Beyond Explainable AI (XAI): An Overdue Paradigm Shift and Post-XAI Research Directions

Current explainability methods have fundamental flaws that can't be fixed by incremental improvements. The paper identifies that XAI approaches contain "two paradoxes, two conceptual confusions, and five false assumptions," suggesting that trying to patch existing explainability techniques may actually make problems worse rather than solve them.

We should shift focus from explaining AI decisions after the fact to verifying AI performance upfront. Rather than attempting post-hoc explanations of why an AI system made a particular decision, the authors propose "verification-focused Interactive AI" that establishes protocols to certify that AI systems work reliably before deployment—similar to how we test medications or aircraft rather than explaining why they failed.

Different users need different types of AI transparency. The paper introduces "User-Sensible AI," emphasizing that a doctor, a patient, and a regulator may all need different information to trust an AI system. One-size-fits-all explanations don't serve diverse communities effectively.

AI development needs stronger scientific foundations. The authors call for "AI Epistemology"—essentially establishing rigorous scientific standards for how we build and validate AI systems, rather than relying on informal explanation techniques that lack clear validation methods.

Technical analysis of AI models should focus on what's actually happening inside them, not post-hoc stories. "Model-Centered Interpretability" means studying the genuine mechanics of how AI systems work, rather than generating explanations that may not reflect the true reasons behind AI decisions.