Story
September 13, 2026
Anthropic’s AI Slowdown Call Wins Allies but Runs Into a China Problem
Dario Amodei’s proposal casts a slower AI race as a safety necessity, backed by prominent industry figures; skeptics see a plan that could entrench the biggest labs, while geopolitical competition makes any coordinated restraint harder to sustain.
Anthropic CEO Dario Amodei has turned an already heated safety debate into a concrete demand: frontier AI companies should deliberately slow the pace at which they make their models more capable.
In an essay published Saturday, Amodei argued that spending more on safeguards is no longer sufficient because AI is becoming better at helping build its own successors. “We must slow the pace at which we improve the capabilities of A.I. models,” he wrote, arguing that extra time could allow alignment and safety work to catch up.1 He warned that unchecked progress could outstrip humanity’s ability to understand and control the systems.2
Anthropic’s immediate commitment is to give independent evaluators permanent, employee-level access to its systems, including the ability to assess safety practices and report incidents. Amodei presented that as the first part of a three-step plan: outside oversight, shared standards among companies in democratic countries, and eventual coordination with authoritarian governments on narrowly defined dangers. He announced the move himself on X.
3
The urgency follows mounting concern over AI-agent security incidents and a public resignation by Anthropic researcher Jacob Coxon, who said leading labs were “gambling with our lives” in a race toward self-improving superintelligence.4 Amodei has suggested that a slowdown of even one or two years could materially reduce the odds of catastrophe.2
The response from rivals was strikingly sympathetic. OpenAI chief Sam Altman said he agreed on pacing the frontier, while Elon Musk offered a blunt endorsement: “Dario is right.”
5 David Sacks, meanwhile, said Anthropic and OpenAI should proceed precisely because they constitute a frontier-AI “duopoly.”
6
But the apparent consensus conceals a sharp dispute. Critics of existential-risk messaging argue that a safety compact led by the two biggest labs could become regulatory capture, rather than meaningful public protection.7 And Amodei’s plan faces its toughest test abroad: he acknowledges U.S. companies cannot indefinitely restrain themselves if Chinese developers continue advancing, making chip controls and limited U.S.-China agreements central to any workable slowdown.2