Defining what constitutes a good explanation

Determine principled criteria that define what constitutes a good explanation in explainable artificial intelligence, particularly in the context of explanations for question answering systems.

Background

In the discussion of explainability, the authors emphasize the importance of faithfulness—explanations that causally reflect the model’s reasoning—as distinct from non-causal rationales. Despite this focus, the fundamental question of what makes an explanation ‘good’ remains unresolved.

Clarifying criteria for good explanations would guide both the design of explanation frameworks like QED and the evaluation of models, shaping standards for trust, debuggability, and human understanding in QA systems.