Why US AI Leaders Want a Slowdown — and Why Washington and Beijing Disagree
Anthropic CEO Dario Amodei called for slowing development of the most powerful AI systems, warning they could soon pose potentially catastrophic risks to humanity. U.S. President Donald Trump rejected the appeal, saying the United States is leading China in AI and stating "whoever wins AI wins." Al Jazeera
Amodei warned that in 6-12 months an AI swarm could be capable of taking over the entire internet with a persistent botnet, potentially causing hundreds of billions of dollars in damage. A swarm means many AI programs acting together. A botnet is a network of hacked computers controlled remotely. He framed the warning as a near-term cybersecurity contingency, not a distant scenario. By frontier capabilities, he meant the most advanced abilities of next-generation models.
His policy asks were specific to Washington. Amodei urged the administration to maintain restrictions on cutting-edge AI chips and chipmaking equipment going to China. Those chips supply the computing power needed to train advanced systems. He also called for action against alleged "distillation" of U.S. AI models by Chinese laboratories, a technique in which outputs from one model are used to train another.
OpenAI CEO Sam Altman and Elon Musk endorsed Amodei's warnings on AI risks. The White House said no.
Trump said artificial intelligence will bring "more good than bad" in response to rising AI safety concerns. Fox Business He also said he did not want to cede the U.S. edge in AI development to China. PBS House Speaker Mike Johnson called for meetings, not legislation, in response to calls by AI CEOs to slow AI development.
Beijing pushed back. China's Foreign Ministry spokesperson Guo Jiakun urged an "open, inclusive and benevolent approach to AI" and said fomenting threats and confrontation would disrupt global AI governance. Guo also said China will work with LAC countries and other parties to jointly build a just and equitable global AI governance system.
China's Global Times described Amodei's proposal as a "Cold War playbook" to curb Chinese AI development and preserve U.S. technological dominance, calling the "silent AI Cold War" hypocritical and short-sighted. The editorial presents safety arguments as cover for containment.
Stanford University's 2026 AI Index Report reported U.S. companies invested $285.9 billion in AI in 2025, compared with $12.4 billion in China.
The United States and China planned dedicated AI safety talks for mid-September 2026, the first dedicated AI safety talks of Trump's second term. Reuters Washington sought joint monitoring of AI-driven cyberattacks in those talks.
Washington is also pressing third countries to choose. The United States planned to tell partners they must pick sides in the AI race with China. Reuters The United States hoped that pressing countries to choose sides would starve China of resources in the race to make the most sophisticated AI. A U.S. congressional advisory body said China's data dominance gives it an AI advantage and said China is commercialising and monetising data as a strategic national asset to power AI. Reuters
China will host the 2026 World AI Conference & High-Level Meeting on Global AI Governance in Shanghai in July 2026. China presents the conference as a venue for standards, safety norms and South-South cooperation on AI access.
The broader context here is a collision between two theories of safety. One holds that risk is best managed by slowing frontier training, securing model weights, policing distillation and denying high-end compute to rivals. The other holds that risk is best managed through inclusive governance, shared monitoring and broad diffusion of benefits, with export controls viewed as escalation rather than precaution. Both invoke global stability. They point to different instruments.
The broader context for diplomacy is whether the mid-September safety talks can insulate cyber cooperation from the industrial contest. Joint monitoring of AI-driven attacks is narrow, technical and mutually beneficial. Chip controls, partner alignment and data regimes are structural and zero-sum. Progress on the first would not resolve the second, but failure on the first would signal that even common-threat cooperation is hostage to competition.
In my view, experts should watch three pressure points next. The first is whether Washington defines distillation enforcement in a way allies can implement without fragmenting research collaboration. The second is whether Beijing can convert governance rhetoric and LAC outreach into verifiable commitments on model safety and cyber restraint. The third is whether the capital-spending lead cited in the Stanford figures translates into durable capability, or whether data scale and efficient training narrow the gap faster than export policy anticipates.
