Talk to Me Kindly: Designing AI Companions to Foster Positive Self-Talk in Young Adults

Abstract

Young adults (aged 18–25) are prone to emotional and mental health struggles, with rumination that can manifest as negative self-talk. Positive self-talk (PST) can counteract this through reframing that promotes well-being. While the use of Large Language Models for mental health has increased, limited work promotes PST for young adults. To address this gap, we interviewed therapists (N = 11) to study how they foster positive self-talk and how these strategies might translate into human-AI interactions. We found that PST is shaped by identity development and demographic factors. Based on these findings, we propose the Human-AI Self-Talk (HAIST) framework to guide the design of non-clinical AI companions. We refined the framework with four AI-experts. Our findings indicate that rapport building can be supported through adaptation of users’ language, interactional style and operationalized through Dialogical Self-Theory. Our work operationalizes self-talk theories within an interactional framework to support learning and self-sufficiency.