Story
September 16, 2026
AI’s Safety Consensus Collides With a Simple Challenge: Slow Down Yourself
Anthropic argues that rapidly advancing AI needs time, outside scrutiny and international guardrails before safety failures outrun human control. Its supporters share the concern, while critics say a call for industry-wide restraint cannot substitute for frontier labs choosing to restrain themselves.
The latest clash over AI safety began with an unusually public rupture inside Anthropic. Researcher Jacob Coxon resigned, accusing leading labs of racing toward self-improving systems while “gambling with our lives”; other Anthropic staff publicly amplified the warning.1
Days later, Anthropic chief executive Dario Amodei escalated the argument from internal alarm to a concrete proposal. He said AI capabilities were advancing too fast for safeguards to keep pace, warning that increasingly capable systems could soon enable destructive autonomous-agent behavior. “We must slow the pace at which we improve the capabilities of AI models,” he wrote.2
Amodei’s first move was deliberately unilateral: Anthropic would give independent evaluators permanent, employee-level access, allowing them to inspect safety practices and report incidents. His broader plan calls for shared limits among frontier labs in democratic countries and, however difficult, coordination with authoritarian governments on narrow dangers such as AI-enabled biological weapons.3 On X, Amodei presented it as a three-part plan and said Anthropic was already committing to the first step.
4
The message found rare allies. Elon Musk’s response was blunt: “Dario is right.”
5 Sam Altman likewise argued that losing control of the future to AI was unacceptable and that increasingly capable systems must serve people.
6
But the agreement fractured over who must act first. David Sacks noted that Anthropic and OpenAI effectively dominate the frontier, telling Amodei and Altman: “go ahead. You guys are the frontier.”
7 His sharper formulation was that companies should make products safe “without demanding any concession from the government.”8
That is the brutal test beneath the apparent consensus: safety advocates want collective rules to prevent a race no one can safely lose; their critics see a simpler standard. If the danger is as immediate as advertised, the labs need no permission to slow down.