Faithfulness and the Notion of Adversarial Sensitivity in NLP Explanations

Current faithfulness evaluation methods in NLP are unreliable: The paper identifies that existing approaches to measuring whether AI explanations accurately reflect how models actually work contain "discrepancies and biases" and often fail to capture true model reasoning, highlighting a critical gap in how we verify explainable AI systems.

Adversarial attacks reveal whether explanations are genuine: By testing how explanations change when models face adversarial inputs (deliberately crafted to fool AI systems), researchers can better determine if explanations truly represent the model's decision-making or are merely superficial descriptions.

Sensitivity to input changes is a key faithfulness indicator: The paper proposes that faithful explanations should respond appropriately when model inputs are adversarially modified—if an explanation claims to show why a model made a decision, it should change when the decision changes under attack.

This addresses an "underexplored paradigm": The research tackles an important but previously neglected angle in AI accountability, suggesting that robustness testing of explanations themselves (not just models) is essential for responsible AI deployment.

Better evaluation methods strengthen AI trustworthiness: By developing more rigorous ways to verify that explanations are faithful to actual model behavior, this work supports the broader goal of making AI systems more transparent and reliable for real-world applications.