Mapping Caregiver Needs to AI Chatbot Design: Strengths and Gaps in Mental Health Support for Alzheimer's and Dementia Caregivers
Jiayue Melissa Shi, Dong Whi Yoo, Keran Wang, Violeta J. Rodriguez, Ravi Karkar, Koustuv Saha · 2026 · ACM Transactions on Computing for Healthcare · doi:10.1145/3803549
Summary
Family caregivers of individuals with Alzheimer's Disease and Related Dementias (AD/ADRD) face significant mental health challenges — including stress, anxiety, depression, compassion fatigue, anticipatory grief, and burnout — yet have limited access to timely, affordable support. This paper investigates how AD/ADRD caregivers perceive and engage with an AI-powered mental health chatbot, using a prototype called Carey (built on GPT-4o) as a technology probe to surface early design insights rather than evaluate real-world effectiveness. The researchers conducted scenario-driven, semi-structured interviews with 16 family caregivers recruited via AD/ADRD online communities (Reddit, alzheimers.net, alzconnected.org). Participants interacted with Carey across eight theory-grounded caregiving scenarios — covering disruptive care-recipient behaviour, lack of support, low self-efficacy, emotional distress, relationship tensions, compassion fatigue, lack of self-care, and burnout. Following each interaction, caregivers discussed their reactions, expectations, and concerns. Analysis used inductive coding and reflexive thematic analysis, generating 420 codes grouped into 24 subthemes and five higher-level themes. The five core themes identified were: on-demand information access, safe space for disclosure, emotional support, crisis management, and personalization. The study also examined privacy and data security as a cross-cutting concern. For each theme, the authors mapped caregiver needs, Carey's demonstrated strengths, current gaps, and actionable design recommendations — summarized in a comprehensive mapping table that can guide future AI system design for caregiver-facing mental health tools. Participants were demographically diverse (age 19–65+, varied race, education, and caregiving duration), and the study was approved by institutional review boards. One participant showed signs of emotional distress during the session, underscoring the sensitivity of the research context.
Key findings
Fifteen of 16 caregivers rated Carey as "helpful" or "very helpful" for addressing caregiving-related mental health concerns. Emotional distress was the most commonly selected scenario (12 participants, 28 interactions), followed by disruptive behaviours (8 participants, 15 interactions). Six key themes emerged with nuanced tensions in each: (1) Caregivers valued Carey's on-demand information access but expressed concerns about source verifiability, hallucination risk, and generic responses. (2) They appreciated Carey as a judgment-free safe space for disclosure, but wanted more natural conversational reciprocity and proactive follow-up questions. (3) Emotional support was valued but felt superficial — caregivers wanted depth of engagement beyond validation, and worried about AI replacing meaningful human connections. (4) Carey's 24/7 availability was seen as emotionally stabilizing for crisis management, but participants flagged the absence of escalation pathways to professional help. (5) Personalization felt surface-level — caregivers wanted long-term contextual memory and adaptation to the evolving caregiving journey, not just session-level tailoring. (6) Privacy perceptions were complex: caregivers felt safer disclosing to AI than humans, but were hesitant about persistent storage of sensitive family information, and desired ephemeral data options. Design recommendations include transparent source citation, conversational turn-taking, tiered crisis intervention mechanisms, consent-driven longitudinal memory, and flexible privacy controls.
Relevance
This paper is directly relevant to accessibility and inclusive design practitioners working on AI systems for older adults, people with dementia, and their support networks. It demonstrates that effective AI mental health tools must go beyond information delivery to address emotional labour, relational dynamics, and the psychological complexity of chronic caregiving. The finding that caregivers value AI as a low-burden, on-demand complement — not a replacement — for human support has clear implications for how AI assistive tools should be positioned. The tension between personalization and privacy is particularly important: caregivers wanted contextually aware support but were deeply protective of care-recipient data, suggesting that opt-in ephemeral memory and clear data transparency are accessibility requirements, not optional features. Limitations include a non-representative convenience sample, scenario-driven rather than naturalistic interactions, and no measurement of clinical outcomes pre/post use. Future work should include longitudinal deployment studies and validated outcome measures (e.g., PHQ-9 for depression, Zarit Burden Interview). The study's design framework and need-mapping table offer a practical starting point for any team building caregiver-facing AI tools.
Tags: AI chatbots · dementia caregiving · mental health · caregiver wellbeing · conversational AI · large language models · human-centered design · technology probe