ChatGPT Health Is Interesting

Not because it “fixes healthcare” but because it changes what’s possible to do with health data

The ChatGPT Health announcement landed and the reactions were predictable. Some people went straight to “this changes everything.” Others went with “we’ve seen this before, it’s just a PHR in new clothes.” Both takes are partly right, but they skip the part that matters.

What is actually new here is not “health data in an app.” We have had that for years. What is new is the idea that health data might show up inside a tool people already use daily, in a format that supports reasoning instead of just viewing. That is a real shift, and it is worth being optimistic about. It is also worth being honest about what it will not solve.

Start with the boring constraint. Even if the connection becomes automatic, this is not going to be your full, longitudinal record. The data that typically flows through patient access APIs is constrained by USCDI and US Core style datasets. That is a minimum set, not “everything in Epic.” It is also filtered by what providers choose to share and what they are technically capable of exposing. Then you run into the real-world fragmentation that patients feel but the product screenshots never show. Claims history does not reliably follow you when you change payers. Specialty systems live outside the core EHR. Referrals and scheduling often sit in separate platforms. External labs and imaging can be messy. Multi-owned groups can split your story into separate fiefdoms. So yes, there will be data. No, it will not be complete. That does not make the whole thing pointless, but it sets the boundary conditions for what “AI on your record” can safely claim.

Now for the part where I disagree with the cynical take.

PHRs were mostly built for display. Tabs, charts, PDFs, and a vague promise that visibility would lead to better care. In practice, they gave patients a new place to look at information and then handed them the same work they always had. Figure out what matters. Decide what to do. Track down the right office. Make the calls. Chase the referral. Fill out the forms. Remember the prep instructions. Show up. Hope the right data made it to the right place.

ChatGPT changes the interaction model. When you put health data into a reasoning system, you stop asking patients to interpret raw artifacts. You can ask better questions in plain language. You can request different views of the same record depending on what you need in the moment. You can get explanations that adapt. That is a meaningful step beyond “here are your labs, good luck.”

I have already seen a preview of this in my own experimentation. I built a small prototype called Plumly to test AI summarization on top of messy health data. The surprising thing was not that it could summarize. Any decent model can do that. The surprising thing was how many reasonable summaries the same record could produce depending on framing. You can write a clinical handoff. You can write a plain-language explanation. You can highlight risk. You can focus on deltas since last visit. You can make it a checklist. None of those are inherently wrong, but they lead people toward different next steps. That is where the hard part starts.

Summarization is not the challenge. Clinical accuracy is. Responsible framing is. Knowing when to stop is. This is going to be the needle ChatGPT Health has to thread. If it is overly cautious, it will feel like a polite FAQ and people will stop using it. If it is too confident, it will wander into advice it cannot safely give. The difficult middle ground is explaining uncertainty without being useless, and giving guidance without pretending to be your doctor. There is no UI trick that solves that. It is a product and safety problem.

Here is the second place where the hype takes drift away from reality.

Even if ChatGPT nails the reasoning layer, that still does not fix the thing that makes patients crazy. Healthcare is not primarily broken because people cannot understand their record. Healthcare is broken because nothing is wired to completion.

I tried to schedule a colonoscopy recently. It took a week, five phone calls, a referral chase, and a form that felt like a DMV exam. I did not need an AI to explain what a colonoscopy is. I needed someone, or something, to coordinate the work without me acting as the integration layer.

This is why I keep coming back to scheduling. Scheduling is the black hole of healthcare value. It swallows intent, motivation, and even clinically correct recommendations. You can tell a patient exactly what they need and why it matters, and still fail because the step from “you should” to “it is booked” is a maze. That maze is not solved by a better summary.

The uncomfortable part is that the system already knows a lot of what people need. Care gaps are computed. Quality engines run nightly. EHRs and population health tools identify who is overdue and what is next. We do not need a model to rediscover that someone is overdue for a screening. The opportunity is turning those insights into execution. Find an in-network option, propose times, generate the referral if required, send the prep instructions, close the loop, confirm completion. That is the difference between information and outcomes.

There is one unglamorous dependency that sits underneath all of this, and it is not discussed enough. Provider directories. If you cannot answer who can do the service, where, when, and at what cost with what coverage, then “knowing what you need” has limited value. Directory data is messy. Networks are messy. Availability is messy. Until that gets addressed in a programmatic way, most “next steps” will still end in a phone number.

So where do I land on this announcement?

I am optimistic, but it is the expensive kind of optimism. I think plumbing health data into a reasoning environment is a real shift. I think it will produce genuinely useful summaries and explanations for patients, even if the dataset is incomplete. I also think it will expose the gap that PHRs never crossed. Reasoning is helpful, but it is not enough. The prize is coordination. The prize is completion. The prize is consented agency that can act across systems rather than just describe them.

That is why I want to try ChatGPT Health as soon as the waitlist opens. Not because I think it fixes healthcare, but because it will test a better question than “can patients access their data.”

Can health data become something you can think with, and then actually use to get care done?

If the answer is yes, even partially, then this is bigger than another portal. If the answer is no, then we just built a nicer interface for the same old maze.

Either way, we are about to learn a lot.