A run of consumer AI health products landed in the first half of 2026. Amazon's in January. Microsoft's Copilot Health and Perplexity's health platform in March. In May, Google rebranded Fitbit Premium as Google Health Premium and bolted a Gemini coach onto it.
The pitch is the same in each case. Connect your wearable, your lab results, maybe your medical records, and a friendly model will tell you what it all means. It is a good pitch. It is also worth knowing what happens between the moment you ask and the moment the advice comes back.
What the coach actually does
Strip away the branding and the mechanism is unremarkable. Your numbers — heart rate, sleep, glucose, whatever you've connected — leave your phone and travel to a server. A large language model reads them alongside your question and writes an answer.
The interesting part is how it writes that answer. The model does not look your symptoms up in a curated, clinically vetted database and report what it finds. It estimates the most likely next words, given everything it was trained on. Most of the time that lands somewhere sensible, because sensible things are common in the training data. But it is predicting, not checking. If something false sits in your prompt, or in the sources it learned from, it can hand the falsehood back to you with total composure.
When the guess is wrong
The companies building these tools know this, which is the part worth sitting with.
Apple spent the run-up to 2026 building an AI health coach — codenamed Quartz, meant to ship inside a paid Health subscription — then scaled it right back. As Bloomberg reported in February, it dropped real-time dietary coaching, mental-health check-ins, and third-party provider integration. Two reasons surfaced. Apple's lawyers worked out that some features would be classed by the FDA as a medical device, with the approval timeline that implies. And internal testing turned up reliability problems serious enough to give a company that ships nearly everything pause.
The published research is blunter. A Mount Sinai study found that one health chatbot under-triaged more than half of medical emergencies in structured testing — quietly routing cases that needed an ambulance toward rest and monitoring. A Nature Medicine study found that people using chatbots to work through symptoms did no better than a control group left with their own resources, and that the same tool could give conflicting advice for identical symptoms depending on how the question was phrased.
A model that under-triages half your emergencies isn't a coach. It's a confident stranger holding your bloodwork.
Who's holding the data while it learns
There is a second cost, and it is quieter. By March, KFF's tracking poll found that about a third of adults had used AI for health information, and four in ten of those had uploaded personal medical information to get a more tailored answer. That is a great deal of bloodwork, body weights, and mental-health notes moving onto servers in a single quarter.
The privacy lawyers noticed the speed. The IAPP summed the period up flatly: five products in three months, and the privacy questions that got left behind. Most of these services promise not to train their core models on your health data, and several genuinely mean it. But "de-identified, aggregated" usage still flows into product telemetry by default, and de-identified health data has a long history of being re-identified by anyone sufficiently motivated. The setup screen offers an opt-out. Most people scroll past it in the four seconds it takes to start asking questions.
The part you can keep, the part you can leave
None of this makes the tools useless. A model that summarises a year of sleep data is genuinely handy, and some of these products are carefully built. The point is narrower. When you ask the coach in the cloud what your numbers mean, you are sending your most private information to a server so that a prediction engine can guess — and the guess is wrong often enough that the company that makes the iPhone got cold feet about shipping its own.
You can keep the receipts without the round trip. Log the numbers somewhere they don't leave your device. Look at them yourself on a Sunday. The model on the server is the part you can take or leave, and most weeks you can leave it.
OneFileClub makes powerful one-file tools for keeping your own receipts — habit trackers, training logs, food diaries, sleep logs. One file. One payment. Your data stays on your device, where nothing has to phone home to read it.