XMENTOR: A Rank-Aware Aggregation Approach for Human-Centered Explainable AI in Just-in-Time Software Defect Prediction

Multiple AI explanations can confuse rather than clarify: When different explainability methods (LIME, SHAP, BreakDown) were applied to the same defect prediction task, they often produced conflicting explanations, increasing developer confusion and cognitive load—highlighting a critical gap between technical transparency and practical usability.

Aggregating explanations significantly improves developer trust: Nearly 90% of study participants preferred unified, aggregated explanations over individual ones, demonstrating that thoughtfully combining multiple AI perspectives into a single coherent view can meaningfully enhance trust and reduce frustration in real-world workflows.

Context matters: embedding explanations into existing tools works better: By implementing XMENTOR as a VS Code plugin rather than a standalone tool, the researchers showed that integrating AI explanations directly into developers' daily workflows (debugging and code review) makes them more useful and actionable than presenting them separately.

Responsible AI requires human-centered design, not just technical accuracy: The paper demonstrates that explainability methods alone don't ensure trustworthy AI—the presentation, aggregation strategy, and integration into human workflows are equally critical to whether developers actually understand and appropriately rely on AI predictions.

Smart filtering prevents information overload: XMENTOR's use of adaptive thresholding and rank-agreement strategies to filter conflicting signals shows that responsible AI sometimes means showing less information more clearly, rather than maximizing transparency at the cost of usability.