What the paper contributes
Proposes evaluating dangerous capabilities and the propensity to use them harmfully, to inform decisions about training, deployment, security and policy.
Why it belongs here
Makes one mechanism behind calls for independent evaluation more concrete.
What this does not establish
A framework and research argument. Passing an evaluation does not establish a general guarantee of safety or settle who should set the threshold.
Read it alongside
AI Licensing for a Better Future
Explore the mechanism
The lab translates a governance question into a hypothetical arrangement. It illustrates assumptions; it does not reproduce the paper’s results or predict an outcome.
Open the accountability lab