Alignment Forum post proposes a legal minimum wait before new AI checkpoints can be used internally
Writing in a personal capacity, annawang suggests no checkpoint be used beyond safety testing until X calendar days after its hash is filed with an auditor, starting at 10 days and possibly rising to 90.
Why it matters
Most safety rules debated so far target external releases, while this proposal targets internal use, where labs increasingly run new models on their own research. It builds on an idea the AI Futures Project floated, that AI research should only be assisted by AIs trained months earlier. The post comes hours after AI Futures Project authors moved their forecast for a superhuman AI researcher to 2027. It is a personal proposal, not a lab policy or bill; the author lists open questions on how it could fail.
Key facts
- Author: annawang, writing in a personal capacity
- Posted on the AI Alignment Forum on 11 November 2026
- Wait starts when a checkpoint's hash is filed with an auditor
- Suggested wait: 10 calendar days, possibly rising to 90
- Safety testing allowed during the wait