Why 1,293 frontier AI employees, Dario Amodei included, asked Washington for tools to pace AI development
More than 1,290 employees of the world’s frontier AI companies have asked the United States government to support an international effort to build the technical and governance tools needed to pace automated AI development. The statement, published in July 2026 under the title Pacing the Frontier, carried 1,293 verified signatures and remains open to employees of frontier AI companies who can prove their employment. Dario Amodei, Chief Executive Officer of Anthropic, signed it alongside co-founders Jared Kaplan, Jack Clark, Chris Olah and Benjamin Mann, with Jakub Pachocki and Mark Chen of OpenAI, Shane Legg of Google DeepMind, Ilya Sutskever of Safe Superintelligence and Shengjia Zhao of Meta AI also on the list.
An industry asking a government for the means to restrain it has almost no precedent in modern technology, and the people asking are the ones with the most to lose from a slower clock. Anthropic endorsed the petition at company level, saying its chief executive had signed it, several co-founders and senior staff, and pointing to its own recursive self-improvement research as pointing to the need for “tools to deliberately pace the frontier of AI development so society can prepare”. OpenAI said it believes acceleration may reach a point at which the world will need to pace the rate of AI advancement, and that it hopes to contribute to US-government-led work alongside other laboratories and the open-source community.
Warnings about AI have usually come from outside the laboratories, which allowed the laboratories to answer that the critics could not see the systems. That defence no longer holds, because the sharpest numbers in this debate are now published by the companies themselves. Anthropic’s June research reported that more than 80% of the code merged into its codebase as of May 2026 was written by Claude, up from low single digits before Claude Code reached research preview in February 2025, with the typical engineer merging eight times as much code per day in the second quarter of 2026 as in 2024. On the most open-ended engineering problems, where the engineer does not know what a correct answer looks like, Claude’s success rate reached 76% in May 2026, an increase of 50 percentage points in six months.
The company also reported movement on research judgement, the capability that separates a fast assistant from a system able to design its own successor. Asked to choose the next step in real research sessions, Anthropic’s November 2025 model beat the human choice 51% of the time, rising to 64% with Mythos Preview in April 2026. Independent measurement points the same way: METR found that the length of tasks models can reliably complete alone has been doubling roughly every four months, up from every seven, and that Mythos Preview could work for at least 16 hours. Anthropic stated in the same research that it would be good for the world to have the “option to slow or temporarily pause frontier AI development” so that societal structures and alignment research can keep pace with the technology.
A brake only works when every driver can see the others braking
What the signatories requested is a capability rather than an immediate halt, and the engineering of that capability is the whole difficulty. The statement argues that industry, government and society may need the option to buy time to address emerging risks, develop security measures and strengthen oversight, and that no company or country will surrender acceleration unilaterally while the world “lacks the technical and governance tools to deliberately pace frontier-wide progress”. Anthropic set out why verification carries the weight here, noting that training runs are far easier to conceal than missile silos, that their inputs are general-purpose, and that whoever continued while others paused could inherit the lead.
Solving that verification problem would leave behind something durable, because instruments built to confirm what a laboratory is training also confirm what it is testing, securing and releasing. Several signatories argued for building those instruments before any statute compels it. John Schulman, Chief Scientist at Thinking Machines, said he signed to establish common knowledge about coordination and added that “I would also like to see labs start designing these mechanisms voluntarily”. Will Yager of Anthropic compared the moment to the splitting of the atom and said that “the world should take the time to do this right”.
Borrowed time is what converts raw capability into public benefit
The most persuasive argument for pacing is an argument about gains rather than about fear, and the clearest example comes from Anthropic’s most restricted model. In the first weeks of Project Glasswing, Mythos Preview found more than 10,000 high and critical severity software vulnerabilities across systems of global importance, enough that the constraint in cyber defence has moved from finding vulnerabilities to patching them quickly enough. A capability of that magnitude protects hospitals, payment systems and power grids only if the institutions around it can absorb what it produces, which is a question of human capacity and public coordination rather than model quality.
Every domain where AI promises the most carries the same dependency. Clinical evidence accrues over years of use, regulators need time to write standards they can defend, and schools and employers need time to rebuild what they teach and how they hire. Dawn Song, Vice President of AI Research at Meta, made the point directly in her comment on the petition, saying that if AI proves as transformational as electricity, “we had better be intentional about how it is introduced to the world”. Anthropic’s own research reached a similar conclusion in different terms, observing that more intelligence cannot learn what a drug does over decades of use and cannot hold elections sooner than a constitution dictates. Pacing, on this reading, is how the science, health and security dividends get delivered rather than lost in a transition nobody had the capacity to govern.
Commercial context deserves recording. Anthropic’s call for a pause option came days after it confidentially filed for an initial public offering, not long after a funding round valuing it at close to $1 trillion, and shortly after it declined to release Mythos publicly because the model was too capable of finding software vulnerabilities. Christian Catalini, writing in Forbes, argued that treaties cannot be enforced and that laboratories which believe their own warnings should fund verification infrastructure rather than delegate the problem to diplomacy. The request also goes to Washington and names no European institution, three years after the EU AI Act became Europe’s principal governance instrument.
Those criticisms describe work still to be done rather than reasons to abandon the attempt, and the second of them is close to what the signatories themselves asked for. Amodei has already argued for statutory rather than voluntary constraint, writing in his essay Policy on the AI Exponential that models have moved in four years from barely producing “a coherent line of code” to writing most of the code at major AI companies, and proposing that frontier models undergo technical testing and auditing on the model of aviation regulation, with release blocked or reversed where safety standards are not met. Anthropic published a legislative proposal on frontier model testing and a policy framework for job displacement alongside that essay. The company has said it will convene policymakers, researchers, civil society and rival laboratories over the coming months to establish what a credible slowdown would require, and has committed to publishing the results. A generation of technology companies insisted that governance would come after the fact. This one has asked for the instruments in advance, on the record, with its own numbers as the evidence.
