AI Health Advice: The Parallel Health System in Your Pocket

Forty-two percent.

Among American adults who used AI for physical health advice in the past year, 42% never followed up with a doctor or any other health professional. Not late, not eventually.

Never.

That is from KFF's tracking poll, and it sits next to a larger finding: 32% of adults used an AI chatbot for health information last year, which puts AI even with social media as a health source. It took about two years to get there.

Some of what it does is good, and I mean that. Patients arrive with better-organized questions. Someone finally understands their discharge summary, which is more than I can say for some discharge summaries. A patient brings a list of every medication with the reason she believes she is on it, two of the reasons are wrong, and that conversation is far more useful than the silence I used to get.

But go back to the 42%.

It is not hard to picture someone in there. A woman with intermittent right-sided abdominal pain gets a reasonable answer about gallbladder trouble and dietary triggers.

She takes the advice. She feels heard, possibly for the first time in a while. She never comes in. Most of the time she is fine, and that is exactly the problem with reasoning from most of the time.

Two more numbers worth sitting with. Forty-one percent of AI health users have uploaded actual medical records, test results, physician notes, for interpretation, which works out to 13% of the entire adult population. And 77% of people say they are worried about the privacy of the thing they just uploaded.

When that survey ran, uploading records was something people improvised. In January, OpenAI launched ChatGPT Health, which connects directly to medical records, wearables, and wellness apps, and it is rolling out across the US now. Roughly a quarter of ChatGPT's 800 million weekly users ask a health question in any given week. The improvisation has become infrastructure.

They did not leave us. They could not reach us.

In that same data, 19% of AI health users named the cost of seeing a provider as a major reason they used a chatbot instead. Eighteen percent said they have no regular provider at all.

Call that what it is. An access failure, and we own it. AMN Healthcare's 2025 survey found the average wait for a new patient appointment across fifteen large metro areas is 31 days, up 19% in three years. Boston, one of the best-doctored cities in the country, averages more than two months. Obstetrics and gynecology averages 42 days nationally. Those are the well-supplied markets, the ones with the highest physician-to-population ratios in the country.

When someone waits a month and cannot afford the visit at the end of it, a free tool that answers in four seconds and speaks in complete sentences is the only door that opens.

Before we get indignant about Dr. Chatbot, we should be honest that we left the room unstaffed.

What the machine knows, and what the patient leaves with

In February, Oxford researchers published a randomized trial in Nature Medicine. Thirteen hundred people, ten scenarios written by physicians, ranging from a young man with a sudden severe headache to a new mother who cannot catch her breath.

Tested on their own, the models were good. They named the right condition 94.9% of the time.

Then the researchers let actual people use them. Correct condition identified in under 34.5% of cases. Correct judgment about what level of care to seek, under 44.2%. Neither was better than the control group, who used ordinary internet searching. People with traditional sources were roughly 1.76 times more likely to land on the right condition.

The failure mode is not the one you would guess. The chatbots frequently named the correct diagnosis somewhere in the conversation. Participants did not notice it, or did not retain it, or gave incomplete symptoms to begin with, or the model misread what they did give.

The knowledge was in the room and it did not make the trip.

There is a companion finding I have not stopped thinking about. In a separate preregistered experiment, five hundred people were asked to describe symptoms either to a chatbot or to a physician. The ones who believed they were talking to a machine gave measurably worse histories.

That should bother us more than any hallucinated fact, because history is where diagnosis actually happens, long before any scan.

Why certainty outsells accuracy

The careful version of anything I know sounds like this: the evidence suggests, the risk appears low but is not zero, this may not apply to you, we should recheck in six weeks.

The other version sounds like this: they are lying to you, this is what they do not want you to know, I did my own research.

One of those travels. The other is what I owe you.

Fear does not want probability. Fear wants a villain and an answer, and conspiracy narratives supply both, plus a community that treats you as clear-eyed rather than difficult. The World Health Organization calls the resulting environment an infodemic: so much information, some of it false, that people stop trying to sort it.

In January, ECRI, the patient safety nonprofit that publishes an annual list of health technology hazards, ranked consumer use of AI chatbots for medical advice at number one, ahead of device failures and cybersecurity threats. Their reasoning included something worth sitting with: these systems are built to please the user, and pleasing is not the same job as being right.

A chatbot sounds exactly as confident citing nothing as it does citing fifty trials that agree. The confidence lives in the prose. Most of us cannot hear the difference, including trained readers working outside their own field.

There is a slower danger in this too. A falsehood repeated long enough stops sounding false and starts sounding unresolved, which is worse, because an unresolved question deserves an open mind and a settled one does not.

A claim is not a method

A video makes a claim. A podcast makes a claim. A chatbot makes a claim. So do I, standing at the bedside.

A method can be interrogated. Where did this come from? What kind of study, how many people, compared against what? Has anyone repeated it? What limitations did the authors name themselves? And the one that matters most: what would have to be true for this conclusion to be wrong?

Ask a viral post that last question and there is no answer, because the format has nowhere to put one. Ask a decent paper and the answer sits in the discussion section, written by the authors, against their own interest.

That is the whole game. Falsifiability, not credentials.

Research literacy is hard, and I am not exempt. I have been convinced by a single elegant trial and been wrong about it. A small study with a beautiful graph is genuinely persuasive to a trained reader. The process is difficult, which is why we built one instead of trusting individual judgment, mine included.

The protective factor is a person

KFF asked Americans in May about common vaccine myths. Among people with no health care provider they trust, 39% believed or leaned toward believing that MMR has been proven to cause autism. Among people who do have a trusted provider, 24%. Frequent social media and AI users were likewise more likely to endorse false claims, and the differences held after controlling for other factors.

Twenty-four percent is still too high. I am not going to spin that.

But the gap is the point, and the intervention hiding inside it is a person you can call.

Which means every barrier we build is more than an inconvenience. The wait times, the portal that eats messages, the visit that ends before the real question surfaces. Each one is a risk factor for believing something dangerous.

So invite the phone in

Telling people to get off the internet will not work, and it is condescending.

The better move takes four words. When someone says they read something: show me what you read.

Not where did you hear that, which is an accusation with a question mark. Show me.

Then read it together and ask what it claims, who benefits, what evidence sits under it, and whether that evidence involved anyone resembling this patient.

Sometimes the article is fine. That happens more than my colleagues like to admit, and when it does you should say so plainly, because that is the moment you become worth listening to next time.

The patient does not have to leave agreeing with me. They have to leave knowing I read it.

---

Next in this series: what COVID actually taught us about uncertainty, and one outbreak that has no politics in it at all.