Development and Evaluation of an Image Generating AI–Based Conversational System for Reminiscence Therapy

抄録

This study explores the use of generative AI to support reminiscence therapy for elderly. We developed a system that combines a chatbot with an image generation model capable of producing new images based on the ongoing dialogue. The goal was to examine whether dynamically generated visuals could stimulate memory recall and reduce caregivers’ burden during therapy. Two experiments were conducted: a preliminary study with elderly and a follow-up study with individuals with moderate to severe dementia. Through detailed analysis of the studies, the results revealed that participants were more likely to choose and discuss people-related topics―such as family members or friends―than those focused on objects or places. The phrasing of AI-generated questions was also found to strongly influence conversational flow, while even partially inaccurate images sometimes encouraged discussion and facilitated reminiscence. Future work will explore the use of voice interaction and embodied avatars, and further fine-tuning of system components to better accommodate the cognitive and emotional characteristics of elderly users.