Position: Explain to Question not to Justify

Two distinct purposes for AI explanations exist: The paper distinguishes between BLUE XAI (explanations designed for humans and stakeholders to understand AI decisions) and RED XAI (explanations designed to validate and improve the AI models themselves), arguing these serve fundamentally different goals that shouldn't be conflated.

Model validation and debugging are critically under-explored: The authors contend that RED XAI—methods for extracting knowledge from well-performing models and identifying bugs in faulty ones—receives far less research attention than it deserves, despite being essential for ensuring AI safety and reliability.

Current XAI progress is hindered by conflicting objectives: The field's advancement is being slowed by incompatible goals between those focused on human-centered explanations and those focused on technical model validation, suggesting the community needs to clarify which problem it's actually solving.

Explanations should "question" models, not just "justify" them: The paper's core argument is that XAI should prioritize using explanations as a tool for interrogating and improving AI systems rather than primarily serving as a post-hoc justification mechanism for already-deployed models.

Safety requires shifting research priorities: The authors identify significant untapped research opportunities in RED XAI that are necessary for building safer AI systems, implying that responsible AI development requires more focus on internal model validation than current XAI literature provides.