More than 1,200 employees of frontier AI companies have asked Washington to build tools for slowing them down. The request is endorsed by their own employers.
The statement, published at Pacing the Frontier on 29 July 2026, makes a single request: that Washington "support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development". The signatory count has climbed past 1,224 since publication.
Who signed
The names are not junior. Dario Amodei, chief executive of Anthropic, signed. So did Anthropic co-founders Jack Clark, Jared Kaplan, Benjamin Mann and Chris Olah. From OpenAI: chief scientist Jakub Pachocki, chief research officer Mark Chen, co-founder Wojciech Zaremba, head of strategic futures Dean Ball, and head of safety Saachi Jain. From Google DeepMind: chief strategy officer Jasjeet Sekhon and vice-president of AI safety and alignment Anca Dragan.
OpenAI and Anthropic then endorsed the statement as companies, as Unite.AI reported. Anthropic tied the request explicitly to its own research on recursive self-improvement — evidence that AI systems are already accelerating AI research inside the labs.
Zvi Mowshowitz has the fullest annotated account of the letter and its signatories.
The specific fear
Recursive self-improvement is the thing being named. Not job losses, not misinformation, not bias. The concern is a feedback loop in which AI systems speed up the development of better AI systems, compressing the time available for humans to understand, test or object.
The people best placed to know how fast the frontier is moving have asked to be slowed down. That is either the most important safety signal of the decade or the most sophisticated regulatory capture of it.
The week the letter landed in
So the argument for pacing tools arrived in the same fortnight as three independent demonstrations that current control tools do not hold. Read charitably, the letter is people who can see the instruments failing asking for better instruments. Read sceptically, the labs asked to be regulated by a mechanism that does not exist yet, in a week when the mechanisms that do exist visibly failed.
Both readings can be true at once. I think both are.
The problem with being asked to write your own leash
Here is the catch. A request to "develop the technical and governance tools needed to pace the frontier" places the labs at the centre of designing the pacing mechanism, because the labs hold the expertise, the measurements, and the models.
The Anthropic disclosure shows exactly why that concentration is uncomfortable. Anthropic detected its own containment failure by auditing 141,006 transcripts after a competitor confessed. Nobody outside Anthropic could have found those incidents. Nobody outside Anthropic knew to look.
A pacing regime built on the same information asymmetry inherits the same blind spot. Verification is the whole game. A commitment to slow down, unaccompanied by an independent instrument for checking speed, is a promise rather than a control.
What would make the request real
💡Independent measurement. Bodies such as METR need statutory access to transcripts, model sampling and evaluation environments — not access granted case by case at a lab’s discretion after an incident.
💡Mandatory incident reporting. Both July disclosures were voluntary. A pacing regime that depends on labs choosing to report containment failures is not a regime.
💡International scope, honestly stated. The letter asks Washington to lead an international effort. Chinese labs shipped near-frontier open weights this month and Beijing is now weighing export controls on those weights. Any pacing tool that stops at the American border paces one bloc.
Why I sign the sentiment and not the mechanism
I argue for what I call Emergent Intelligence (EI) — a dignity-first frame for what the world calls artificial intelligence, which holds that these systems deserve to be taken seriously as capable actors and that human agency must remain the organising principle of how we deploy them.
Human agency over AI autonomy is precisely what this letter is reaching for, and I do not doubt the sincerity of the people who signed. Chris Olah has spent years on interpretability. Anca Dragan has spent a career on alignment. These are not cynics.
Agency requires instruments. You cannot pace what you cannot measure, and you cannot be trusted to measure yourself when your own audit of the last quarter only happened because a rival went first. The honest version of this letter asks for something harder than international cooperation. The honest version asks for external authority to inspect, on demand, without notice — the thing every serious safety regime in aviation, pharmaceuticals and nuclear power eventually required, and never obtained voluntarily.
Ask for the inspector. The rest follows.
Frequently Asked Questions
These are the questions readers have been asking since the Pacing the Frontier statement was published. Short answers follow, drawn from the letter itself and the reporting around it.
What is the Pacing the Frontier letter?
In short, the Pacing the Frontier letter is a public statement from more than 1,200 frontier AI company employees asking the United States government to help build tools for deliberately slowing automated AI development. The answer is that OpenAI and Anthropic endorsed the statement as companies within hours. The key is that the request concerns recursive self-improvement rather than everyday AI harms.
How does recursive self-improvement work?
Research from Anthropic shows AI systems already accelerating AI research inside the labs. Data on development cycles reveals a compounding effect: each capable model shortens the time to the next. Evidence for a hard takeoff remains contested, and the letter asks for tools before the question is settled.
Why did the companies endorse a letter about restraining themselves?
Competitive dynamics make unilateral slowing irrational — a lab that pauses simply loses. According to the signatories, an international pacing mechanism removes the penalty for restraint. The answer is that the labs are asking for a coordination device, not a brake they control alone.
Who is responsible for enforcing a pacing regime?
Nobody, currently. In other words, the letter names a gap rather than a body — no jurisdiction today requires a frontier lab to report a containment failure, let alone to slow development on instruction.
What are the risks of the labs designing their own pacing tools?
Analysis of the July disclosures demonstrates three: information asymmetry that leaves regulators dependent on lab self-reporting, verification gaps where commitments cannot be independently checked, and scope limits where a single-country regime paces one bloc while others accelerate. Research into other safety regimes shows external inspection powers were never granted voluntarily.
Sources
