Personality as Relational Infrastructuring: User Perceptions of Personality-Trait-Infused LLM Messaging

Abstract

Digital health interventions often rely on static message templates for health behaviour support, which can struggle to sustain engagement over time. Large language models (LLMs) offer a promising alternative by enabling context-sensitive message generation with personality-trait-based language adaptation across repeated interactions. Yet most prior work evaluates generated messages in isolation, leaving unclear whether such adaptation improves perceptions of individual messages or whether its benefits emerge through cumulative exposure. To examine this question, 90 participants rated hypothetical messages generated with four LLM strategies, with and without Big Five personality-trait-based language adaptation, in a controlled online study. Using ordinal multilevel models with within-person and between-person decomposition, we found no message-level effect on perceived personalisation. However, greater exposure to personality-trait-adapted messages was associated with increased perceived personalisation, increased appropriateness, and decreased negative affect. These findings suggest that the benefits of personality-trait-based language adaptation may emerge through sustained interaction rather than single-message optimisation.