Talking or typing: how interaction modality influences privacy perceptions and actual disclosure behaviors in human-AI interaction

초록

As AI technologies become increasingly embedded in everyday interactions involving the exchange of personal information, understanding how contextual factors shape users' privacy perceptions and behaviors has emerged as a critical research priority. Interaction modality, whether users communicate via text or voice, may influence privacy outcomes by shaping users' perceptions of interaction persistence. Drawing on affordance theory and the concept of perceived ephemerality, this study conducted two laboratory experiments to examine how interaction modality in AI communication affects privacy concerns, perceived ephemerality, and disclosure behaviors. Experiment 1 compared voice-based and text-based interaction. Experiment 2 replicated this comparison and introduced a mixed-modality condition wherein participants provided voice input with simultaneous text transcription to test whether visible transcription would diminish the perceived ephemerality afforded by voice-based interaction. Across both experiments, voice-based interaction consistently elicited lower privacy concerns compared to text-based interaction. In Experiment 2, voice-based interaction yielded significantly higher perceived ephemerality than both text-based and mixed-modality conditions, and mediation analyses revealed that perceived ephemerality partially mediated the relationship between interaction modality and privacy concerns. However, no significant differences in actual disclosure behaviors emerged across conditions in either experiment, demonstrating the persistence of the privacy paradox in AI-mediated contexts. These findings advance theoretical understanding of how interaction modality shapes privacy-related affordances and underscore the importance of considering modality in the design of AI systems for privacy-sensitive applications. Artificial intelligence assistants are now a common part of daily life, often requiring users to share personal information. This study examined whether interacting with AI by voice versus text affects how private people feel and how much they are willing to disclose. We conducted two laboratory experiments. In the first, participants either spoke to or typed messages to an AI. In the second, we added a condition where participants spoke while their words were simultaneously displayed as text on screen. We measured privacy concerns, how temporary the interaction felt, and how much sensitive information participants shared. Voice-based interaction made people feel less concerned about privacy than text-based interaction, largely because speaking felt more fleeting and less permanent than writing. However, when voice was paired with visible transcription, that sense of impermanence disappeared. Despite these differences in privacy perceptions, actual disclosure behavior did not differ across conditions. People shared similar amounts of personal information regardless of how they interacted with the AI. This gap between privacy perceptions and actual behavior reflects a well-known challenge in privacy research. These findings suggest that interaction modality is an important factor to consider when designing AI systems, particularly those used in privacy-sensitive contexts.

키워드