The Map.

01 · The blank map

Millions use AI for their health. No one has mapped it.

We know the territory exists, the way early geographers knew a continent was there. The interior is blank.

Who are these users? What do they actually do? In what circumstances, with what skill, to what effect? The honest answer, today, is that nobody knows, because nobody has properly looked.

That blank is not a curiosity. It is a scandal. We would never tolerate a disease this widespread with no epidemiology, no map of who it touches and how. Yet the largest shift in how ordinary people engage with their own health in a generation is happening almost entirely off the chart. Policy is already being written about it, on terrain no one has surveyed. The map is the thing we are building, and this page is what it will measure.

02 · Why there is no map yet

The existing research asks the wrong questions.

It is not for lack of papers. Most of the literature is normative — should patients use AI? — or narrowly experimental: can a person score correctly on a medical quiz with a chatbot in a lab? Neither is cartography. Neither tells you what is actually happening in the world, at scale, in the kitchens and waiting rooms and 2 a.m. bedrooms where the real use occurs.

And the little descriptive work that exists is too slow and too thin to count. By the time a survey of attitudes toward last year's model clears peer review, the tools have changed twice and the behavior with them. The terrain is unmapped not because it does not matter, but because the field's instruments were pointed somewhere else. We are pointing them at the ground.

03 · What the map measures

Seven axes.

A real topography has dimensions, contours, terrain. The map of patient AI use has seven axes, and a finding only means something when it is located on all of them.

Who. The user.

Patients themselves, but also the caregiver tracking a parent's symptoms, the parent decoding a child's results, the patient-builder writing software for a disease the market ignores. Sorted by condition — acute, chronic, rare, undiagnosed, mental health — and by situation: uninsured, remote, dismissed, excluded by language or cost.

What for. The use.

A whole taxonomy no one has drawn: comprehension (decoding a lab, translating a diagnosis), preparation (getting ready for a visit), navigation (appeals, second opinions, finding care), decision support, self-tracking and management, building, and the quiet, enormous category of company and reassurance when there is no one else to ask.

How. The practice.

Which tools, in which language, by voice or by text. Whether the user brings their own context or asks cold. Whether it is a single question or a sustained relationship across months of a chronic illness.

With what judgment. The quality of use.

Does the user interrogate the answer or accept it? Cross-check against a second source? Know the line between understanding something and acting on it? This is the axis the warnings assume is always at zero. It is not, and measuring where it actually sits is among the most important things the map can do.

To what immediate effect. The consequence in the moment.

An error caught, a decision changed, an appeal won, an anxiety eased — or a false reassurance, a missed signal, a harm. This is the effect you can observe close to the interaction, and the axis the field most wants to talk about yet least knows how to observe, because you cannot see it in a lab full of people pretending to be sick.

To what health outcome. The downstream effect on health itself.

Over months and years: diagnosed sooner or later, treated better or worse, healthier, sicker, harmed, or unchanged. This is the gold-standard question and the one that finally decides whether patient AI use is, on balance, good — and for whom, and under what conditions. It is the hardest thing the map must eventually measure, and the most important.

Against what barrier. The context of access.

Whether the use happens instead of care that cannot be reached, alongside care, or in the long gaps between visits. The map is drawn from the patient's position, and that position is usually defined by what they could not otherwise get.

04 · How we build it

From primary observation, in the open.

Not the way the failed literature builds it. The map is assembled from primary observation of real use, held to the REAL-PAIR standard: real patients, real questions, real stakes, not paid strangers role-playing symptoms from a script.

Its first instruments already run. The Radar is the observatory: a daily, forensic census of the software patients actually build, the most unambiguous record of real use there is, because code does not exaggerate. The Casebook is its counterpart for the literature: every paper we run through the instruments, assessed and on the record, so the claims made about patients can be weighed as carefully as what patients do. The Daily Scan is the longitudinal record. The frameworks are the measurement methodology, including the instruments that audit the existing literature so its errors do not contaminate the map. Everything is built in the open, dated, and falsifiable.

05 · The instruments we still need

Most of them don't exist yet. That is the work.

These are the early instruments, and they are honest about their reach. The Radar sees only patients who publish code. The frameworks mostly measure other people's research rather than use itself. Together they survey a few corners of a vast territory. To map the whole of it, most of the instruments still have to be built — and building them is not a detail of the work. It is the work.

The science needs instruments that reach the users who never publish anything, which is almost everyone: ways to study real conversations and real use directly, with consent, ethically, at scale. It needs instruments that measure the quality of use across a population. It needs instruments that capture immediate effect, and — hardest of all — that follow real patients over real time to observe health outcomes. It needs panels and cohorts held to the REAL-PAIR standard. It needs instruments that work in the languages and places the existing research ignores entirely, and that capture caregivers, harms, and false reassurance as carefully as benefit.

None of this exists yet, anywhere. That is precisely why it must be built, and built patient-directed: a scientific toolkit designed from the patient's position, for questions patients actually have, rather than retrofitted from instruments built to study something else.

06 · What is visible, and what is dark

A good map shows where it ends.

Some contours are emerging. We can already see that patients build, in large numbers, for conditions no company serves. We can see that experienced users cross-check and interrogate far more than the warnings assume. We can see patients winning insurance appeals with AI-drafted arguments.

Far more is dark. The users who never publish anything, and so never appear in the Radar. Everyone using AI in a language the research ignores. The actual distribution of judgment across the population. And above all, health outcomes — what difference any of this finally makes to whether people get better, over the months and years a snapshot cannot see. Naming the dark regions precisely is not an admission of failure. It is the research agenda.

07 · Whose map it is

Made with patients, not about them from a distance.

Whoever maps a territory defines how the world understands it. Right now, in the absence of a map, patient AI use is being defined by the people most anxious about it, on the basis of evidence that never observed a patient. The decisions that follow — what to permit, what to warn against, what to build — are being made on a blank chart.

The map is infrastructure. Patients need it to see they are not alone. Builders need it to know what to build. Clinicians and policymakers need it to make decisions grounded in reality instead of fear. And it must be drawn from the place patients actually stand, which is the meaning of the four words under our name: patient-directed, AI-assisted.

That is the science we are building. It does not yet exist. It should. We intend to make it.