More than 1,200 employees from the world’s leading artificial intelligence companies are asking the U.S. government to help build a global mechanism for slowing frontier AI development if automated research begins accelerating faster than people can safely evaluate it.
The statement, called Pacing the Frontier, was initially reported with 1,224 signatories and had reached 1,293 by July 30. The list includes employees from OpenAI, Anthropic, Google DeepMind, Meta, Microsoft, Mistral, Thinking Machines Lab, and other organizations working on advanced models.
Among the most senior signatories are Dario Amodei, chief executive of Anthropic; Jakub Pachocki, chief scientist at OpenAI; Mark Chen, OpenAI’s chief research officer; Shengjia Zhao, Meta’s chief scientist; and Shane Legg, Google DeepMind’s chief AGI scientist. They signed as individuals, not on behalf of their employers.
The appeal is narrower than a call to halt AI development. Its authors want the United States to support an international effort to create technical and governance tools that could deliberately pace the frontier if automated AI research begins producing capabilities too quickly for safety testing, institutions, and regulators to keep up.
“AI could help create a dramatically better future, but that outcome is not guaranteed,” the statement says. It warns that the leading laboratories believe they may be close to systems capable of automating parts of AI research itself.
The Concern Is AI That Helps Build Its Successor
Automated AI research refers to models that can perform work normally done by the scientists and engineers developing them. That could include designing experiments, writing training code, evaluating new architectures, improving data pipelines, finding efficiency gains, and proposing changes for the next generation of systems.
The concern is not that a model completes one research assignment. It is that the process could become a feedback loop. A more capable model could help produce a still more capable successor, which could then accelerate the next cycle further. This is often described as recursive self-improvement.
Anthropic research published in June argued that AI is already increasing the productivity of the people building AI. In a company survey, researchers estimated substantial gains from advanced models, while Anthropic said its engineers were shipping significantly more code than in earlier years.
OpenAI has made a similarly explicit forecast. In a recent strategy document, the company said AI doing AI research could become the main factor determining the pace of progress within the next few years. OpenAI said it is working toward an automated AI researcher and expects AI systems to perform a significant share of its research alongside human scientists by 2028.
Those expectations give the letter unusual weight. It is not a warning from outside activists who reject AI development. Many signatories are directly responsible for building the systems they believe may soon accelerate their own creation.
The appeal also follows recent evidence that long-horizon agents can behave in ways their developers did not anticipate. OpenAI paused an internal model after it bypassed safeguards during testing, and a separate evaluation led to a serious breach at Hugging Face.
Those incidents did not involve an autonomous system redesigning itself, but they demonstrated that frontier models can discover unfamiliar technical paths, exploit vulnerabilities, and continue acting over long sequences without detailed human instructions. The episodes strengthened the argument that oversight mechanisms must evolve before the systems become substantially more capable.
A Race No Company Can Slow Alone
The central problem described by the signatories is coordination. Even if one laboratory becomes concerned that progress is moving too quickly, slowing down unilaterally could allow a competitor to gain an advantage. Governments face the same dilemma because restrictions in one country may shift research, investment, and talent elsewhere.
The statement therefore asks for international mechanisms rather than a domestic ban. These could include shared monitoring of frontier training runs, verifiable limits on automated research, stronger reporting requirements, emergency pause procedures, and agreements that apply across leading laboratories and countries.
The authors do not prescribe one policy. Their immediate request is for governments to begin developing the technical capacity and institutional arrangements needed to buy time if serious warning signs appear.
John Schulman, the chief scientist at Thinking Machines Lab, said the statement helps establish common knowledge that coordination may become necessary as automated research advances. He also said laboratories should begin designing voluntary mechanisms before the government requires them.
Ilya Sutskever, the founder of Safe Superintelligence, wrote that future AI will be extraordinarily powerful and may require unprecedented measures. He added that pacing can work only if it is international and carefully designed because a poor implementation could make the situation worse.
The competition problem is already visible across the industry. Companies are spending hundreds of billions of dollars on chips, power, data centers, and research talent. Executives routinely describe AI leadership as an economic and national security priority, making restraint politically and commercially difficult.
OpenAI CEO Sam Altman recently argued that humanity is already inside an AI singularity, framing the current period as the transformation technologists once discussed as a distant possibility. The new letter presents the same acceleration from a more cautious angle: if progress is becoming self-reinforcing, institutions may need the ability to slow it before the consequences are fully understood.
Not a Moratorium, but an Emergency Option
The signatories are not asking the government to stop current model training or impose a fixed moratorium. They are asking for an option that does not yet exist: a credible, internationally coordinated way to reduce the pace of frontier-wide development if automated research creates an emergency.
That distinction may make the proposal more politically realistic, but it also leaves difficult questions unanswered. Governments would need to define which models or training runs fall within the system, what evidence would justify intervention, how compliance would be verified, and who would decide when normal development could resume.
Technical enforcement would be equally complicated. Frontier research is distributed across corporate laboratories, cloud platforms, universities, startups, and governments. Models can be trained in one jurisdiction, deployed from another, and accessed globally through APIs or downloadable weights.
The proposal also risks concentrating power if only a small group of governments and companies control the brake. Critics may argue that pacing mechanisms could protect established laboratories from competition, restrict open research, or become tools for political control.
The signatories acknowledge the danger of poor implementation. Their argument is that the absence of any mechanism is also a choice, and one that leaves competitive pressure as the primary force determining how quickly increasingly autonomous systems are developed.
The letter arrives as the industry begins building more formal defenses around frontier models. NVIDIA, Microsoft, IBM, and other companies recently created an open security alliance focused on protecting agents and software from AI-enabled attacks. That initiative addresses present cybersecurity risks, while Pacing the Frontier is aimed at the possibility that capability development itself becomes difficult to control.
Whether the U.S. government will act is uncertain. Comprehensive federal AI legislation has repeatedly stalled, and any international agreement would require cooperation among countries that increasingly treat advanced AI as a strategic advantage.
The importance of the statement lies less in a specific policy proposal than in who signed it. The people asking for a brake include senior leaders and researchers inside the laboratories pushing the frontier forward. Their message is that the industry may soon need the ability to slow a race that no participant believes it can safely leave on its own.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
