A Framework for how People use Generative AI for Self-Care
Domain
Consumer AI, Wellbeing, Digital health
Problem
Trust, privacy, AI dependency, personalisation
Methods
Diary study, first-person research, interviews, workshops
People are already using general-purpose AI to take care of themselves, without anyone designing for that use — so using researcher diaries, a two-week diary study, and interviews with 24 participants, we mapped five distinct ways this is happening and surfaced the trust, privacy, and dependency risks that no product team has yet addressed.
The Problem
People are already using AI to take care of themselves. Not because a product was designed for that purpose, but because tools like ChatGPT are available, conversational, and — crucially — non-judgmental. The self-care AI market is attracting significant product investment. But very little is known about what people actually do when they use general-purpose AI for their health and wellbeing — what they ask, what they trust, what helps, and what risks they take without knowing it. We set out to find out. Not through surveys or hypothetical scenarios, but by working alongside people across three weeks of real use — watching, listening, and asking hard questions about trust, privacy, and the line between useful and problematic.
Who we worked with
Five members of the project team — including Tara Capel from the University of Edinburgh — began with a first-person approach, using AI tools for their own self-care and documenting their experiences. This is a methodological choice we stand behind: it means our team understood the tools from the inside, and it ensured we acknowledged our own perspectives rather than pretending to a false objectivity. We then recruited 24 participants across a broad range of ages, occupations, and self-care concerns — students and working adults across Australia, aged 19 to 64, with 15 participants born outside Australia. Self-care concerns ranged from pregnancy, sleep and chronic health conditions to work-life balance, emotional regulation, and social anxiety.
How we did it
- 01
Workshops
Participants explored how they might integrate AI with their existing self-care, with examples ranging from asking for coping strategies to generating images to express feelings they could not put into words.
- 02
Two-week diary study
Participants used AI tools daily and kept brief, low-burden diary entries. We deliberately kept the format minimal so the diary did not interfere with the self-care itself.
- 03
Follow-up interviews
One-on-one conversations reviewing diary entries and probing hard questions about trust, accuracy, privacy, and what participants would change.
What we found
Finding 01
AI as personalised advisorThe successor to Dr. Google
The most common use was advice-seeking — asking ChatGPT specific, personal questions and receiving tailored responses. 22 of 29 participants did this. They asked about sleep schedules, pregnancy nutrition, coping strategies for anxiety, how to handle difficult workplace conversations, and potential diagnoses for unexplained symptoms. One participant, running out of options after two years of undiagnosed symptoms, used ChatGPT to generate a differential diagnosis list she then took to a specialist appointment. Personalisation depends entirely on what people choose to disclose — broad questions returned generic answers; specific, contextualised prompts returned genuinely useful ones.
If you are building AI for health or wellbeing
We can design and run research that examines how your specific users experience the gap between AI personalisation and AI accuracy — and what that means for trust, disclosure, and sustained engagement with your product.
Finding 02
AI as ongoing mentorAnd where that becomes complicated
Five participants used AI not for one-off questions but for sustained, evolving conversations about a self-care concern over days and weeks. One used ChatGPT as a triathlon coach. Another turned to it as a counsellor she could reach in the middle of the night when her therapist was unavailable. One participant began using Replika as a virtual friend and, over time, found herself texting her human friends less. Her friends noticed. She had substituted the easier, frictionless relationship for the harder, more complicated ones. The therapeutic potential and the risk of AI companionship are not separate questions — they are the same question.
If you are building AI for mental health or emotional wellbeing
We design research programmes specifically structured to examine how sustained use changes the relationship between your product and your users — including what happens to human relationships alongside it.
Finding 03
AI as creative resourceMaking things people could not make alone
12 participants used AI's generative capabilities to create personalised self-care resources: guided meditations, children's bedtime stories, sleep schedules, prayer plans, colouring images, and music. One parent generated a bespoke bedtime meditation story for her four-year-old, adapted to her daughter's sense of humour. Another participant composed music with an AI tool to work through emotions she found hard to name — describing it as a judgment-free jam.
If you are building creative or expressive AI tools
We can research how your users navigate the relationship between personal relevance and creative distance in AI-generated content — and translate those findings into interaction design that gets the balance right.
Finding 04
AI as expressive mediumFeeling in images and music, not words
Nine participants used AI to express and reflect on their emotional states through generated images and music. One researcher created a visual mood diary — generating an image each day that reflected how he felt, mapped onto a digital whiteboard over two weeks. He had previously tried writing his feelings down and found words inadequate. The iterative process of generating, reviewing and refining images deepened his understanding of his own emotional state in ways prose had not. The modality mattered — but our findings also caution against adding multiple modalities as a default. For advice-seeking and mentoring, text was often exactly enough. The value of images and music was specific to contexts where words could not carry the full weight of what participants needed to express.
If you are building multi-modal AI products
We can research which modalities genuinely serve your users' self-care needs and which add complexity without value — giving you an evidence base for product decisions that are otherwise made by assumption.
Finding 05
The trust questionWhat people believe, what they check, and what they do not
Participants found AI-generated content largely accurate — more than the research team expected. But trust was not uniform, and its limits were instructive. Participants contested AI advice when they had the expertise to notice it was wrong. One researcher identified AI-propagated misinformation about seed oils and challenged it. But she only noticed because she had relevant background knowledge. Most users do not. Cultural context was a consistent blind spot. Unless explicitly prompted, AI defaulted to Western assumptions about food, family structures, and what healthy looks like. Participants from non-Western backgrounds had to actively work around this. Privacy concerns were present but did not prevent engagement — people shared intimate health information with AI they would not share with the research team, because the AI felt non-judgmental. This is both the product's strength and a significant risk.
If you are deploying AI tools across diverse user populations
We can research how cultural context shapes your users' experience of AI advice and build the evidence base you need to address normative bias before it becomes embedded in your product's design.
“I feel like I'm vulnerable in relaying my feelings towards an AI.”
— Participant, reflecting on privacy after sharing personal health information with ChatGPT
Working with Forth Story
If you are building or deploying AI products for health, wellbeing, or consumer self-care contexts and want research that goes beyond usability to understand what your product does to the people who use it — get in touch. We worked with a wide project team to deliver this work.