The Only Slowdown the AI Labs Want Is One Washington Imposes
For six weeks the AI labs warned the world to slow down, and every slowdown they offered was one only a government could impose — even on themselves.
OpenAI is targeting full research automation by March 2028, and Anthropic reports Claude now writes 80 percent of its own code [1]. Neither clock is slowing. Yet in the same stretch, the people running both companies spent six weeks telling governments the frontier must be paced — and never once volunteered to go first. On what the evidence actually demands, Anthropic co-founder Jack Clark was careful.
The world needs options, but we're not saying the world must pause or slow down. That's not what the evidence says. — Jack Clark
The same month, OpenAI's chief scientist Jakub Pachocki reached for the opposite.
We actually believe this should be slowed down … We need some sort of international norm to be able to control this. — Jakub Pachocki
Two labs, two clocks, one shared alarm. Over the weeks that followed, every slowdown put on the table turned out to be the same kind: one a government compels, and applies to everyone at once. In August, 1,367 researchers from the labs themselves — OpenAI, Anthropic and DeepMind staff among them — urged Washington to lead an international effort to pace automated AI through licensing and registration. The headline was a willingness to halt. The fine print was a clause: Amodei and DeepMind's Demis Hassabis signaled they would stop only if competitors agreed to the same [2]. A halt that binds everyone at once is not a unilateral slowdown; it is a request for one. On September 9, OpenAI took a framework to Congress: mandatory national safety rules, audits under parameters the labs would set themselves, evaluators the labs would choose, and a push for state laws in California, New York and Illinois to assemble a national baseline [3]. Every rule in it was mandatory. None of it bound OpenAI unless Washington made it bind everyone. Hassabis, meanwhile, floated a self-regulatory body on the FINRA model — an industry-run overseer — and OpenAI and Anthropic backed that too [4]. On September 17, Sam Altman said he agrees with Dario Amodei that the frontier should be paced [4]. Amodei, the same day, said the quiet part aloud.
Some forms of coordination that would be impactful for pacing are legally challenging, and will require government support. — Dario Amodei
In Amodei's telling, the labs cannot simply agree among themselves to go slow; that is the coordination he names as legally challenging. So the mechanism he proposed is explicit about it. The most effective pacing, he said, is regulation aimed at every U.S. frontier company.
The most effective method of pacing is via regulation that targets all U.S. frontier AI companies, as that covers even those who are unwilling to cooperate voluntarily. — Dario Amodei
The state, in other words, abolishes the first mover — for everyone, including the labs making the ask. Read cynically, all of this looks like moat-building, and the record has material for that read. It has a cost line too. Anthropic paid for its safety stance: after Amodei refused to let Claude be used for domestic mass surveillance or autonomous weapons, the Pentagon designated the company a supply-chain risk, a federal appeals court upheld the ruling, and Claude was pulled from military systems [5]. Even the nationalization offer — the single ask that would cost the asker everything — still routes through the state. Amodei said at that same event he would work with a government that took a 50 percent stake, if safety stayed the priority [4]. The outside world has read the same record. Florida's attorney general went to court this month to stop OpenAI from building new models without outside oversight.
Defendants claim they cannot stop barreling forward with their potentially civilization-ending endeavors unless they are forced to do so by the government. — James Uthmeier
And the party being petitioned has, for now, declined.
Whoever wins A.I. wins! — Donald Trump
The only commitment in these six weeks that carries a date is the automation target [1]. The ask — however many signatures and Security Council briefings it gathers — still has no bill, no sponsor, no signature. The frontier's pace is being set by the only parties who ever volunteered a number: the labs themselves, running clocks they keep telling someone else to stop.
