Explainable AI in Usable Privacy and Security: Challenges and Opportunities

LLM explanations can be unreliable in high-stakes decisions: The paper found that Large Language Models used to evaluate privacy policies suffer from inconsistency, hallucinations, and transparency issues—problems that are particularly concerning when users rely on these explanations to make trust decisions about their personal data.

Users have different needs for AI explanations: Based on a study with 22 participants, the researchers discovered that people vary significantly in how much detail they want in AI explanations and how engaged they want to be with the tool, suggesting that one-size-fits-all explanations are insufficient for responsible AI deployment.

Transparency about AI uncertainty matters: The paper identifies uncertainty estimation as a key strategy to build user trust, implying that AI systems should openly communicate when they're less confident in their judgments rather than presenting all conclusions with equal confidence.

Privacy and security are ideal testing grounds for responsible AI: The authors argue that usable privacy and security represents a promising application area for developing Human-Centered Explainable AI, because the stakes are high (personal data protection) and users genuinely need to understand and trust AI recommendations.

Technical safeguards alone aren't enough: The research suggests that mitigating LLM concerns requires combining multiple approaches—structured evaluation criteria, uncertainty estimation, and retrieval-augmented generation—rather than relying on any single technical solution.