Launching Health in ChatGPT

The real story behind Health in ChatGPT

OpenAI has begun rolling out Health in ChatGPT to U.S. users, letting people securely connect their medical records and Apple Health data so the model can reason over their actual history. On the surface it looks like a convenience feature: ask about a lab result, prepare for a doctor's visit, get a straight answer about your insurance options. Useful, but easy to dismiss as one more chatbot skill.

At Skylab we think that framing misses the point. The feature is small. What it signals about the next decade of healthcare software is not.

A thirty-year data problem, quietly conceded

For as long as digital health records have existed, they have failed at the one thing they were supposed to do: move. Every hospital, clinic, and lab bought a system. Almost none of those systems could speak cleanly to the next one. The data was captured, then stranded, locked inside schemas that refused to agree with each other about what a patient, a diagnosis, or even a date should look like.

The result is a workflow most people would not believe if they had not seen it. Clinicians print records and re-fax them, because paper is easier than getting two databases to reconcile. Patients carry folders between specialists. Interoperability standards were written, mandated, and then quietly worked around. The information existed the whole time. It simply was not legible.

This is the backdrop that makes the ChatGPT announcement more interesting than it first appears. The industry did not solve the schema problem. It found a way to route around it.

Why AI is suited to messy records

Traditional integration assumes structure. You map field to field, enforce a shared format, and reject anything that does not fit. That approach has consumed enormous budgets in healthcare IT and delivered far less than promised, because human medical data is stubbornly unstructured. Notes contradict each other. Units differ. The same condition appears under five different labels.

Large language models invert the assumption. They do not need a shared schema to find meaning. They read inconsistent, human text and extract what matters from it, which is exactly the shape of the problem that defeated a generation of integration projects. The very quality that made health data useless to conventional software, its lack of structure, is the quality modern AI handles best.

That is the shift worth naming. We are moving from a world where data had to be perfectly formatted before it was useful, to one where imperfect data can be interpreted on demand.

The longer arc: from records to a map of health

Now extend the trajectory. Wearables already track heart rate, sleep, and activity continuously. Layer on weight, diet, medication, and lab work, then multiply across millions of people over years. For the first time, medicine has the raw material for something it has never truly had: an evidence-based picture of what a healthy life actually looks like, drawn from real outcomes rather than from guidelines built on small studies.

The value is not the single conversation about one test result. It is the compounding dataset underneath it, and the fact that AI can finally read it. That is where genuine breakthroughs in prevention and personalized care will come from.

Where caution is the right instinct

None of this argues for moving fast and breaking things. Healthcare should move deliberately, because the cost of an error is measured in lives, not in churn. The models are still wrong sometimes, and a confident answer is not the same as a correct one. Consent, privacy, and clear boundaries on how health data is used matter enormously. OpenAI's own framing, that connected health data is not used to train foundation models or to target ads, exists precisely because trust is the gating factor here.

So the right posture is neither hype nor dismissal. It is clear-eyed attention. The direction is set, and it is not toward a smarter chatbot. It is toward the moment health data finally becomes legible, and what a well-designed system can do once it is.

How Skylab thinks about this

We build with AI for organizations that cannot afford to be careless with data, and the lesson generalizes well beyond healthcare. Most enterprises are sitting on the same problem in a different costume: valuable information trapped in incompatible systems, waiting for a costly integration project that never quite finishes. The opportunity now is to interpret that data where it lives rather than forcing it into one rigid schema first.

If your organization is sitting on data it cannot yet use, that is the conversation worth having. Let's talk.