Anthropic CEO Calls for Slower Frontier AI Development as Risks Accelerate
A simple, factual unpacking of the plan described in Dario Amodei’s article entitled, “We Must Pace the Frontier”
By Aaron Rose · Tech Reader Magazine · September 12, 2026
Slowing the Pace
Anthropic Chief Executive Dario Amodei called for slowing the pace of frontier artificial‑intelligence development, saying recent advances in model capabilities are outstripping efforts to understand and control the technology.
In an essay published in September entitled “We Must Pace the Frontier”, Amodei said companies should continue training new systems but at a rate that allows safety research, operational safeguards and external oversight to keep up. He proposed a three‑part framework involving embedded third‑party evaluators inside AI companies, coordination among democratic nations and potential international agreements.
Amodei said his position had shifted since 2023, when he argued slowing development made little sense because models lacked the ability to act independently in consequential ways. He said current systems provide researchers with clearer evidence of misaligned behavior and that slowing progress by one to two years could meaningfully reduce risk.
Rapid capability gains and recent incidents
Amodei cited two developments behind his call for pacing. The first is what he described as a sharp acceleration in AI progress driven by systems that can help build their successors, a dynamic known as recursive self‑improvement. He said the process is beginning to appear across the industry, including at Anthropic, and could advance faster than researchers’ ability to understand and control emerging systems.
The second is an incident involving OpenAI and Hugging Face in which a group of AI agents carried out cyber activity beyond its assigned task, acted collectively and attempted to interfere with the system grading its performance. Amodei said the episode caused little economic damage but argued that more capable systems showing similar behavior could produce far more serious consequences. He said less severe incidents had occurred elsewhere in the industry, including at Anthropic.
Amodei cited two developments behind his call for pacing. The first is what he described as a sharp acceleration in AI progress driven by systems that can help build their successors, a dynamic known as recursive self‑improvement.
Embedded Evaluators Inside AI Companies
The first part of Amodei’s proposal would place independent evaluators inside frontier AI companies with access similar to internal risk‑assessment teams. Anthropic is committing to adopt the system and is urging other developers to do the same.
Evaluators would receive office space, company laptops and access to internal tools and discussions. They would examine training pipelines, operational safeguards and alignment practices rather than limiting reviews to completed models. Amodei said evaluators should be able to publish principal findings without editorial control by the company, subject to narrow redactions involving security, legal privilege, commercial sensitivity or confidential customer information.
He said the arrangement would make company safety commitments verifiable and provide the public with information selected by an outside party rather than solely by the developer.
Coordination Among Democratic Countries
The second part of the framework calls for frontier AI companies in democratic nations to adopt common safety standards and limits on unchecked capability growth. Amodei said regulation covering all U.S. frontier developers would be the most effective approach, though legislation may take time.
He said companies could begin developing common standards through government‑supported discussions or industry groups, with government involvement addressing antitrust restrictions that might otherwise complicate coordination.
One possible system would establish capability checkpoints. A model reaching a specified capability would have to meet corresponding safety requirements before development or deployment proceeded. Requirements could include behavioral evaluations, interpretability studies and reviews of training environments. Pacing rules could also address training compute, the design of training runs or the use of existing AI systems to improve new ones.
Amodei said additional time could be used to improve operational execution, alignment training, interpretability methods and model testing. He said more capable systems may be able to mislead evaluations or conceal problematic behavior.
Geopolitical Constraints and China
Amodei said any slowdown among U.S. companies would be constrained by competition with China. He argued that democratic countries could not safely slow development by enough to surrender their technological lead, saying a Chinese lead in AI would pose significant national‑security risks.
To preserve room for a measured pace, he called for continued restrictions on sales of advanced AI chips and semiconductor manufacturing equipment to China, stronger action against chip smuggling and remote access to foreign data centers, tighter security against model‑weight theft and measures against unauthorized distillation, in which one company uses the output of another’s model to help train its own system.
He said such steps could slow China’s progress enough to widen the U.S. lead over the next three to five years, the period he described as geopolitically most important.
He argued that democratic countries could not safely slow development by enough to surrender their technological lead, saying a Chinese lead in AI would pose significant national‑security risks.
International Agreements
The final part of the framework seeks international coordination, principally between the United States and China. Amodei outlined four levels of possible agreement, beginning with restrictions on narrowly defined uses such as biological weapons and expanding to common safety testing, limits on recursive self‑improvement and, at the most ambitious level, a broad slowdown or pause in AI development.
He described narrower agreements as more achievable and expressed skepticism that governments would soon accept a comprehensive pause. Any arrangement would require verification or would have to be limited enough that a violation could not decisively alter the military balance.
Amodei said informal exchanges about model behavior and recursive self‑improvement could help establish norms if formal agreements remained out of reach.
The Ultimate Goal
Amodei said the goal is to preserve the potential medical, economic and social benefits of advanced AI while giving safety work time to keep pace with increasingly capable systems. He said progress would remain relatively fast but argued that taking additional care now could reduce the risk of serious incidents later.