Uncertainty Gating for Cost-Aware Explainable Artificial Intelligence

Explanations from AI systems can be unreliable: The paper demonstrates that post-hoc explanation methods—tools used to help humans understand AI decisions—don't always provide faithful or stable interpretations, meaning the same model might generate different explanations for similar inputs.

Uncertainty can signal when to trust explanations: The researchers found a strong negative correlation between epistemic uncertainty (a measure of how confident the model is about its decision boundary) and explanation stability across four datasets and multiple AI architectures, suggesting uncertainty is a practical indicator of explanation quality.

Cost-benefit tradeoffs are possible in AI interpretation: Rather than always using expensive explanation methods, the paper proposes "routing" samples to cheaper or more expensive XAI methods based on predicted reliability—potentially reducing computational costs while maintaining explanation quality where it matters most.

This approach works across different AI types: The findings held consistently across tabular data, image classification, five different neural network architectures, and four different explanation methods, suggesting the uncertainty-reliability relationship is a generalizable principle rather than a narrow technical artifact.

Responsible AI deployment requires resource awareness: By enabling systems to defer explanations for uncertain cases or allocate computational resources strategically, this work addresses the practical reality that organizations have budget constraints when implementing explainability in real-world AI systems.