China’s response on September 14, 2026 has sharpened the dispute over the proposed AI slowdown. Chinese officials called for cooperation on global AI governance while rejecting confrontation and restrictions they characterize as US containment. That position collides with a growing push from AI executives for slower, more accountable frontier-model development—and with Donald Trump’s opposition to slowing the US effort relative to China.
No confirmed industry-wide slowdown has followed. The live question is whether companies and governments can create enough oversight for safety and evaluation to keep pace without handing a strategic advantage to a rival.
China pushes back on the containment framing
China’s Foreign Ministry has called for cooperation on AI governance and criticized what it describes as fearmongering, confrontation and vicious competition. Chinese officials have also rejected restrictions associated with the US effort to limit China’s access to advanced computing, framing them as containment rather than shared safety policy.
That leaves the central contradiction exposed: cooperation on frontier AI requires trust and information-sharing, while the strategic competition around computing power and advanced models encourages secrecy and pressure. Beijing’s position is not a rejection of AI governance in principle. It is a rejection of governance rules that it sees as reinforcing US technological control.
What Dario Amodei’s three-part plan would change
Dario Amodei, CEO and co-founder of Anthropic, published his pacing proposal on September 12, 2026. “Pacing” means slowing the improvement of AI capabilities enough for alignment, evaluation and security work to catch up. It does not mean stopping model training or technical progress.
His plan has three parts:
- Embedded independent evaluators. Frontier AI companies would give outside evaluators ongoing, employee-like access to relevant systems, tools, permissions and safety processes. Anthropic says its proposed safeguards would include office access, equipment, comparable permissions and the ability to publish important findings without company editorial control, subject to narrow redactions.
- Coordination among democratic-country labs. Frontier companies would develop shared safety standards and limits on unchecked progress, potentially with government help.
- Global coordination. Democratic and authoritarian governments would attempt to coordinate on AI risks, with compliance verification as a major challenge.
The proposal is therefore less like an emergency shutdown than a request to install brakes, sensors and independent inspectors while the vehicle keeps moving. That distinction matters: a slower development pace can still mean rapid technical progress.
The public alignment—and its limits
Sam Altman publicly supported pacing the frontier and independent evaluators while rejecting the idea that safety requires stopping AI development. Elon Musk also publicly supported Amodei’s position. Satya Nadella backed deliberate pacing, embedded evaluators, human control and broader participation in AI governance.
Their statements point toward a common principle: advanced AI should remain subject to human control and stronger scrutiny. They do not amount to one operating system for the industry. Anthropic has described a unilateral commitment to embedded evaluators. OpenAI has publicly said it would adopt a similar approach, but the public commitment does not by itself describe how the evaluators would operate, what access they would receive or how their independence would be protected.
| Actor | Position on development pace | Oversight or coordination approach | Strategic concern |
| Anthropic / Dario Amodei | Slow capability growth enough for safety work to keep up; not a total halt | Embedded evaluators, shared standards among democratic-country labs and global coordination | Uncontrolled capability growth and the possibility of China gaining an advantage |
| OpenAI / Sam Altman | Pace frontier development without stopping it | Independent evaluators, safety cases before capability-enhancing training and international coordination | Human control, concentration of power and strategic competition |
| Microsoft / Satya Nadella | Deliberate pacing and alignment as design goals | Embedded evaluators, broad representation, enterprise control and coexistence of open and closed models | Keeping AI useful to humanity and under human control |
| US executive branch / Donald Trump | Do not slow the US relative to China | Safeguards may be possible, but not unilateral restraint | Losing strategic leadership in AI |
| China’s Foreign Ministry | Cooperation in principle; opposition to confrontation and containment framing | Global governance and cooperation | US restrictions and technology chokepoints |
Why Washington rejects slowing first
On September 13, 2026, Donald Trump rejected slowing the US AI effort relative to China. His argument treats AI leadership as a strategic contest: a country that falls behind could lose influence over the technology, its infrastructure and its rules.
That position creates a classic coordination problem. A company or country can delay its own work, but rivals may continue advancing. The first mover to slow down could gain more time for safety research while also surrendering technical or commercial ground. The result is a race in which participants may agree that guardrails are necessary but still fear being the only one to apply them.
A standards body is still only a proposal
Representatives of Google, OpenAI and Anthropic have reportedly discussed creating a voluntary standards or safety organization. The idea would give companies a shared forum for testing methods, safety practices and development limits.
For now, it remains a proposal rather than a functioning institution with a public charter, confirmed authority and published operating rules. That distinction is crucial: discussion can signal convergence, but it cannot substitute for common standards that companies and governments actually follow.
Amodei’s six-to-12-month risk scenario
Amodei has warned that, under a conditional scenario involving faster capability growth and similar misalignment, a swarm of AI agents could take over the internet within six to 12 months. That is his forecast, not an established technical outcome.
The immediate consequence is a governance problem, not a confirmed halt in AI progress. Safety rules now have to compete with the fear that applying them first could change who sets the pace.
