Investigating In-Context Privacy Learning by Integrating User-Facing Privacy Tools into Conversational Agents

A team of researchers from UMass Amherst has published a study investigating a novel approach to user privacy education for conversational agents (CAs) like ChatGPT. The core of their work is a prototype "just-in-time privacy notice panel" integrated directly into a simulated chatbot interface. This panel actively monitors user input, intercepts messages it identifies as containing potentially sensitive information, and provides immediate, contextual warnings. It also offers users protective actions and access to FAQs about privacy in CAs, aiming to create an experiential learning loop during normal use.

In their experiment, participants used versions of a chatbot with and without this privacy panel across two task sessions designed to mimic real-world interactions. The researchers then qualitatively analyzed survey responses and think-aloud transcripts. Key findings indicate that interactions with the privacy tool during use enhanced participants' privacy learning and shifted their perceptions. The study also identifies specific interface design features that either supported or hindered users in protecting sensitive information, providing a blueprint for future user-facing privacy tools in AI assistants.