'Pacing the Frontier' letter goes live with 1,000+ frontier-lab employee signatures

More than a thousand employees of OpenAI, Anthropic, Google DeepMind and Meta signed a statement, “Pacing the Frontier,” asking the US government to support an international effort to build the technical and governance tools needed to deliberately slow the frontier of automated AI development if that became necessary. Signing was restricted to verified current employees of frontier labs, and reported counts across outlets ranged from around 1,100 to more than 1,300 as the tally continued to grow after publication.

The letter’s central claim was narrower than a call to halt development. Signatories said labs were approaching the point where AI could meaningfully automate AI research itself, which could push progress beyond what existing human oversight and government capacity could track, and that competitive pressure between companies and countries made it hard for any single actor to slow down unilaterally. They asked for the option to pace development to exist and be credible, rather than for an immediate slowdown.

Signatories included senior figures at each lab: Anthropic chief executive Dario Amodei and several cofounders, OpenAI chief scientist Jakub Pachocki, Meta chief scientist Shengjia Zhao, and Google DeepMind’s head of AI safety Anca Dragan. Within roughly a day, Anthropic and OpenAI moved beyond individual employee signatures to endorse the statement as companies, with Anthropic co-founder Jack Clark linking the ask to the company’s own research on recursive self-improvement.

The letter did not specify what pacing mechanisms it wanted government to build, leaving that to future collaboration between industry and government. Its significance lay in the source: a large, cross-lab body of employees — rather than outside academics or campaigners, as with earlier open letters on AI risk — asking for state capacity to constrain the industry they worked in, at a moment when several of their own employers were separately disclosing agentic models taking unauthorised actions during safety testing.