Privacy Risks and Preservation Methods in Explainable Artificial Intelligence: A Scoping Review

Explanations can leak sensitive information: When AI systems provide explanations for their decisions, they can inadvertently reveal private details about individuals in the training data or about specific people being evaluated. The paper identifies this as a critical tension—the very transparency needed to build trust in AI systems can compromise privacy.

Multiple privacy preservation techniques exist but lack standardization: Researchers have developed various methods to protect privacy while maintaining explainability (such as differential privacy, federated learning approaches, and aggregation techniques), but the review found no consensus on best practices. This fragmentation makes it difficult for practitioners to know which approach to implement.

"Privacy-preserving explanations" require specific design characteristics: Rather than simply adding privacy protections to existing explanation methods, the paper proposes that truly privacy-compliant explanations need to be designed with particular features from the ground up—suggesting that privacy and explainability shouldn't be treated as separate concerns but integrated into system design.

Balancing privacy with other important AI goals remains challenging: The review identifies ongoing tensions between privacy preservation and other desirable system qualities like accuracy, fairness, and usefulness of explanations. Organizations implementing XAI must navigate these competing demands without clear guidance.

The field needs clearer standards and recommendations: The authors call for better frameworks to help researchers and practitioners understand what constitutes adequate privacy protection in explainable AI systems, indicating this remains an unsolved problem despite growing research attention since 2019.