Balancing Safety and Autonomy: Accessibility-Oriented Interventions in Generative AI for Cognitive Impairment

Key Findings

• — Participants were classified by clinical assessment as mild (n=22), moderate (n=12) and severe (n=11) impairment
• — Medical question-answering assistants were the most frequently encountered system type (mentioned by approximately half of participants); 28 used smart health-management apps and 12 used conversational interfaces designed for emotional companionship, mainly those with mild impairment or living alone
• — Some systems limited continuous conversation, interrupted prolonged use or redirected users to family and offline activities; one participant described a clinician-recommended chat interface where 'the robot will actually push me away'
• — Users with mild impairment reported safety-oriented restrictions as a loss of autonomy, for example being told to consult family or 'ask a doctor' for decisions they felt able to make; one said 'I am an adult, and I should have my own dignity'
• — As impairment progresses, tensions between safety and autonomy become less visible because user participation diminishes and decisions move to caregivers, not because the tensions are resolved

Methodology Notes

Semi-structured interviews (35-55 minutes) with participants recruited by purposive sampling through community health service centres and collaborating hospitals; doctors screened eligibility; written consent; institutional ethics approval; participants paid 15 dollars. Caregivers served as proxy informants for participants who could not be interviewed. Systems were self-reported and span LLM-based medical Q&A assistants, cognitive-training dialogue tools, health-management apps and companionship interfaces; no system logs were analysed. The recruitment country is not stated in the text (limitations: 'specific geographic contexts'). Single time point; retrospective accounts; proxy accounts add interpretive distance. Open access (CC BY 4.0). Verification: VoR abstract via OpenAlex (identical to arXiv v1); full text read from arXiv HTML v1 (2608.17175), which carries the ASSETS DOI; dl.acm.org is behind a Cloudflare challenge from this host.

Sources