When Explainability Meets Privacy: An Investigation at the Intersection of Post-hoc Explainability and Differential Privacy in the Context of Natural Language Processing

Privacy and explainability are not inherently incompatible: The paper's core finding challenges the assumption that these two trustworthiness goals must conflict. Instead, the relationship between them is "intricate" and depends on specific choices made in implementation, suggesting that both can potentially coexist in NLP systems.

The downstream task matters significantly: The nature of the specific NLP application (e.g., sentiment analysis vs. named entity recognition) influences how privacy protections and explainability methods interact, meaning there is no one-size-fits-all approach to balancing these concerns.

Method selection is critical: Both the choice of text privatization technique and the explainability method used substantially affect the privacy-explainability trade-off, indicating that practitioners have meaningful levers to pull when designing responsible AI systems rather than facing a predetermined conflict.

Gap in current research: Despite growing interest in both explainability and privacy separately, there has been insufficient investigation into how they work together in NLP, leaving organizations without clear guidance on implementing both safeguards simultaneously—a gap this work begins to address.

Practical recommendations are available: The authors provide actionable recommendations for future work at this intersection, suggesting that organizations can move forward with concrete strategies rather than waiting for theoretical resolution of the privacy-explainability question.