What FemTech gets Wrong about Diverse Users, and What it Means for AI
Domain
Femtech, Health AI, Consumer AI
Problem
Personalisation, inclusion, bias, marginalised users
Methods
Cultural probes, interviews, card sorting, feminist HCI
Personalisation is one of AI's most compelling promises and one of its most frequently broken ones — so using cultural probes, card sorting, and interviews with nine LGBTQ+ participants living with menstrual health conditions, we examined where period tracking apps perform personalisation without enacting it, and found a failure pattern that applies to any AI system that solicits user disclosure without genuinely responding to it.
The Problem
Personalisation is one of AI's most commercially compelling promises — and one of its most frequently broken ones. This research examined what happens when health tracking tools claim to personalise but do not, working with people routinely excluded by the default design of menstrual tracking apps: those who are LGBTQ+, those living with menstrual health conditions including endometriosis, adenomyosis and PCOS, and those who are both. What we found has direct implications for any AI system that promises personalisation as a feature.
Who we worked with
Nine participants who identified as LGBTQ+ took part, six also reporting menstrual health conditions, aged 18 to 34. We used three methods designed to surface what fixed-choice questions could not — the sorting exercise in particular was generative, with what participants said as they sorted revealing more about identity, health goals and system assumptions than any survey could.
How we did it
- 01
Cultural probe
Participants reviewed and annotated commercial period tracking apps over one to two weeks, positioned as designers contributing to a femtech startup. This reframing shifted them from users tolerating gaps to critics naming them.
- 02
Semi-structured interviews
90 to 120 minute conversations unpacking probe outputs and exploring what meaningful personalisation would look like for each participant's body, identity and health context.
- 03
Card sorting activity
Participants sorted tracking metrics into what they personally needed, what others like them would need, and what had no place in the app at all — revealing the implicit assumptions embedded in current feature sets.
What we found
Finding 01
The illusion of personalisationDisclosure without response
The most consistent frustration was a specific experience: spending significant time answering onboarding questions, only to find that none of the answers changed anything. One participant spent ten to fifteen minutes working through setup — disclosing preferences, answering health questions — only to find the app entirely unchanged when she entered it. Another, asked to disclose her menstrual health conditions, found none of the six listed options matched her diagnosis of adenomyosis. The only alternative was 'none of the above', which registered her as having an unaffected cycle. The apps solicited disclosure without creating any back-end mechanism to respond to it. This is not a femtech problem — it is an AI personalisation problem. Any AI system that asks users to disclose preferences and then fails to visibly respond to those disclosures will produce the same outcome: distrust, abandonment, and a sense of having been manipulated.
If you are building AI that personalises
We can design research that examines whether your personalisation is real — whether user inputs are actually changing outputs in ways users can detect and value — before those gaps become embedded in your product.
Finding 02
A singular view of the userWhen the default user is the only user
Every app participants reviewed assumed the same user: cisgender, heterosexual, in a relationship with a man, and interested in pregnancy. Every participant fell outside this picture in at least one significant way. Queer participants encountered apps where a shared tracking feature for couples was specified — in the app's own FAQ — as available only for heterosexual couples. LGBTQ+ content, when it appeared, was sexualised and frequently paywalled, positioning queer identity as a monetised bonus rather than a baseline consideration. Participants with complex cycles found themselves defaulted into a residual 'track my period' category that changed nothing about the fertility notifications, pregnancy prompts, and ovulation content they continued to receive. The singular user assumption is not a design aesthetic problem — it is a model problem. Any AI system optimised for a dominant user profile will systematically underserve everyone outside it.
If you are building AI for diverse user populations
We can research who your actual users are, where your system's implicit assumptions about the default user are generating exclusion, and what genuine inclusivity would require — before those assumptions scale.
Finding 03
Tracking that does not trackFeature abundance is not adaptivity
Participants with complex or irregular cycles found the available tracking metrics simultaneously too broad and too shallow — full of generic inputs like alcohol intake and travel logs, while lacking the nuance to document their actual health. One participant with adenomyosis could not select both medium flow and blood clots simultaneously, despite that combination being medically significant for her condition. Another found her contraceptive method — an intravenous injection — not listed at all. Several encountered alarmist irregularity warnings that turned out to be paywalled PCOS assessments — with no mechanism to disclose an existing diagnosis and disable them. An AI health tool that offers a hundred tracking options but cannot adapt its output to a user's disclosed condition is not personalised. It is comprehensive and useless.
If you are building AI health or tracking tools
We can research what meaningful depth looks like for your specific user population — and what the difference is between feature abundance and genuine adaptivity.
Finding 04
The broader lessonPersonalisation as a category of trust
This research is, on the surface, about period tracking apps. But its findings apply to any AI system that makes personalisation claims. A system that collects user preferences and produces identical outputs regardless is not intelligent — it is a static template with a conversational interface. And when users have shared sensitive personal information on the assumption that doing so will help the system help them, the failure to respond is not just a product flaw. It is a specific kind of betrayal.
“It felt like any of the information I put in… it just wasn't listened to, and everything that spat out was just generic.”
— Participant, reflecting on a period tracking app after completing onboarding
Working with Forth Story
We worked with a wide project team to deliver this work. More details can be found in our academic outputs.