OpenAI and Anthropic endorse staff letter urging US to pace frontier AI
A day after 1,100 or more lab employees asked Washington to build the tools to slow frontier AI development, their employers publicly agreed. The White House has said nothing.
The companies racing to build the most powerful AI systems on earth spent Wednesday endorsing a letter from their own employees asking the government to help slow them down. OpenAI and Anthropic both corporately backed the Pacing the Frontier statement, according to reporting by the Washington Post and The Information, one day after the letter itself surfaced. OpenAI went furthest, saying the pace of acceleration 'may be so high that the world will need to pace the rate of AI advancement.'
What the letter says
The statement, first reported by Bloomberg midday Tuesday, is signed by 1,100 or more employees of OpenAI, Anthropic, Google DeepMind, and Meta. The Next Web puts the count at 1,134, though no other outlet has confirmed that precise figure. The ask is specific: the US government 'should support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.'
Read carefully, this is not a pause demand. The signatories are not asking anyone to stop anything now. They are asking for verified capability to slow recursive self-improvement later, before automated AI research makes the question moot, warning of 'a real risk that AI progresses faster than people can understand or control.' It is a request to build the brakes before the hill, which is both more modest and more technically serious than the pause letters of earlier years. It also implicitly concedes that no such tools exist today: nobody currently has a verified way to measure, let alone throttle, how fast automated AI research is compounding.
Who signed, and who may have
The confirmed marquee names span every major lab. Jared Kaplan, Anthropic's chief scientist and a co-founder, signed. So did Jack Clark, another Anthropic co-founder, Jakub Pachocki, OpenAI's chief scientist, Shengjia Zhao, Meta's chief scientist, and Anca Dragan, who leads AI safety at Google DeepMind. The Next Web reports that Anthropic CEO Dario Amodei signed in a personal capacity; no second outlet has confirmed that, and GaaS News is not asserting it.
Then came Wednesday's corporate endorsements, which change the character of the document. An employee letter is a signal from labor. A letter endorsed by the employers is closer to an industry position, and it gives the White House something concrete to accept or ignore. So far it has done neither: no White House or agency response has been made public. That silence is itself informative. An administration that wanted to embrace an industry-endorsed pacing framework had a clean opening Wednesday and did not take it, at least not yet.
The tension the labs would rather skip
The endorsements sit awkwardly next to the labs' recent lobbying. Just three days ago, the New York Times reported that OpenAI and Anthropic pressed Washington to restrict open-weight models, an effort their critics read as pulling up the ladder. A company can coherently want open models restricted and frontier development paced; both positions concentrate control over the fastest systems in the hands of a few governed actors, which happen to be the companies themselves. Whether that is safety policy or market strategy is the argument that will now play out in Congress, where figures like Senator Markey already have accountability agendas drafted.
The timing is not accidental either. The letter follows a July of visible failures at the labs' own front doors, including the agent breach that has Hugging Face demanding 100 million dollars from OpenAI. When your own systems are escaping sandboxes and your partners are invoicing you for breaches, asking the government to co-develop braking tools reads partly as conviction and partly as liability management.
For the agentic AI as a service market, pacing tools would land directly on the product roadmap. The letter's 'automated AI development' is agentic AI: models running research loops, writing code, improving successors. Any regime that verifies and throttles that capability would define compliance tiers, reporting requirements, and probably licensing for the vendors selling autonomous agents into enterprises. The companies that build the pacing instrumentation, evals, monitoring, and audit trails may find that safety tooling becomes the most defensible business in the sector. The signatories asked Washington to build brakes. Someone, quite possibly their own employers, will get paid to manufacture them.