Position: Explainability Research Must Prioritize Foundations over Ad-hoc Methods

Explainability tools often go unused in practice: Despite the rapid development of numerous XAI techniques (feature attributions, sparse autoencoders, and others), explanations are frequently generated and then discarded without actually guiding meaningful decisions or actions in real-world workflows.

The field lacks foundational frameworks: The paper identifies three critical structural gaps preventing progress: unclear problem formulations about what explanations should accomplish, underspecified evaluation objectives for measuring explanation quality, and the absence of established pipelines for using explanations to improve AI systems through feedback loops.

Current research prioritizes novelty over usability: Analysis of recent papers from top ML conferences (ICML, NeurIPS, ICLR) and surveys of XAI practitioners reveal that the field has focused on developing new ad-hoc methods rather than addressing the fundamental challenges needed to make explanations actually useful in human-centered systems.

Explanations need to be integrated into human-in-the-loop systems: Rather than treating explanations as standalone outputs, the paper argues they must be designed as part of end-to-end systems where humans can act on the explanations and provide feedback to improve the AI system over time.

A human-centered approach is essential for responsible AI: The paper proposes shifting toward an "action-oriented paradigm" with practical guidelines, emphasizing that responsible AI requires explanations designed specifically to support human decision-making and accountability, not just technical transparency.