Regulator proposes disclosure rules for frontier training runs
The draft rule would require labs above a compute threshold to report training runs before they begin.
A regulator has published draft rules that would require AI labs training models above a set compute threshold to disclose the training run before it begins, part of a broader push toward pre-deployment oversight.
What triggers the disclosure
- A compute threshold measured in floating point operations
- Any model intended for public API release above that threshold
- An exemption for pure research runs not intended for deployment
Industry response has been mixed, with larger labs broadly supportive of a clear threshold and smaller developers warning the rule could still capture more routine work than intended.
