A Human-Centered Validation of the Explainability-Performance Coefficient

Explanation quality can be measured objectively: The paper introduces the EPC score, a metric that evaluates how well AI explanations work by balancing two competing needs—keeping explanations simple (using few features) while maintaining the model's accuracy. This addresses a key challenge in responsible AI: having a standardized way to assess whether explanations are actually good.

Human judgment validates AI explanations: The researchers tested their metric against real human evaluations and found that higher EPC scores correlated strongly with human assessments, including sentiment judgments and visual annotations. This is significant because it suggests the metric captures what humans actually find helpful in explanations, not just what looks good mathematically.

One metric works across different AI applications: The EPC score was validated on three different types of data—tabular (spreadsheet-like), text, and images—and performed consistently well. This suggests the approach could be broadly applied across various high-risk domains where explainability matters, from healthcare to finance.

Explainability doesn't require sacrificing performance: By explicitly measuring the trade-off between explanation simplicity and model accuracy, the EPC score helps practitioners understand that they don't necessarily have to choose between having trustworthy explanations and having accurate AI systems—both can potentially be achieved together.