Does Faithfulness Conflict with Plausibility? An Empirical Study in Explainable AI across NLP Tasks

Faithfulness and plausibility don't have to be opposing goals: The study found that traditional methods like Shapley value and LIME can achieve both high faithfulness (accurately reflecting how the AI model actually works) and high plausibility (making sense to domain experts) simultaneously, challenging the assumption that these two qualities inevitably conflict.

Perturbation-based methods outperform alternatives: Across three common NLP tasks—sentiment analysis, intent detection, and topic labeling—the researchers found that perturbation-based explainability methods (which test what happens when you remove or change inputs) performed better than other approaches at meeting both faithfulness and plausibility standards.

Dual-objective optimization is achievable: Rather than forcing developers to choose between explanations that are technically accurate or explanations that make sense to humans, the research suggests that explainability algorithms can be designed with both goals in mind, potentially improving both the accuracy and user-friendliness of AI explanations.

Expert validation matters for responsible AI: The study's comparison of algorithmic explanations against expert-level interpretations demonstrates the importance of involving domain specialists when evaluating whether AI systems are truly explainable—a key requirement for building trustworthy AI systems.