Proactive, But Not Creepy: Legitimacy and Disclosure Boundaries in Generative IR Assistants

초록

Proactive generative AI-based information retrieval (IR) assistants can reduce user effort by initiating suggestions without explicit requests. However, proactive initiation can raise privacy concerns and uncertainty about system use of personal context, making acceptance difficult to predict. This paper investigates proactive IR assistant acceptance through a mixed-methods design. A quantitative study (N=112) tests a structural equation model showing social presence influences proactivity acceptance through trust, privacy concern, intention to disclose, and satisfaction, while goal criticality and information sensitivity shape key indirect effects. Results show that social presence increases trust and decreases privacy concern, with both intention to disclose and satisfaction predicting acceptance. However, structural relationships alone cannot explain how users interpret proactive interventions at the moment they occur, or how they negotiate disclosure boundaries in practice. To address these gaps, a qualitative study with semi-structured interviews (N=12) reveals that acceptance forms through legitimacy judgments, conditional trust calibration, and boundary-based disclosure management. Design implications emphasize clear explanations at initiation, incremental disclosure with user control, and context-adaptive proactivity intensity.

키워드