[论文解读] Examining the Effect of Explanations of AI Privacy Redaction in AI-mediated Interactions

Roshni Kaushik, Maarten Sap+1 more|arXiv (Cornell University)|Mar 25, 2026

Explainable Artificial Intelligence (XAI)被引用 0

一句话总结

该论文研究在 AI 介导的沟通中对隐私 redaction 的解释如何影响用户信任和偏好，显示解释提高了感知隐私和信任，且情境与个体差异调节作用。

ABSTRACT

AI-mediated communication is increasingly being utilized to help facilitate interactions; however, in privacy sensitive domains, an AI mediator has the additional challenge of considering how to preserve privacy. In these contexts, a mediator may redact or withhold information, raising questions about how users perceive these interventions and whether explanations of system behavior can improve trust. In this work, we investigate how explanations of redaction operations can affect user trust in AI-mediated communication. We devise a scenario where a validated system removes sensitive content from messages and generates explanations of varying detail to communicate its decisions to recipients. We then conduct a user study with $180$ participants that studies how user trust and preferences vary for cases with different amounts of redacted content and different levels of explanation detail. Our results show that participants believed our system was more effective at preserving privacy when explanations were provided ($p<0.05$, Cohen's $d \approx 0.3$). We also found that contextual factors had an impact; participants relied more on explanations and found them more helpful when the system performed extensive redactions ($p<0.05$, Cohen's $f \approx 0.2$). We also found that explanation preferences depended on individual differences as well, and factors such as age and baseline familiarity with AI affected user trust in our system. These findings highlight the importance and challenge of balancing transparency and privacy in AI-mediated communications and suggest that adaptive, context-aware explanations are essential for designing privacy-aware, trustworthy AI systems.

研究动机与目标

• Motivate privacy-preserving AI mediation in high-stakes communications.
• Explore how redaction quantity affects user trust and explanation preferences.
• Assess whether explanations mitigate information asymmetry between communicating parties.
• Examine how user demographics and AI familiarity influence trust and preferences.

提出的方法

• Designed an AI mediator that redacts private information from researcher messages before sending to a collaborator.
• Implemented two explanation styles (general and thorough) and a no-explanation condition.
• Varied redaction level (high, moderate, low) to test contextual impact on trust and understanding.
• Conducted a between-subjects and within-subjects user study with 180 participants recruited via Prolific.
• Used LLM-based generation and evaluation to create questions, redactions, and explanations, with an LLM judge to verify privacy preservation.

实验结果

研究问题

• RQ1 Does providing an explanation to collaborators improve their trust in and understanding of the system?
• RQ2 How does the amount of sensitive information removed (redaction level) affect collaborators’ explanation preferences?
• RQ3 Do individual differences (demographics, AI familiarity) influence trust and explanation preferences?

主要发现

• Explanations increased perceived privacy preservation (p<0.05, Cohen's d ≈ 0.3).
• Participants relied more on explanations and found them more helpful when the system performed extensive redactions (p<0.05, Cohen's f ≈ 0.2).
• Trust in the system was influenced by individual characteristics such as baseline familiarity with AI/LLMs.
• Explanations also correlated with higher engagement and more positive open-ended feedback.
• The study confirms a trade-off between explanation detail and privacy leakage risk, with thorough explanations risking more leakage in text, though redaction procedures mitigated this in explanations.