AI Didn’t Replace My Doctor. It Replaced Everything In Between.

What a root canal, a heart nudge, and two weeks of low-grade panic taught me about using AI where it actually helps.

I finally had some time off after the most intense stretch of my career. My body took that as its cue to fall apart.

First came the tooth. A flicker of sensitivity to cold that I told myself was nothing, the way you do, right up until it stopped being nothing and started following me through the day like a bad mood. Then, because misfortune enjoys company, a long-overdue health check (the kind you keep postponing for roughly five years) came back with a polite but firm note about my heart. Eat better. Move more. Possibly reconsider the twelve hours a day I spend in front of a screen, which, given what I do for a living, felt less like advice and more like a lifestyle intervention.

So there I was. Off work, in pain, and mildly alarmed. And for the first time in a long time, I had the hours to actually pay attention to my own health instead of triaging it between meetings.

Here’s what I learned. Not that AI became my doctor. It didn’t, and you should be deeply suspicious of anyone who tells you theirs did. What actually happened is stranger and more useful.

The problem nobody names

Modern healthcare has a memory problem. Not the doctors. The system.

Every appointment starts from zero. You sit down, you try to remember when the pain started, which pills you’re on, what the last person said, what that report from eighteen months ago showed. The clock is running. The doctor is kind but busy. You leave with a plan you half-remember and a nagging sense that something got lost in translation. Then you move to a different clinic, or a different city, or a different specialist, and none of them can see what the others saw. You become the only thread connecting your own care, and you’re a terrible thread, because you’re anxious, forgetful, and not medically trained.

Call it appointment amnesia. It’s the space between visits where all the value leaks out.

That gap, the connective tissue around the diagnosis, is where AI turned out to earn its keep.

The dental saga, or: four dentists and one thread

It began as a whisper. Cold drinks set the tooth off, then even plain water did, the kind of thing you assume will settle on its own.

The first dentist did the sensible, conservative thing. Rather than reaching straight for a root canal, we tested it and put a filling at the base of the tooth, where the dentine was exposed. The hope was that would calm it down.

It did not calm down. It got worse. The sensitivity spread to heat, then hardened into a constant throb, and I settled into taking painkillers every single day. Not the holiday I’d had in mind.

By the third dentist, it took one quick look to name it: Cracked Tooth Syndrome. A hairline fracture you often can’t see, that behaves like a hypochondriac’s dream because it hurts inconsistently and hides from X-rays. The catch: that dentist was about to travel and couldn’t do the work.

That left me a choice I didn’t love. Sit in limbo for two or three weeks, or let someone less experienced do temporary work I wasn’t confident about, all with an international work trip closing in. So I took it into my own hands, found someone else equally qualified who could move quickly, and pushed (politely) for the option that preserved the tooth rather than the most drastic one. We landed on a root canal and a crown, finished before the trip, most of the original molar still mine.

Here’s the part that made all of it manageable. Through the whole saga, across all four dentists, I kept one long ChatGPT conversation running. Symptoms as they shifted. What each dentist said and did. What was tried, what worked, what didn’t. Dates, which tooth, sharp or dull, hot or cold.

That thread let me walk into each new chair as a participant instead of a stranger to my own case. I could hand over an accurate history in seconds, ask sharper questions, and understand the tradeoffs well enough to make decisions with the dentist rather than just nodding along.

Two things worth being honest about here.

The AI did not diagnose the crack. A dentist did, at a glance, with experience I will never have. What the AI did was hold the story together across four people who couldn’t see each other’s notes and X-Rays, and help me move faster and decide better when the system left me to join the dots myself.

And the quieter benefit, the one I didn’t expect: reassurance. When you’re mid-treatment and something aches at midnight, you can’t exactly ring your dentist and ask “is this normal or am I dying.” In the old days you’d either stew until morning or, worse, open a search engine and spiral into a diagnosis of something exotic and terminal. Being able to ask “is this expected at this stage” and get a calm, contextual answer, then verify it at the next appointment, took the edge off the waiting. That’s not a small thing. Anyone who’s lain awake catastrophising about their own body knows exactly how not-small it is.

From a pile of reports to a single page

The health check produced the opposite problem: too much information, not too little.

Lab reports. Historical check-ups. Scans and questionnaires, some as clean PDFs, some as photos of printouts, some as scanned pages that looked like they’d been faxed during a storm. On their own, each was a fragment. Together, they were a mess only I was motivated to untangle.

I fed them into a single ChatGPT project, one by one, adding new reports as they trickled back in, and asked it to do the boring, valuable work: pull out the findings, translate the medical language into something a human can hold in their head, compare this year against previous years, flag which files held duplicate data I could safely delete, and separate the genuinely reassuring from the few things actually worth raising with a doctor.

