responsible scaling policy
A lab's published commitment to pause or test larger models if certain safety risks emerge during training.
Think of it like
An alarm system that trips if the house gets too hot—you've committed to checking before it burns down.
Example
Anthropic's RSP: 'If we detect certain red-line capabilities during training, we'll pause, do further testing, and only proceed if we can mitigate risks.'
How it actually works
RSPs are a credible commitment device—a lab announces in advance that it will slow down if specific risk thresholds are crossed. Hard to enforce (who checks?), but signals seriousness about safety. Different labs have different thresholds and procedures.
For product teams
Signals safety commitment to regulators and public; reduces reputational risk.
For engineers
Gating on evals; capability measurements tied to go/no-go decisions.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome