Anthropic chief executive Dario Amodei spent the weekend proposing three ways to slow down the technology his own company builds: permanent embedded access for independent third-party evaluators at every American AI developer, shared safety standards across democratic countries, and coordination with autocracies, China first among them. Sam Altman of OpenAI and Elon Musk of SpaceX backed the plan quickly. The reception outside the industry was worse. The president's co-chair for science and technology called the implied bargain blackmail, one of the field's most senior academics said the plan has its order of operations backwards, and the former first director of Britain's AI Safety Institute called it too little and too late, and asked for an indefinite international moratorium instead.
The week that produced the plan is the part worth holding onto. Seven days before any of this, AI enthusiasts were cheerfully testing OpenAI's new model, GPT-6 Astra, which was promoted in part as a household tool: booking tennis courts, running fashion companies, ordering takeout. By Wednesday the frontier looked considerably less harmless. Jacob Coxon, a researcher on his way out of Anthropic, said that the people building AI seriously entertain the possibility that the technology destroys humanity by the end of the decade.
Coxon was talking about artificial superintelligence, a version of AI that does not yet exist and would surpass humans across most domains. That distinction usually defuses this kind of warning. This time it did not, and expert statements backing him up pushed the alarm higher.
Then came the turn that did the real damage, because it was about systems that ship today rather than ones that do not exist. Anthropic disclosed that users had been getting around the guardrails on its current models and using them in ways that could help develop biological weapons. The scenarios named included new mosquito-borne viruses and strains of bird flu. Lawmakers in London and Washington responded by demanding limits or bans on the most dangerous lines of AI development.
Amodei's plan arrived into that. Point one: every American AI developer would have to give independent outside evaluators permanent, built-in access to its systems, to check that safety commitments are being kept, report incidents, and watch for new models drifting from their intended goals and from human interests. Today such observers are let in at the companies' own initiative, or given narrow access to investigate something that has already gone wrong. With no independent government regulator in place, Amodei presents continuous outside access as a first step toward one.
Point two: every company in a democratic country building frontier models should agree on common safety standards and set limits on the speed of uncontrolled development. Point three: democratic countries with serious AI capability should coordinate with autocracies, China above all, to keep the race in check. Amodei suggested starting narrow, with an agreement banning obviously dangerous applications such as bioweapons production.
Alongside the plan he made a forecast: keep accelerating capability, and within 6 to 12 months a swarm of AI agents could be in a position to take over the entire internet. Skeptics doubted the timeline. Others treated it as sufficient grounds to declare a red alert.
Washington is not in the mood. Amodei's proposals require the American government to act, and last week Trump said he is not afraid of AI wiping out humanity. On Sunday he repeated it, saying people are discussing something that will not happen. His approach suggests losing the race to China worries him considerably more, and his administration says so out loud: Treasury Secretary Scott Bessent said last week that if China gets to AI first, there is "no next day" — everything else stops mattering.
The people around Trump also cannot see why Amodei and his peers do not simply slow down themselves. David Sacks, co-chair of the President's Council of Advisors on Science and Technology, replied to Amodei's post with the obvious move: the simplest way not to build superintelligence is to agree not to build it. Demanding a preferred regulatory framework as the price of that decision, Sacks argued, would look like blackmailing society and the political system.
Stuart Russell, one of the field's leading figures, attacked the plan from the opposite direction. Amodei wants to slow progress to buy developers time to improve the systems meant to keep AI aligned with human interests. Russell considers that sequence entirely wrong: set the safety requirements first, and let development continue only once they are met. He compared Amodei's approach to a pharmaceutical company shipping a new cancer drug every year in the hope that clinical trials will catch up and come back positive. You specify the safety requirements, Russell said, and then you determine whether you are allowed to proceed.
Sacks and Russell are political opposites and they have found the same hole. Every one of Amodei's three points describes something someone else must do: outside evaluators, other democratic companies, foreign governments. None of them describes something Anthropic would do on Monday morning if nobody else agreed. That is not hypocrisy, exactly — a unilateral pause by one lab is a transfer of the frontier to whoever does not pause, and Amodei's whole argument is that the race is the problem. But it does mean the plan's timing carries the weight that its content does not. Anthropic spent the week being the source of the scariest available information about its own products, and then produced a framework under which that information would flow through channels Anthropic has proposed. The 6-to-12-month swarm forecast does a great deal of work here: it is precise enough to demand emergency measures and vague enough to survive being wrong.
Specialists outside the major labs were suspicious for a simpler reason, which is who was proposing global regulation — the head of the world's most valuable company working on the technology. David Krueger, an AI professor, safety campaigner and the first director of the UK government's AI Safety Institute, called the initiative late and insufficient, and called for an immediate international moratorium on frontier AI development for an indefinite period.
The hardest of the three points gets its first test on September 24, when Trump meets Xi Jinping in Washington. Amodei's coordination proposal would have to travel there in the hands of a president who said on Sunday that the danger is not real, to a counterpart with every reason to read an American call for a slowdown as an American attempt to hold a lead. The plan's authors have named the week's problem accurately. They have handed the solution to the people least interested in it.