Anthropic CEO Dario Amodei's call to slow frontier AI pairs safety requirements with restrictions designed to preserve America's lead over China. Beijing's response contests the competitive framing, while its official documents recognize advanced-model risks.
The big change
- What changed: The slowdown debate joins two questions: what makes advanced models safe, and who can obtain the resources to build them.
- Why it matters: Countries seeking access face a different bargain from developers asked to demonstrate safety. Rejecting the first does not establish opposition to the second.
- What to watch: Look for specified tests, verification arrangements and participating institutions, rather than treating shared risk language as an agreement.
Two obligations in one proposal
Amodei's September essay proposes common pre-release testing and reliably verifiable pacing. It also urges restrictions on China's chips, manufacturing equipment and remote compute access.
Proposed measure | Who or what it constrains |
|---|---|
Common model testing | Developers' release decisions |
China-focused supply controls | Access to computing resources |
China's official documents recognize frontier risk
China's Foreign Ministry separated its objection to competitive framing from its safety position. On September 14, spokesperson Guo Jiakun rejected threat narratives, confrontation and competition as a basis for global AI governance. Asked again the next day, he said safeguards against loss of control should improve as the technology advances and pointed to China's newly released safety framework.
That official AI Safety Governance Framework 3.0 says autonomous learning and recursive self-improvement may outrun human anticipation and control. It calls for regular tests of technological loss-of-control risks and international recognition of assessment methods and benchmarks.
A 2026 World Artificial Intelligence Conference statement likewise calls for guardrails at leading AI companies, traceability for agentic systems, tiered risk governance and transparent international standards. The same statement stresses equal participation, open access and inclusive supply chains. These records establish official concern about advanced-model risk. They do not establish Chinese acceptance of Amodei's proposed pace or verification regime.
Different speakers, different authority
Rest of World reports a more varied Chinese debate. Alvin Wang Graylin of the Asia Society Policy Institute told the publication that rejecting the current terms does not rule out a future pacing agreement. Governance researcher Zilan Qian argued that U.S. and Chinese institutions make different risks and voices more prominent, even when many underlying concerns overlap.
The report also cites a personal post by Tao He, head of R&D at MiniMax Code, an essay by DeepSeek engineer Liu Shengyu and comments from Huawei rotating chairman Eric Xu. They are individual or company voices, not statements of Chinese government policy.
Shared vocabulary is not an agreement
China's framework supports international recognition of evaluation methods. That identifies a subject for technical discussion, but the reviewed official responses do not accept a pacing agreement or its verification regime.
