Exploring Older Adults' Reminiscing with ChatGPT and Text-to-Image Technology
Yuxiang Zhai, Jiawen Zhang, Jihong Jeung · 2024 · Proceedings of the 26th International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS '24) · doi:10.1145/3663548.3688521
Summary
This short paper explores the potential of generative AI technologies — specifically ChatGPT for conversational elicitation and text-to-image generation (DALL-E) — as tools to support reminiscence storytelling by older adults. Reminiscence, the act of recalling and sharing personal memories, is widely recognised in gerontology and occupational therapy as beneficial for older adults' psychological well-being, sense of identity, and social connection. However, current technology-mediated reminiscence tools in HCI tend to position older adults as passive subjects — viewing photo albums, watching curated slideshows — rather than as active narrators who shape their own stories. The researchers conducted an experiment with seven older adults in China, aged 61 to 76, using a two-stage process: first, participants had conversations with ChatGPT (mediated by a researcher) where the AI asked follow-up questions to draw out detailed narratives about significant life memories; second, the narrative details were used to generate corresponding images via DALL-E, which participants then reviewed and discussed. The study used a structured approach based on Erikson's psychosocial development theory, prompting memories from different life stages (childhood, youth, middle age, later life). The researchers analysed both the storytelling process and participants' reactions to the generated images.
Key findings
The study identified several influential factors in AI-mediated reminiscence. On the positive side, ChatGPT's follow-up questions successfully elicited richer and more detailed narratives than participants would typically share unprompted — the AI's neutral, patient "listener" role helped some participants feel comfortable sharing memories they had not discussed in years. The conversational structure helped participants organise fragmented memories into coherent narratives, and several reported the process itself felt therapeutic. The generated images served as powerful emotional triggers — even when not perfectly accurate, they evoked additional memories and prompted further storytelling. However, significant challenges emerged. Cultural and historical accuracy in text-to-image generation was poor: images of 1960s rural China often contained anachronistic elements (modern clothing, Western-style architecture), and participants found these inaccuracies jarring and sometimes upsetting, as they misrepresented formative life experiences. ChatGPT sometimes asked questions that were culturally insensitive or that touched on painful memories without appropriate sensitivity — the AI lacked the emotional intelligence to recognise when a participant was becoming distressed and adjust its approach. Participants with lower technology literacy needed significant researcher mediation to interact with the system, raising questions about independent use. The paper proposes design guidelines including: using culturally and historically grounded image models, implementing emotional sensitivity detection in conversational AI, providing participants control over the depth and direction of reminiscence, and designing for mediated rather than fully independent use with older adults.
Relevance
This paper sits at the intersection of ageing, AI, and accessible design. For accessibility practitioners, it highlights both the promise and risks of deploying generative AI with older adult populations. The promise is significant: AI can serve as a patient, always-available conversational partner that supports cognitive and emotional well-being through structured reminiscence — particularly valuable for older adults who are socially isolated or who have limited access to human listeners. The risks are equally important: AI hallucinations in image generation and cultural insensitivity in conversational prompts can cause real harm, particularly when dealing with emotionally significant memories. The paper reinforces that AI tools designed for older adults must account for varying technology literacy, cultural context, and emotional vulnerability. Limitations include the very small sample size (seven participants), the researcher-mediated interaction (which masks independent usability challenges), and the specific Chinese cultural context, but the design guidelines around emotional sensitivity, cultural accuracy, and user control are broadly applicable to any AI-mediated accessibility tool for older adults.
Tags: older adults · generative AI · reminiscence therapy · well-being · storytelling · text-to-image · conversational AI