Explanation-by-Example Based on Item Response Theory

Black-box AI models pose a trust problem: Many machine learning systems used in everyday applications cannot explain their own decisions, raising legitimate questions about whether users should trust their predictions—a gap that explainable AI (XAI) aims to address.

Explanation-by-Example is a practical XAI approach: The paper examines "Explanation-by-Example," a technique that helps users understand AI decisions by showing similar past examples, which is more intuitive for non-technical audiences than complex mathematical explanations.

Item Response Theory can measure explanation reliability: The research demonstrates that Item Response Theory (IRT)—a statistical method originally developed for educational testing—can be repurposed to assess how reliable an AI system's explanations actually are, not just whether they exist.

The approach successfully identifies unreliable predictions: When tested on a Random Forest model across multiple datasets, IRT flagged 83.8% of the model's errors as coming from instances where the system itself was unreliable, suggesting this method could serve as an early warning system for when to distrust AI recommendations.

Reliability assessment matters for real-world deployment: By distinguishing between cases where an AI system is confident versus uncertain, this work addresses a critical need for responsible AI: helping users know not just what a system decided, but when that decision is trustworthy.