Evaluating Search System Explainability with Psychometrics and Crowdsourcing

Explainability is multidimensional, not one-size-fits-all: The paper challenges the common practice of treating explainability as a single concept, demonstrating through psychometric analysis that it actually comprises multiple distinct human-centered factors. This means that effective explanations in search systems need to address several different aspects simultaneously rather than focusing on just one type of explanation.

User studies reveal what actually makes explanations work: Using crowdsourced evaluation with real users, the researchers developed the SSE (Search System Explainability) metric and showed it can reliably distinguish between systems that users find interpretable and those they don't. This suggests that explainability evaluation should be grounded in actual human understanding rather than purely technical measures.

Current explainability approaches lack consensus and rigor: The paper identifies a significant gap in the field—despite advances in explainable AI, there is "no consensus on the definition of explainability." This fragmentation makes it difficult to hold AI systems accountable, which is critical for ensuring fairness and reducing bias in widely-used systems like search engines and conversational agents.

This work provides a template for evaluating explainability across AI domains: The authors position their psychometric and crowdsourcing methodology as a "blueprint for similar explainability evaluation efforts" beyond search systems, suggesting this human-centered approach could improve accountability standards across machine learning and natural language processing applications more broadly.