Addressing the Selection Problem in Explainable AI

The "Selection Problem" is a real barrier to XAI effectiveness: User studies show that existing explanation techniques often fail in practice, not because the techniques are flawed, but because users struggle to choose which explanation method matches their specific question. The paper formalizes this as a systematic gap between what users naturally want to know and the technical tools available to answer them.

Current XAI interfaces place an unfair burden on users: Conventional explainability systems require users to first understand their own uncertainty well enough to translate it into a technical method selection—a challenging prerequisite. For example, a user asking "why did the system deny my loan?" must somehow know whether they need feature importance, counterfactual explanations, or example-based reasoning before getting help.

A practical solution exists using AI orchestration: The paper proposes using large language models (LLMs) as intermediaries that can translate natural-language questions into appropriate XAI techniques automatically. Rather than forcing users to pick from a menu of explanation methods, an LLM could understand the user's intent and route them to the right explanation tool.

This addresses a critical gap in responsible AI deployment: For AI systems to be truly accountable and trustworthy, stakeholders need explanations they can actually use. The selection problem reveals that even well-designed explanation techniques fail if users can't access them effectively—highlighting an often-overlooked implementation challenge in responsible AI.