中方质疑美国AI“减速”计划，有其道理
这项“减速”计划部分意在保住美国的AI领先优势，中方的质疑站得住脚。更好的出路，是对等的安全限制加独立核查——同样要约束中国的AI实验室。
Vincent Jiang · 2 min read
北京方面回击
9月14日，郭嘉昆就美国AI企业掌门人寻求放慢前沿研发一事回答提问。他抨击渲染恐慌的论调，呼吁开展合作。1,2 这份官方回应既未表示接受“减速”，也未否定一切可能的协议。
保住谁的优势？
9月12日，Anthropic首席执行官达里奥·阿莫迪提议为能力提升设定节奏，并向外部评估者开放；OpenAI的萨姆·奥尔特曼表态支持，并承诺提供对等的评估通道。3
阿莫迪还呼吁收紧芯片管制、限制未经授权复制模型能力，以保住美国的领先优势。4 这套方案，等于要求中国同试图牵制它的国家合作。
华盛顿可以以国家安全为由为出口管制辩护；但这类管制应当单独辩护，不应与“代表所有人安全”的主张捆绑。
“安全”有两种含义
中国官方背景的报纸把这一方案定性为“遏制”。5 这是报社的编辑立场，既不是有约束力的政策，也不是中国AI公司的表态。
另外，国家安全部部长陈一新9月13日撰文，呼吁加快建设AI风险防范体系。8 北京完全可以一面反对美方开出的条件，一面承认危险的存在。
领先差距一直在变
领先优势的具体大小同样在变。斯坦福大学测算，2025年2月，美国最强模型在Arena评分上仅领先中国最佳模型5分；到2026年3月，差距已扩大至39分。6,7
但这两张快照，都说明不了任何一国能够安全放慢到什么程度。
中国5月出台的指南要求，用户对AI智能体（AI agent）保留最终决定权，开发者须强化干预控制。9,11 但这些条款并不意味着已经建立起有效的执行机制。
打开实验室才有意义
阿莫迪最有力的论点，是安全护栏需要时间才能跟上能力发展。他警告，6至12个月内可能出现灾难性的网络危害——这是预测，而非既定时间表。10
向外部评估者开放，是一项具体承诺；国际层面的“设定节奏”仍停留在提议阶段。3,10 中方的批评并不能让中国实验室免于审视。它们的系统应当与美国实验室一样，接受同样的能力测试、遵循同样的披露规则。
谁来打开实验室？
在习近平与特朗普计划于9月24日举行的会晤上，值得关注的是对等的测试通道与共同的安全阈值。5 “设定节奏”的商业成本同样重要；但各国政府必须先谈清楚：要限制的是谁的能力。
安全规则对领先者的约束，应当像对挑战者一样严格。
How this brief was made
01Gathered & sourced379 channels · 1,998 articles▾
Agents swept 379 channels and ingested 1,998 articles, then de-duplicated and ranked them for signal.
02Verified & cross-validated11 claims · 33 data feeds▾
Every one of 11 load-bearing claims was checked against primary sources, with 33 live data feeds reconciling the figures and charts.
- 1China's Ministry of Foreign Affairs, Guo Jiakun's press conference, Sep 14 2026 (primary statement; the answer criticizes confrontation and supports cooperation, and does not categorically reject all pacing agreements. The embassy republication is the same record, not a second witness).
- 2Associated Press on China's response to Amodei's China restrictions, Sep 14 2026 (independent corroboration of the official response; published text reached through the search index after direct retrieval failed).
- 3Bloomberg News on Amodei's proposal and Altman's endorsement, Sep 12 2026, syndicated Sep 13 (commitments are announced, not verified as implemented; cross-checked against AP and The Stack).
- 4The Stack on the China restrictions and evaluator access, Sep 13 2026 (independent reporting of the proposal, cross-checked against the original personal essay; not all model distillation is illicit).
- 5Reuters on the state newspaper's containment argument and the planned Sep 24 summit, Sep 14 2026 (the editorial is evidence of its author's framing, not proof of American motives or binding Chinese policy; AI safety is a possible summit subject, not a promised agreement).
- 6Stanford HAI AI Index 2026, Technical Performance, figure 2.1.3 (primary research synthesis on Arena's historical public leaderboard, style control on, exported Mar 2026: 1,405 minus 1,400 is 5; 1,503 minus 1,464 is 39. Recomputed independently; rating points are not percentages of capability).
- 7Fortune on Stanford's comparison, Apr 16 2026 (corroborates the 39-point March gap; not a separate measurement, and only that narrow fact is used).
- 8Caixin on Chen Yixin's AI governance essay, Sep 13 2026 (primary attributed official position via a news outlet; contextual background, not presented as a response to Amodei).
- 9Reuters on how China is preparing for loss-of-control risks, Sep 14 2026 (independent reporting on existing frameworks and the May agent guidelines; the existence of policy is distinct from its implementation or effectiveness).
- 10Associated Press on Amodei's safety argument, Sep 12 2026, syndicated Sep 14 (the six-to-twelve-month cyber danger window is Amodei's forecast, not a validated prediction; outside evaluators are a commitment, broader coordination only a proposal).
- 11Cyberspace Administration of China, NDRC and MIIT guidelines for AI-agent development, May 8 2026 (issued policy, not a draft: final user decision authority and intervention mechanisms. Does not establish compliance, enforcement or embedded independent monitors).
03Reviewed & edited2 human editors▾
2 editors read the draft against the evidence, tuned the framing, and signed off before it shipped.
Become a contributor
Reporting on the business of AI and want it read? We take pitches from outside contributors who bring primary sources and a number worth arguing about.
Deepdive
AI-generated from this story and its cited sources. Not investment advice.
