Interpretability is not Explainability: New Quantitative XAI Approach with a focus on Recommender Systems in Education

Interpretability and explainability are not the same thing: The paper establishes that these terms, often used interchangeably, have distinct meanings in AI systems. A model can be interpretable (its internal logic is understandable) without being explainable (able to justify its decisions to users), an important distinction for anyone evaluating AI transparency claims.

There's no standard way to measure explainability: The paper identifies a critical gap in the field—there is currently no widely accepted framework for quantitatively evaluating how explainable an AI system actually is. This means organizations may claim their systems are "explainable" without objective evidence, making it difficult for stakeholders to assess whether AI decisions are truly understandable.

Five key dimensions should define explainability: The authors propose that explainability should be evaluated across five measurable aspects: transparency (how open the system is), interpretability (how understandable its logic is), completeness (whether all relevant factors are shown), complexity (how simple the explanation is), and understandability (whether users actually comprehend it). This framework provides a concrete checklist for responsible AI implementation.

Educational recommender systems can be made more explainable: The paper demonstrates that these five dimensions can be quantitatively measured in real-world systems, using a case study of an educational resource recommender. This shows that the framework is practical and applicable to systems that directly affect student learning outcomes.

Shared vocabulary is essential for progress: The paper emphasizes that without common definitions and measurement standards across the AI field, it's impossible to compare systems, hold developers accountable, or build genuine trust in AI systems used in sensitive domains like education.