Customising Synthetic Personality Traits of Conversational Agents for Self-Reflection

This research project investigates how user-driven customisation of Conversational Agents (CAs) affect participants’ self-disclosure and self-reflection levels during a self-reflection task. To this end, we developed an interactive web-based interface that allows participants to customise a CA’s synthetic personality along five trait dimensions. Personality traits were selected as the customisation variable as prior work demonstrates that prompting can effectively imbue CAs with distinct personality profiles. During the study, participants interact with the CA by sharing a life event, defined as a meaningful moment from their life. This type of task provides an opportunity for introspection, connects to traditional journaling practices, and relates to theories of narrative identity—all of which are linked to well-being. We compare three levels of user-driven customisation in a between-subject design: no customisation (baseline), a placebo customisation in which participants were led to believe they could alter the CA's personality traits, and a participant-specified customisation in which participants’ chosen traits shaped the CA’s synthetic personality and its subsequent responses. To understand the effects of these conditions, we employ a mixed-methods approach combining quantitative and qualitative data. The main quantitative outcome variables are self-reflection and self-disclosure. Qualitative insights are gathered through open-ended questions focusing on themes such as reflection or perception of the CA. In an exploratory component, we examine whether participants’ chosen synthetic personality traits for their CA align with traits generally preferred in people such as high agreeableness. Furthermore, we consider participants’ sense of self as a potential moderating factor in their choice of synthetic personality traits. This work is expected to have implications for the design of customisable CAs in mental health and well-being contexts. Based on the study findings, we will propose design recommendations on appropriate levels of user-driven customisation. Additionally, our analysis of the relationship between the participants' personality traits and their chosen CA traits will contribute to a deeper understanding of the underlying mechanisms of the human-CA interaction.