The request, in plain language.
A public, verifiable pause of at least six months on training systems more powerful than GPT-4, alongside shared safety protocols.
The proposal concerns frontier training, not all AI development. It asks governments to intervene if a voluntary pause cannot be arranged.
The implementation question
Who could verify a pause across competing laboratories?
A question for inquiry, not a finding about the authors’ intentions.
Keep the source and the reading apart.
This record summarizes a published position. The links below connect it to related documents, available research and an open follow-through record. Inclusion is not an endorsement, and a request is not proof of action.
Explore the evidence
Model evaluation for extreme risksRead the inquiry
Agreement is a beginning. A plan needs an owner.What happened afterwards?
Open the follow-through recordTry the mechanism
The accountability labRecord history
25 Sept 2026 · Source-linked summary added to the archive. Original publication: 22 Mar 2023. Publisher date; wider reporting followed later in March. Full source text has not been republished or preserved as a historical capture in this edition.