CMU Study: AI Privacy Redaction Trust Boosted by Explanations

Researchers from Carnegie Mellon University (Roshni Kaushik, Maarten Sap, Koichi Onoue) explored how different explanation styles influence user trust in AI systems that redact private information during mediated conversations. In the study, an AI mediator redacted sensitive content from messages and then provided recipients with explanations ranging from minimal to detailed. With 180 participants, the team measured how trust and preference changed based on redaction amount and explanation detail.

Key results: Providing any explanation made the system seem significantly more effective at preserving privacy (p<0.05, Cohen's d ~ 0.3). When the system performed heavier redactions, participants relied more on explanations and found them more helpful (p<0.05, Cohen's f ~ 0.2). However, explanation preferences varied by individual — age and baseline familiarity with AI strongly moderated trust. The findings underscore a critical tension: too little explanation leaves users suspicious; too much may overwhelm. The authors advocate for adaptive, context-aware explanations to balance transparency and privacy, with implications for tools like AI assistants, customer service bots, or any system that filters sensitive data on behalf of users.