More than 1,300 technology employees and executives have signed an open letter urging the U.S. government to support an international effort to control the pace of advanced AI development. The signatories argue that increasingly capable systems could accelerate AI research faster than governments, companies and safety teams can establish effective safeguards.
The letter focuses on the prospect of automating substantial parts of AI research. Frontier models are already used to write and optimize code, analyze experiments and assist with technical work. If those capabilities improve enough, AI developers could use models to shorten the cycle for designing and training their successors, potentially causing capabilities to advance more quickly than expected.
Among the named signatories are Anthropic CEO Dario Amodei, OpenAI Chief Scientist Jakub Pachocki, Meta AI Chief Scientist Shengjia Zhao and Google DeepMind Chief AGI Scientist Shane Legg. Employees working in model training, alignment, evaluations, software engineering and preparedness also added their names and individual statements.
The appeal does not propose a specific licensing regime, computing threshold or mandatory pause. Instead, it asks Washington to back the creation of international technical and governance mechanisms that could deliberately slow frontier development if emerging risks require more time to address. Such mechanisms would build on existing efforts to monitor the release and capabilities of advanced models.
“There is a real risk that capability development rapidly accelerates beyond our ability to understand or control the resulting systems,” the letter says.
A central concern is that no individual laboratory or country has a strong incentive to slow down on its own. Developers face commercial pressure from rivals, while governments increasingly view advanced AI as a strategic technology. The signatories contend that coordinated measures are therefore necessary if policymakers want to preserve the option of reducing the pace across the frontier rather than asking one participant to accept a unilateral disadvantage.
The proposal also highlights the difference between monitoring AI systems and having the ability to respond to evidence of danger. Evaluations can test models for capabilities such as cyber operations, autonomous planning or assistance with hazardous activities. Monitoring alone, however, does not create an agreed process for delaying deployment, restricting access or allowing safety work to catch up when models cross a concerning threshold.
Several signatories described that gap from the perspective of people building the systems. Anthropic alignment lead Ethan Perez said safety teams are repeatedly racing to address new risks and may eventually encounter problems that need more time. Nicholas Joseph, who leads pretraining at Anthropic, pointed to models taking on a growing share of infrastructure, optimization and experimental work. Other contributors warned that international competition could reduce the margin available for responding to misuse or failures of model control.
The letter arrives as AI governance remains divided across jurisdictions and levels of government. Policymakers have pursued rules covering deepfakes, chatbot safety, children’s use of digital services, healthcare applications, consumer protections and oversight of frontier models. Broader proposals aimed at the most capable systems have proved more contentious, particularly when they could affect research speed or impose obligations based on computing resources and model capabilities.
The request leaves major implementation questions unresolved. An international pacing framework would need credible ways to identify relevant development activity, assess when intervention is warranted and verify compliance across companies and countries. It would also have to distinguish ordinary AI research from work capable of materially accelerating the frontier. The signatories are asking the government to begin developing those tools before automated AI research makes the underlying coordination problem more urgent.
Sources: Meta AI