Measuring the Fragility of Trust: Devising Credibility Index via Explanation Stability (CIES) for Business Decision Support Systems

Explanations from popular AI interpretation tools (SHAP, LIME) can be unstable: The paper demonstrates that the reasons AI systems give for their decisions may change significantly when data is slightly altered, even though the predictions themselves remain the same—a critical concern for high-stakes business decisions like credit approval or employee hiring.

How you prepare data affects explanation reliability, not just accuracy: The study found that SMOTE, a common technique for handling imbalanced datasets, impacts the stability of explanations alongside predictive performance. This suggests businesses need to evaluate their data preprocessing choices through a credibility lens, not just a performance one.

The most important features deserve extra scrutiny: The CIES metric weights instability in top decision drivers more heavily than marginal features, reflecting real business priorities—a change in why a loan was denied matters more than a change in why it ranked 15th on a risk list.

A practical credibility warning system is now possible: The paper provides a deployable metric that businesses can use to flag when they should be cautious about trusting an AI system's explanations, offering a concrete tool for responsible AI deployment in customer churn, credit risk, and hiring decisions.

Model complexity creates credibility trade-offs: More complex models showed different explanation stability patterns, suggesting organizations face a potential tension between predictive power and explainability reliability that requires deliberate evaluation.