Black-Box Access is Insufficient for Rigorous AI Audits
Apollo Research argues that rigorous AI safety audits need white-box, interpretability-based methods rather than black-box input-output testing alone, citing a FAccT paper on the limits of black-box evaluation.