The Pour Demain, a non-profit AI think tank, released a policy brief that addresses the evaluation of General Purpose AI (GPAI) models under the EU AI Act’s Codes of Practice. The brief argues that current black-box testing methods are insufficient for thorough AI model evaluation. Instead, it advocates for a more comprehensive approach that combines black-box, de facto white-box, and outside-the-box methodologies.
OpenMined was proud to see our work with the UK AI Safety Institute highlighted as one of the technological solutions that can assist in secure GPAI model evaluation. As noted in the brief, such work underscores the viability of these emerging solutions and highlights the critical role these technologies play in advancing GPAI model evaluations.
Read more here.