Story
September 16, 2026
AI’s Safety Pivot Runs Into Washington’s China Alarm
AI’s leading labs say a deliberate slowdown is needed before more capable systems outrun human control; their critics see the same agenda as a potential moat for dominant companies and a strategic risk in the contest with China.
The push to “pace” frontier AI gathered force after Anthropic chief executive Dario Amodei argued that commercial competition could turn into “a race to the bottom” as increasingly capable systems create harder-to-control risks. His prescription was not a halt, but time for alignment work, independent scrutiny and common safety standards.1
The most immediate concrete step came from the labs themselves. Anthropic said it would give outside evaluators employee-like access to inspect safety practices and report incidents. OpenAI’s Sam Altman quickly backed the proposal, saying independent evaluators were “a great idea” and that OpenAI would adopt the approach too.
2 Altman separately framed pacing as progress made slower than it otherwise would be—not a stop—and warned both of losing control to AI and of excessive concentration of power.3
That convergence brought unusual allies. Google DeepMind’s Demis Hassabis said Amodei’s direction was right; Elon Musk also endorsed him. Yet the political response exposed the fault line beneath the safety consensus. House Speaker Mike Johnson called for a balance between national security and safe models, arguing that a moratorium could allow China to “overlap us.” President Donald Trump dismissed new guardrails, insisting US leadership in AI must be preserved.3
Other skeptics questioned the messengers as much as the policy. Vice President JD Vance said the industry’s appeal for regulation felt “a little bit” like a Trojan horse. Former White House adviser David Sacks told Amodei and Altman to proceed on their own, noting their firms’ frontier-AI dominance and challenging the claim that a slowdown was purely altruistic.
4 Former Meta AI chief Yann LeCun mocked the renewed alarm, recalling Anthropic’s earlier warnings about open-source models.3
Amodei acknowledges the strategic dilemma: any democratic-world pacing agreement is vulnerable if China refuses to join. China, for its part, has said fearmongering and confrontation would undermine global AI governance.1 The result is a debate in which nearly everyone invokes safety, but no one agrees whether slowing down makes America—or the world—safer.