Explainability won't save AI

Explainability techniques are currently developed and incorporated by machine learning engineers, and not surprisingly, their needs (and companies' desire to avoid legal trouble) are being prioritized.Realizing a broader set of XAI objectives will require both greater awareness of their existence and a shift in incentives for accomplishing them. XAI standards and policy guidelines should explicitly include the needs of users, stakeholders, and impacted communities to incentivize this shift. Explainability case studies are one pedagogical tool that can help practitioners and educators understand and develop more holistic explainability strategies. Diverse organizational guidance documents, recommendations, and high-level frameworks can also help guide an organizations' executives and/or developers through key questions to support explainability that is useful and relevant to different stakeholders. While there has been some work done to evaluate AI explanations, most attempts are either computationally expensive or only focus on a small subset of what constitutes a "good explanation" and fail to capture other dimensions.