Impact of data minimization approaches on users’ privacy perceptions and behaviors

Determine how different approaches to data minimization in large-language-model-based conversational agents influence users’ privacy perceptions and their privacy-related behaviors during interactions with these systems.

Background

The paper introduces Rescriber, a browser extension that supports user-led data minimization by detecting and sanitizing personal information in prompts to LLM-based conversational agents. The authors note broader challenges in balancing privacy, utility, and convenience, and highlight uncertainty about how different data minimization strategies affect users’ perceptions and behaviors.

To address this gap, the system offers two methods—redaction (replacement with placeholders) and abstraction (rewriting to generalize sensitive information)—and evaluates them through user studies. The quoted sentence explicitly states uncertainty regarding the effects of varying minimization approaches on user perceptions and behaviors.