Then I asked for the output I’ve come to rely on most: a brief, and pointedly not a diagnosis.

The prompt is roughly this: turn all of this into a one-page, non-diagnostic summary I can show a doctor. Reason for the visit, timeline, current medications and supplements, relevant history, my top questions ranked by importance, and anything uncertain or worth flagging. Tone: “Here’s what happened. I’d value your judgement on X.”

I put that on a note on my phone and showed it at appointments. Most doctors, and I genuinely did not expect this, appreciated it. It saved everyone the ten-minute verbal fumble and gave them the signal without the noise. It’s especially useful when you move between hospital networks that don’t share records, which, if you live an expat life like mine, is most of the time.

Somewhere in this I realised the point of a health check isn’t the number of tests you run. It’s whether anyone actually interprets, prioritises, and acts on the results. A drawer full of reports nobody reads is just expensive paper. The AI didn’t add tests. It made the ones I already had usable.

It also helped me keep four different jobs separate, which I hadn’t realised were four different jobs: understanding a result, asking a professional for advice, making a decision, and monitoring what happens after. We tend to blur all of those into one anxious “what does this mean.” Pulling them apart made me a calmer, clearer patient.

The app I built, then rebuilt, then simplified

Here’s where the geek in me took over.

A while back I’d built myself a small web and mobile app that pulls in my Oura ring data through their API and lets me track headaches. Nothing fancy. But with these recent health nudges, I wanted more: daily meals, exercise, supplements, the works. One of my goals is to get more protein in while staying vegetarian and easing off cholesterol, which is its own small culinary puzzle.

A few years ago, building this would have meant a spec, a designer, a developer, and several weeks. Instead I described what I wanted to ChatGPT, in plain conversation, and evolved the thing on the fly. Add a feature. Remove one. Tweak another. When my thinking changed, the app changed with it. That flexibility is the real unlock. Off-the-shelf health apps are built for the average of everyone, which means they fit no one in particular. This one fit my meals, my habits, my kitchen.

But the story I want to tell you is about a feature I killed.

My clever original idea was to photograph every meal and have one of OpenAI’s models do live nutritional analysis inside the app. Very futuristic. Very much the kind of thing that demos well. In practice it added cost (paying for API calls on top of a subscription), added friction, and added surprisingly little. So I dropped it. Instead I use a ChatGPT project on the side to lock down the nutrition of a meal once, from a photo, and then just log it as a favourite in the app. Over a few weeks you build up a bank of your common meals, and the daily logging becomes a two-tap affair instead of a chore.

The lesson is one I keep relearning in every corner of my life with these tools: the cleverest system is rarely the one you’ll actually keep using. The best health tracker is the one that survives a travel day, a bad night’s sleep, and zero motivation. Elegant loses to sustainable, every time.

Same logic shaped the whole build. I let ordinary code handle the things that must be boring and correct: dates, totals, whether a checkbox saved. I let the AI handle the things where judgement and synthesis add value: spotting patterns, asking questions, doing a weekly review. You do not want a probabilistic language model deciding whether your Tuesday entry got saved, and you don’t need expensive inference for every tap. Ambiguity is where AI belongs. Consistency is where plain software belongs. Knowing the difference is most of the skill.

And yes, it broke in instructive ways. Imports stalled. A dashboard once cheerfully declared everything complete while quietly missing half the records. A green “import successful” message, I can now report, is not the same as your data actually being correct. AI can generate software at a speed that still startles me, but data plumbing and validation remain stubbornly, boringly real.

The worrier’s trap

Now the part I have to be careful about, because it’s the part that can quietly do harm.

Give a capable AI a stack of medical reports and it will read every single line with the intensity of a detective who suspects everyone. That is genuinely wonderful when you’ve missed something. In one of my reports, a nurse had recorded the wrong gender and age for a test. I’d skimmed straight past it. The AI caught it instantly. In another, it flagged a discrepancy I’d never have spotted.

But the same trait has a shadow. If you are, like me, a bit of a worrier, an AI that flags every mildly-off value and then helpfully speculates about what might be causing it is not always your friend. It can turn a routine result into a late-night investigation of seven possible conditions, most of which you do not have.

This is the Dr House problem. In the rare case where everyone’s genuinely stumped, that relentless “but what if it’s this” is exactly what you want. In the other ninety-something percent of cases, the boring answer is the right one. Good doctors are brilliant at this. Liver enzyme slightly up? Yes, it could theoretically be a dozen things, but looking at everything about you, the most likely explanation is the obvious one, and here’s why you can relax. Occam’s razor, delivered with a reassuring bedside manner. AI can do the razor. It’s still learning the reassurance.

So I learned to steer it. Ask for the most likely explanation first, not the scariest. Ask what would actually change the decision. Ask for the strongest case that this is nothing.

