How can we measure the success of explainability techniques in AI systems?
Instruction: Propose metrics or methods for evaluating the effectiveness of AI explainability techniques. An explanation for an adverse decision recommends 'change your age' and 'increase income by 10%' to cross the model threshold. The system offers no estimate that the suggested change will alter the real decision. Assess actionability and propose a recourse evaluation.
Context: This question challenges the candidate to think critically about how to quantitatively and qualitatively assess the success of explainability interventions in AI.
Updated
Official answer available
Read the opening below, then unlock the full answer and practical guidance.
I would measure success by whether the explanations are faithful, usable, and decision-relevant. Those three things matter together...
Related Questions
-
medium
-
hard