Shared Stories, Shared Bonds: People with Dementia Exploring Generative AI Together
Teis Arets, Maarten Houben, Fleur van Haeren, Wijnand IJsselsteijn, Giulia Perugia · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3791055
Summary
This paper reports on a qualitative workshop study examining how people living with mild to moderate dementia engage with three generative AI tools - Microsoft Copilot (text), Midjourney (images), and Suno (music) - in a group setting, and what forms of social connectedness emerge from that engagement. Rather than positioning GenAI as a one-on-one reminiscence delivery mechanism (the dominant framing in prior dementia-tech work), the authors treat it as a shared medium and investigate whether people with dementia can actively shape, critique, and negotiate its outputs together. Seventeen participants across six workshops at three Dutch daycare facilities co-created prompts using stacks of tangible cards (colors, objects/animals, environments, emotions, plus a wildcard), with moderators entering the composite prompts into the models while participants discussed and refined results on a shared screen. Four formal caregivers were interviewed afterwards as a triangulation source, with care taken to keep caregiver perspectives secondary to the participants' own voices. The data - video, audio, and transcripts enriched with behavioural annotation (gaze, posture, mirroring, gesture) - was analysed using reflexive thematic analysis following Braun and Clarke, with the first and third authors maintaining reflexive research diaries. The paper frames the work within participatory HCI traditions in dementia research and contributes design guidelines for ethically responsible GenAI targeted at this population.
Key findings
Three themes emerged. (1) Co-constructing and evaluating outputs through iterative prompting: participants actively judged GenAI outputs against personal memory and sense of 'what feels right' (verisimilitude) and, when satisfied, used quirky outputs as springboards for fantasy narratives; both modes drove further prompting. Examples included rejecting a windmill in favour of a traditional mill, and repeatedly refining an attic-window image that still missed the mark. (2) Building social bonds through reciprocal engagement: the negotiation space opened by GenAI surfaced personal memories (e.g., a pink-bike discussion triggering childhood reminiscence), collective musical preferences (a shared love of Fats Domino), and prosocial behaviours - caregivers noted participants P4, P5, P6 who 'normally wouldn't really seek each other out' were engaged in mutual conversation and laughter after sessions. Humour and self-mockery lowered barriers; participants with stronger cognitive capacity scaffolded peers who lost track. (3) Emerging understanding and critical awareness: initial awe gave way to discernment, with participants spotting inconsistencies and raising standards over time. Moderator scaffolding remained important but became more distributed peer-to-peer. The authors identify GenAI as a 'temporal anchor' that externalises conversation threads to compensate for working-memory loss - distinct from static media because outputs can evolve with shifting interpretations.
Relevance
For accessibility practitioners this is one of the clearest recent demonstrations that people with dementia can be active shapers of AI interactions, not passive recipients - a reframing with significant implications for how GenAI tools are designed and deployed in care settings. Four design guidelines are directly usable: (1) provide scaffolds (tangible probes, demonstrations) that can fade as skill grows; (2) keep prompts and outputs persistently perceptible; (3) implement fail-safes against harmful or stereotyping outputs; (4) support personalization through interpretation, not personal-data profiling - a privacy-preserving stance aligned with the authors' concern about vulnerable-population profiling. Limitations matter: all participants were Dutch and in mild to moderate stages, the workshop context was controlled, and moderators retained a prominent role, so autonomous use by people with dementia or applicability to later-stage dementia remain open. The paper's careful positionality statement and ongoing consent model are worth studying as a template for ethical accessibility research with cognitively vulnerable populations.
Tags: dementia · generative AI · social connectedness · cognitive accessibility · aging · participatory design · qualitative research · thematic analysis · reminiscence · shared reality · personhood · co-creation