The other guardrail is verification, and I learned it the embarrassing way. On one scanned questionnaire, the AI read the crosses in the checkboxes as problem flags, when they meant the opposite: all clear. A small misread with a large potential to mislead. Handwriting, ticks, faded scans, a page rotated the wrong way: multimodal AI is impressive, and it is also perfectly capable of confidently getting a smudge wrong. On anything that matters, I now check the critical fields against the original page myself. Fluency is not accuracy. A beautifully formatted summary can be beautifully wrong.

What I’d actually tell a friend

If you strip all of this down to what’s worth stealing, it’s a short list. None of it is medical advice, just how to make the tool pull its weight without letting it run away with you.

1. Keep one running thread for an ongoing issue, not a graveyard of disconnected chats. Continuity is the whole point.
2. Feed it primary sources. The actual report, the label, the photo. Not your half-remembered paraphrase.
3. Ask for a doctor-safe brief, never a diagnosis. One page. Reason, timeline, meds, top three questions, uncertainties. Let the human make the call that matters.
4. Ask what would change the decision. What needs action, what can wait, what would make this urgent, when to reassess. It stops research from becoming its own anxious hobby.
5. Demand ranges, not fake precision. “Probably 15 to 25 grams of protein” is useful. “This meal contains 21.7 grams” is theatre.
6. Ask for the strongest alternative explanation, and the strongest case for doing nothing yet. AI anchors hard on the first plausible story. Make it argue against itself.
7. Make logging easier than skipping it. If it doesn’t survive a bad day, it won’t survive at all.
8. Verify anything that matters against the source. Especially anything scanned, handwritten, or too neat to question.
9. Bring your doctor three good questions, not your entire research dump. Nobody’s judgement improves under a pile of twenty technically-valid queries.

Which tool? Ask me again next week

Let me be specific, because vague “I used AI” claims help nobody.

For this kind of work I mostly landed on ChatGPT in Work mode (5.6 Sol Max, at the time of writing), inside a custom project that held all my context, reports, and running chats in one place. That combination stayed balanced and grounded in my actual situation, which is exactly what you want when the stakes are personal.

I ran the same inputs through Claude and Gemini too. Claude (I used Fable 5, and cheerfully burned through a pile of credits chasing maximum intelligence) is genuinely brilliant at some things. Ask it to code up a working app or a quick game on the fly and it’s close to magic. For this messy, context-heavy, high-stakes synthesis, I found it could reach for the elaborate answer when a simple one would do. Different tools, different sweet spots.

One practical wrinkle if you go deep: ChatGPT projects cap out around 25 files on a Plus plan (more on the higher tiers), so you have to be deliberate about what context you feed it. Claude’s approach of working with files straight off your computer sidesteps that limit, though it still sends them to the cloud to do the work, so don’t mistake the convenience for a privacy upgrade.

Now the caveat, and please read this bit twice: everything above is a snapshot, and a blurry one. These tools change on a weekly basis. Whatever I tell you is best at what today will very likely be wrong by the time you read this. The only real way to know is to test them yourself, on your own actual problems, with the same inputs, and build up a feel for it. Which tool for which job is a skill you earn by doing, not a spec sheet you can look up.

The bit that isn’t really about health

Here’s the wink, for those of you who, like me, spend your working life thinking about how organisations should use this stuff.

Everything above is the same lesson we keep learning at work, just wearing a hospital gown. The value never came from the model on its own. It came from the data I fed it, the context I maintained, the workflow I built around it, the rules that kept the boring things boring, the feedback I looped back in, and, crucially, the human judgement I kept firmly in the loop. The model was maybe twenty percent of it. The system around the model was the rest.

The genuinely new capability isn’t intelligence on tap. We’ve had clever answers for a while. It’s orchestration at the level of one person: having a record-keeper, a translator, a researcher, a prep assistant, and a slightly overzealous analyst all available at once, holding the same context, at a cost that would have been unthinkable to assemble a few years ago. That’s the shift. And it’s exactly the shift our industry is still fumbling toward, one pilot at a time.

Where this leaves me

The doctors still do the doctoring. The dentist found the crack. The physician read the heart. None of that changed, and I wouldn’t want it to.

What changed is the space in between. I stopped walking into appointments empty-handed and stopped walking out of them alone. The two weeks of pain still hurt. The heart note still means I need to move more and eat better, screen time and all. But I spent that stretch as a participant in my own health instead of a slightly panicked spectator.

That’s the part worth keeping. Not “AI will see you now.” More like: the waiting room in your pocket, the one that remembers everything, argues with itself when you ask it to, and hands you a clean page to give the person who actually knows what they’re doing.

Use it for the connective tissue. Leave the diagnosis to the humans. And for the love of all that is holy, when it hands you a seven-part differential diagnosis for a simple headache, ask for the boring explanation first.