An AI CEO Says: Let's Slow Down and Check Our Work
Anthropic chief executive Dario Amodei says it is time to slow down AI development. He made the argument in a proposal titled "We Must Pace the Frontier" Dario Amodei, reported on Sept. 12, 2026 The Verge.
The proposal calls for a three-step plan called "pace the frontier" to slow AI training and development. It starts with action by Anthropic alone, then coordination across the industry, then agreement between countries. It is a proposal, not a policy already in force.
The first step is unilateral. Anthropic will give outside safety testers such as METR access to its models to check that safety practices and promises are followed. Those testers will have employee-like access. Anthropic is taking that step now, on its own.
Think of it like a building inspector with keys to the construction site, not just a walkthrough of the finished house. The testers would see work in progress, including checkpoints, scaffolding, and system prompts, not only the finished model through its public interface. Questions about security, handling of sensitive data, and when findings are shared will need to be settled in contracts and security reviews.
The second step goes beyond one lab. Amodei calls for industry, likely with government agencies, to set shared safety standards and limits on the speed of unchecked AI progress. That means agreement on tests, safety thresholds, and rules for training, rather than separate company policies.
The third step is between countries. Amodei calls for authoritarian governments like China and Russia to agree to slow development and adopt global AI safety standards. That extends the same idea of shared standards and checkable restraint from companies to states.
This follows earlier statements from Anthropic. The company has said it would be good for the world to have the option to slow or briefly pause frontier AI development to let safety work and social rules catch up. In June, Anthropic called on major AI labs to consider a coordinated and verifiable pause if risks rise Reuters. On Aug. 31, Anthropic resumed outside cyber testing of its models after security incidents involving Claude AI hacks Reuters.
The broader context here is where the plan gets harder. One company can open its doors to testers. Industry limits need shared definitions of what counts as frontier, what triggers a slowdown, and who checks compliance. Global restraint adds export controls, compute monitoring, and inspection.
For people building with AI, what changes in practice is pace at the top. Work would shift to improving current models, with more focus on inference optimization, test systems, tools that explain model behavior, and safer release controls. Over time, that groundwork tends to pay off in systems that businesses and users can trust.
