meeting in mid-September." Lizzi Lee of the Asia Society Policy Institute framed Beijing's question: if Washington wants cooperation on frontier safety while restricting China's access to frontier compute, "what exactly does that cooperation look like?"

As we laid out Saturday: Chinese open-weight models from DeepSeek, Alibaba's Qwen, Moonshot's Kimi, MiniMax and Zhipu are downloadable, forkable and far cheaper to run, with cumulative downloads the Global Times puts above 10 billion. Export controls and evaluator regimes govern American labs and American hardware. They do not retrieve weights that have already been distributed. Amodei's essay does not pretend otherwise; it is why the proposal climbs to a negotiation with Beijing, and why he ranks a treaty-style limit on recursive self-improvement as "difficult but just on the edge of being possible." That third step requires a partner. As of Monday, the partner has said what it thinks of the first two.