A petition titled "Pacing the Frontier" displayed 1,178 verified signatories from frontier AI companies when checked on July 28, including OpenAI Chief Scientist Jakub Pachocki, Anthropic CEO Dario Amodei, and Meta Chief Scientist Shengjia Zhao. It asks the U.S. government to support an international effort to build the technical and governance tools needed to deliberately slow automated AI research. It specifies no trigger, threshold, enforcement body, compute limit, or treatment of open-weight models.
Key facts
- 1,178 signatories displayed on pacingthefrontier.com on July 28; earlier figures of 1,100 and 1,122 are stale snapshots of a live counter.
- All signatories are current employees - the signing form requires a corporate email or proof of employment and has no alumni path.
- Roughly four in five early signers chose to sign publicly under their own name.
- Signature verification was handled by Guidelight AI Standards; organizational support came from Encode AI.
The gap between what this petition says and what it is being reported to say is unusually wide, and the gap is the story.
What it says is compact. Automated AI research - AI systems doing the work of improving AI systems - could accelerate capability faster than societies can understand or control it. Competitive pressure means no single company or country can slow down alone without simply losing to whoever does not. Therefore the United States should help build an international capability to deliberately pace that development.
That is a request for an option, not for a decision. The document asks that a brake be built. It does not ask for one to be pulled, and it does not say who would be allowed to pull it, under what conditions, verified how, or with what consequence for a lab that refused. There is no named threshold, no proposed treaty, no enforcement body, no compute ceiling, and - notably, given who objected loudest - no statement on how any of it would apply to open-weight models that cannot be recalled once released.
The most persistent factual error in coverage concerns who signed. Nearly every writeup described "current and former" employees. The intake form asks "At what frontier AI company do you work?", requires a corporate email address or alternative proof of employment, and lists 18 eligible companies spanning OpenAI, Anthropic, Google, Meta, Mistral, xAI and Thinking Machines. There is no former-employee option. The confusion is understandable - someone can be former-OpenAI and current-Thinking-Machines - but it materially changes the weight of the document. This is not an alumni letter written from outside. It is people signing while still holding the badge.
That is what makes the roster consequential. The publicly featured list includes OpenAI's Jakub Pachocki and Chief Research Officer Mark Chen; Anthropic's Dario Amodei, Co-Founder and Chief Science Officer Jared Kaplan, Jack Clark, Chris Olah and Ben Mann; Meta's Shengjia Zhao and Dawn Song; and Google DeepMind leadership. These are the people who would be personally slowed down.
Two caveats on that roster, both from the site itself. First, the featured list is curated, not a random sample, so its seniority should not be extrapolated to all 1,178. Second, the site states plainly that comments and signatures are personal and do not represent company views. A chief executive signing a personal statement is not a corporate commitment, and treating it as one would be a mistake in both directions.
The organizational backing is disclosed and worth reading carefully. Guidelight AI Standards, which handled verification, describes itself as an independent nonprofit founded by former OpenAI safety leaders Page Hedley and Steven Adler; it says it accepts no AI-company funding while disclosing that Adler retains vested OpenAI equity. Encode AI is an AI-policy advocacy organization. Both are self-descriptions rather than audited disclosures. Verification is also meaningful but not reproducible: the form states that proof-of-identity material is permanently deleted once a statement goes public, so no outside party can recount the signatures.
The 78 personal comments on the site explain why the shared text is so thin. OpenAI's John Schulman frames signing as an act of creating common knowledge that coordination mechanisms may eventually be needed, and wants voluntary lab design before government involvement. Another OpenAI researcher, Leo Gao, describes the situation as a race toward an intelligence explosion. Those are very different beliefs about urgency. A statement broad enough to hold both is necessarily a statement about preparedness rather than about action.
The strongest objection is not that safety does not matter. It is that a pacing regime with no specified mechanism is not yet a policy - and that the version most likely to be implemented first domestically, in the form of licenses, release gates or costly compliance, would advantage exactly the large closed labs whose employees signed it, while burdening smaller competitors and open-weight developers. The LocalLLaMA thread went at that immediately, along with the missing thresholds and the unaddressed China question. The China objection, in fairness, is not something the letter forgot - international coordination is its answer to unilateral defection. The letter simply supplies no mechanism for achieving that coordination.
The most striking corroboration of that criticism came the same day from Sam Altman, who told an interviewer that any approach to pacing must avoid becoming regulatory capture or collusion among frontier labs - naming precisely the failure mode the petition leaves unaddressed.
The accurate summary: a large, organizer-verified coalition of people currently building frontier AI wants the United States to help construct an international ability to slow automated AI research if it ever becomes necessary. Their insider credibility is real. Their proposed brake ships without an operating manual.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)