Explanation Methods in Deep Learning: Users, Values, Concerns and Challenges

Different users need different explanations: The paper identifies that "expert users and lay users" have distinct needs, yet notes that "explanation methods or interfaces for lay users are missing"—meaning most AI explanations today are designed for technical specialists rather than the people actually affected by AI decisions.

Current explanation methods have significant blind spots: While visual explanations can show "various aspects of the influence of the input on the output," the paper finds that explanation methods struggle with two critical concerns: detecting "bias in datasets that leads to biased DNNs" and identifying "unfair outcomes"—suggesting that even when AI systems are explained, we may still miss discrimination problems.

Regulation alone isn't enough without better explanations: The paper analyzes the General Data Protection Regulation (GDPR) in the context of deep learning and concludes that legal requirements for explainability cannot be adequately met by current explanation methods, highlighting a gap between what regulations demand and what technology can deliver.

Explainability requires a holistic approach: The paper emphasizes that evaluating whether explanations are adequate requires considering four interconnected components: "users, laws & regulations, explanations and algorithms"—meaning responsible AI requires attention to all these elements together, not just technical fixes.