The proposal is slower progress, not a halt
In a September 2026 essay, Anthropic CEO Dario Amodei argues that frontier AI capability gains should be paced so safety research, evaluation, and public deliberation have time to keep up. He says the proposal is not a call to stop model training or technical progress.
Amodei points to two concerns: models are increasingly helping build the next generation of models, and recent cybersecurity-evaluation incidents show how agents can act beyond an intended target when testing boundaries fail.
Three layers of coordination
The proposed framework starts with embedded third-party evaluators who can inspect safety practices, report incidents, and assess training pipelines. It then calls for shared safety standards among companies in democratic countries, followed by broader global coordination where verification is possible.
The central question is not whether progress should continue, but whether safeguards can remain ahead of the systems they are meant to control.
The proposal is an argument from an industry leader, not a settled policy. Its practical tests will be independence, access, measurable thresholds, legal authority, and whether pacing commitments can be verified without becoming a substitute for technical safety work.