How It Works
Responsible Scaling Policies
Public commitments by AI developers that tie the training and release of more capable models to specific safety tests and safeguards. If a model crosses a defined danger threshold, the developer commits to stronger protections — or to pausing — before going further.
Origin
The term was promoted by the evaluation group ARC Evals (now METR) in September 2023, the same month Anthropic, led by Dario Amodei, published the first policy under that name. Other labs have since published similar frameworks under different names.