Logo via Wikimedia Commons; treatment-A cover, license to verify on approval
Anthropic's CEO outlines a framework to pace frontier AI advancement with embedded safety checks and global coordination
Dario Amodei wants the AI industry to tap the brakes, but not slam them. The Anthropic CEO published an essay on September 12, 2026 titled “We Must Pace the Frontier,” laying out a three-step framework designed to keep AI development moving forward while making sure safety research doesn’t get left in the rearview mirror.
The proposal arrives at a moment when anxiety about AI’s trajectory has spread well beyond the usual circle of doomers and ethicists. In July 2026, more than 1,100 AI professionals signed an open letter calling for measures to moderate the pace of frontier development, arguing that capabilities are advancing faster than the guardrails meant to contain them. Amodei was among the signatories.
The three steps, unpacked #
Amodei’s framework breaks down into three sequential components, each escalating in ambition and coordination difficulty.
Step one: Embedded Evaluators. This is the only piece Anthropic plans to implement unilaterally, and it’s starting immediately. These evaluators would assess and verify safe practices during both model training and deployment, creating a continuous feedback loop rather than a one-time audit.
Step two: Democratic Coordination. This moves beyond any single company’s walls. The idea is to get major AI labs and democratic governments aligned on shared safety standards, essentially creating a common rulebook so that no single company gains a competitive advantage by cutting corners on safety.
Step three: Global Coordination. The most ambitious tier extends safety alignment to the international stage. Amodei envisions a world where frontier AI development is governed by international norms, not just the policies of whichever country hosts the most GPU clusters.
The word “pacing” is doing a lot of work here. Amodei is careful to distinguish his approach from calls to halt AI development entirely. His framework is designed to create adequate time for alignment research and verification of AI technologies to keep up with capability gains, not to freeze progress in place.
Context and evolution #
Anthropic previously developed its Responsible Scaling Policy, which uses AI Safety Levels (ASL) as a kind of tiered warning system. Those levels incorporate “if-then” safeguards: if a model reaches a certain capability threshold, then specific safety protocols kick in. The new framework appears to build on that foundation while adding layers of external coordination that the RSP, as an internal policy, couldn’t achieve on its own.
Amodei co-founded Anthropic in 2021 after leaving OpenAI. His prior essays, including “Machines of Loving Grace” in 2024 and “The Adolescence of Technology” in January 2026, laid intellectual groundwork for the argument that powerful AI can be beneficial, but only if its development is managed with genuine discipline.
The July 2026 open letter provided additional momentum. Having over 1,100 AI professionals publicly state that the industry’s pace is outstripping safety research made it harder for any lab to dismiss these concerns as the worries of outsiders who don’t understand the technology.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our