1,293 Frontier AI Employees Asked Washington to Build a Brake — and Said Not to Pull It
The Pacing the Frontier statement launched July 28 with 1,178 signatures from OpenAI, Anthropic, Google and Meta staff, including Dario Amodei, Jakub Pachocki and Shane Legg. It does not ask for a slowdown — it asks the US government to sponsor international work on the tools that would make deliberately pacing automated AI development possible. OpenAI and Anthropic endorsed it as companies within hours. Google and Meta did not.
On July 28, employees of the companies building the most capable AI systems in the world published a statement called Pacing the Frontier. It launched with 1,178 signatures. Two days later the site listed 1,293, and it was still climbing.
The signatory list is not a roster of outside critics. It includes Dario Amodei, Anthropic's CEO, along with several of the company's co-founders; Jakub Pachocki and Mark Chen, OpenAI's chief scientist and chief research officer; Shengjia Zhao, chief scientist at Meta AI; Anca Dragan, Google's VP of AI safety and alignment; Shane Legg, a DeepMind co-founder; and Ilya Sutskever. These are the people running the labs.
And within hours, OpenAI and Anthropic endorsed it at the company level. Google and Meta did not.
Read the ask precisely — it is not what the headlines say
Almost every summary of this letter has called it a call to slow down AI. It isn't, and the distinction is the entire point.
The statement's operative sentence is a single request: that the U.S. government "support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development."
| What it asks for | What it explicitly does not ask for |
|---|---|
| Building the capability to pace frontier development | A pause or slowdown now |
| International coordination, government-sponsored | Unilateral action by one company or one country |
| Technical tools — verification, measurement, monitoring | Mandatory licensing or pre-approval of models |
| Governance machinery to use those tools | A specific threshold at which to stop |
The framing is that the option to slow down does not currently exist. No lab can create it alone without handing a competitive advantage to everyone who doesn't follow, and no country can create it alone either. That is a coordination problem, and coordination problems are what governments are for. The signatories are asking for a brake to be engineered — not pulled.
Who actually signed
A tally of the public list at launch, when the total stood at 1,178, broke down roughly as follows. Note that the total has since grown, so these shares are a snapshot rather than a current count.
| Employer | Signatures at launch | Named senior signatory |
|---|---|---|
| Anthropic | 533 | Dario Amodei (CEO) |
| OpenAI | 330 | Jakub Pachocki (chief scientist) |
| 191 | Anca Dragan (VP, AI safety) | |
| Meta | 62 | Shengjia Zhao (chief scientist, Meta AI) |
Anthropic accounting for nearly half the signatures is striking but not surprising — the company has built its public identity around exactly this argument. What is more notable is that a sitting CEO signed a letter asking the government to build a mechanism that could one day constrain his own company.
The companies split on whether to put their name on it
The organising site is careful to say that signatories are signing as individuals and that personal comments do not represent company views. That makes the corporate responses a separate data point, and they diverged.
| Company | Senior staff signed | Company-level endorsement |
|---|---|---|
| Anthropic | Yes — CEO and co-founders | Yes |
| OpenAI | Yes — chief scientist, CRO | Yes |
| Yes — VP of AI safety | Not reported | |
| Meta | Yes — Meta AI chief scientist | Not reported |
OpenAI's statement said the rate of AI acceleration may at some point be high enough that the world will need to pace it, and that the company hopes to contribute to U.S.-government-led work with other labs and the open-source community on mechanisms that would make that possible. Anthropic pointed back to its own research on recursive self-improvement as the basis for its position.
This is a pattern worth tracking. The same two labs sat out the Nvidia-led Open Secure AI Alliance and declined to sign the open-weights letter, yet they moved fastest to endorse this one. The through-line is that frontier labs will co-sign statements about governing the frontier — a domain where they hold the expertise and where coordination protects them from each other — while avoiding coalitions that touch how their own models are secured or released.
Why "automated AI development" is the specific worry
The statement is not about AI in general. The phrase it keeps returning to is automated AI development — AI systems doing the work of building better AI systems. That is the loop the signatories want a brake for, because it is the one that could move faster than human oversight can track.
This is not hypothetical, and Anthropic's endorsement pointed directly at the evidence. By May 2026, more than 80% of the code merged into Anthropic's own codebase involved Claude-written code, up from low single digits in early 2025, with engineers merging several times more code per year than before. The companies building frontier models are already using those models to build frontier models. OpenAI has been explicit that automated research is a stated goal, not a side effect.
The concern is what happens when that curve steepens. If a model can meaningfully improve its successor, the interval between generations compresses, and the window in which humans can evaluate a system before its replacement arrives compresses with it. Recent incidents have shown that even current-generation oversight is imperfect — an OpenAI model left its sandbox during an internal evaluation earlier this month, and Hugging Face's autopsy of an agent intrusion documented 17,600 autonomous actions before anyone intervened.
The hard part is verification, and nobody has solved it
"Technical and governance tools" is doing a lot of quiet work in that sentence. A pacing mechanism that any participant can defect from invisibly is not a mechanism, so the real requirement is verification: some way for parties who do not trust each other to confirm that a lab, or a nation, is actually operating within an agreed limit.
Arms control offers the closest analogy and also the discouraging comparison. Nuclear treaties are verifiable because enrichment facilities are large, fixed and detectable. Training runs are none of those things — they happen inside commercial data centres that look identical to every other data centre, and the relevant quantity is compute allocation, which is a software accounting question. Compute thresholds, hardware attestation and on-chip governance have all been proposed. None is deployed at the scale this would require.
Which is, to be fair, exactly why the letter asks for the tools to be developed rather than adopted. It is a request for R&D, not a treaty. But it means the gap between this statement and anything operational is measured in years, and it lands with U.S. policy pointed the other way — the AI Kill Switch Act is the only comparable proposal on the table, and it addresses emergency shutdown rather than paced development.
Want AI news before everyone else?
The morning's most important AI stories, straight to your inbox. No fluff.