How we use AI in health apps, and the one place we keep it out of
Almost every health app in the store now says it uses AI. Very few say what for, which is convenient, because in a lot of them the honest answer is a chat box bolted to the side. Here is the specific version for our apps: the three jobs the model actually does, where each one runs, and the job we deliberately will not give it.
We use AI for three things: reading a photo of a meal into an ingredient list, condensing months of logs into a one-page brief for an appointment, and finding patterns in your own data. Some of that runs entirely on your phone; where it does not, the app's privacy policy names exactly what is sent. We do not use it to diagnose, to estimate the odds you have a condition, or to interpret results. Those are a clinician's job.
Job one: reading what you photograph
Typing a meal into a food diary is the reason food diaries get abandoned in week two. It is slow, and it is also inaccurate in a specific way: you write down chicken and rice, not the marinade, the stock cube, the thickener or the onion in the sauce. For someone hunting a food trigger, the ingredient you did not think to write down is very often the one that matters.
So in Velora you photograph the plate. The image is compressed on your phone, sent for analysis, and broken into ingredients, with FODMAP and histamine tags where they apply. The image is analysed and then discarded: what is kept is the ingredient list, not the picture.
This is the one job where a model is not a nice-to-have. No amount of interface design makes typing out every hidden ingredient tolerable twice a day, and without those ingredients the correlation engine downstream has nothing worth correlating.
Job two: condensing months into one page
The second job exists because of a structural problem with chronic illness care: you live with the condition for 90 days and then get 15 minutes to describe it. People arrive with a phone full of notes and leave having discussed the last bad week, because that is what memory serves up first.
Several of the apps produce a monthly brief that compresses everything recorded into a single page written to be read in under a minute: Dew for Sjogren's, Ember for fibromyalgia, Mello for gastroparesis, Willow for lipedema, and on the veterinary side PawBeat and RenalPaw.
What the brief contains is your data, summarised and ordered. What it does not contain is an opinion about what your data means. That distinction is the whole design of the feature, and it is the reason the reports are safe to hand over: a clinician reading one is reading you, not us.
Job three: patterns, without leaving the phone
The third job is finding what tends to precede a bad day. Aster and Aura do this on the device itself, which means no upload, no account, and nothing sitting on a server to be breached later.
Running locally costs something: a phone is not a datacentre, so the analysis is simpler than it could be. We think that is the right trade for a cold sore log or a menopause diary, where the sensitivity of the data outweighs the sophistication of the maths.
What leaves your phone, per app
| What you do | What leaves the device | Where it is stated |
|---|---|---|
| Photograph a meal in Velora | The compressed image, analysed then discarded | Velora privacy policy |
| Generate a report in Willow | Aggregated numbers only. No name, no raw daily logs, no photos | Willow privacy policy |
| Use trigger analysis in Aster or Aura | Nothing. It runs on the phone | Aura privacy policy |
| Track in SebDerm | Nothing. The app has no networking code at all | SebDerm privacy policy |
There is no single sentence that covers all of the apps honestly, which is why there is no single sentence here. Each app's own policy is the authority for that app, and where an app sends something we would rather name it than hide behind a portfolio-wide claim.
The job we will not give it
None of our apps diagnose anything. Not a symptom checker, not a percentage likelihood, not a ranked list of conditions you might have, not an interpretation of a lab result. This is a deliberate product decision and not a limitation we are working around.
The reason is a specific failure mode. A model that produces confident, fluent, well-formatted medical text is extremely persuasive, and it is wrong often enough to matter. The dangerous outcome is not someone being told something incorrect; it is someone being reassured and not making the appointment. An app cannot examine you, cannot order a test, and does not carry the consequence of being wrong.
So the apps run the other way. Several of them will interrupt what you were doing to tell you to contact someone: Diverticulitis Coach shows red flag guidance rather than a food list when you report the symptoms that need urgent care, and PawBeat is built around a threshold that means call the vet now.
How to read the word AI on any health app listing
Since the label is now on almost everything, three questions separate a feature from a sticker:
- What job does it do that a form could not? If the answer is that it rephrases what you typed, it is decoration.
- Where does it run, and what is sent? A privacy policy that names the data and the processor is a good sign. One that says we may use third-party services to improve your experience is not an answer.
- Does it tell you what is wrong with you? If it does, be more careful with it, not less, however confident it sounds.
We would rather you applied those three questions to our apps than took our word for any of this. The privacy policies linked above are written to survive that reading.
Frequently asked questions
Do Velora Health apps use AI? +
Yes, for three specific jobs: reading a photographed meal into an ingredient list, condensing months of logs into a one-page brief for an appointment, and finding patterns in your own data. Some of that runs entirely on the device.
Can an AI health app diagnose my condition? +
Ours will not, by design. None of our apps diagnose, estimate the likelihood of a condition or interpret results. A model producing fluent medical text is persuasive and wrong often enough to matter, and the risk is that someone is reassured and does not seek care.
Does my health data get sent to an AI company? +
It depends on the app and the feature, which is why each app's privacy policy states it specifically. Trigger analysis in Aster and Aura runs on the device with nothing uploaded, SebDerm has no networking code at all, Willow sends aggregated numbers with no name or raw logs, and Velora sends a compressed meal photo that is analysed and then discarded.
How does the meal photo recognition work? +
You photograph the plate, the image is compressed on your phone and analysed, and it comes back as a list of ingredients with FODMAP and histamine tags where they apply. The point is the hidden ingredients, the marinade or the thickener, that nobody types into a food diary by hand.
Is the AI report written by a doctor? +
No. It is your own recorded data, summarised and ordered into a page that is quick to read. It contains no clinical opinion, and it is designed to be handed to your own clinician rather than to replace them.
Is an AI symptom summary accurate enough to show a doctor? +
It is a summary of what you recorded, so its accuracy is the accuracy of your own logs. That is exactly why it is useful in an appointment: it replaces trying to recall three months in a waiting room, and the clinician is still the one interpreting it.
This article is about how Velora Health works, not medical advice. Our apps are wellness and self-management tools; they do not diagnose, treat or replace care from a qualified clinician.