Expert Medical Care for Everyone

On its benchmarks, medical AI can look close to solved. Models answer difficult medical questions, draft radiology reports, interpret images, and pass examinations. At the bedside, the same models can come apart.

They miss the finding they just described. They produce convincing language for the wrong reason. They succeed on normal cases and fail on abnormal ones. They perform well at one hospital and lose accuracy at another. They agree when they should disagree. They do not know when important information is missing.

That gap is the research problem.

I want everyone to have access to expert medical care. I am working to build AI doctors that can make this possible.

I use AI doctor to name the destination, not the technology we have today.

An AI doctor should be able to gather a patient’s context, interpret images and language, reason over time, explain its uncertainty, and know when to ask a human clinician for help. It should make a useful second opinion available where the relevant expert is not. It should take on more responsibility only when the evidence supports it.

Today’s systems can do pieces of this. The autonomy is not there yet.

A medical answer is not medical intelligence

Medicine does not arrive as one question with one correct answer. It arrives as a sequence. A symptom changes. A scan adds evidence. A treatment fails. A clinician notices that the history does not fit the image. The next decision depends on what happened before it.

A useful system therefore has to do more than produce an answer. It has to understand the whole medical picture, decide what information is missing, update its view, and connect what it sees to what should happen next. It has to distinguish confidence from fluency. It has to recognize when the safest action is to stop.

This is why the current work spans multimodal models, clinical agents, dynamic evaluation, clinician and AI collaboration, and embodied systems. These are not five disconnected trends. They are parts of the same attempt to build medical intelligence in the form that care actually takes.

The hard capabilities have become clearer. Description is one. Grounding, measurement, reasoning in three dimensions, and calibrated doubt are others. Fluent reports are getting easier. Reliable ones remain the research problem.

The first decade changed the questions

Ten years ago, I was a first-year PhD student working on machine reading comprehension. The question was simple: could a model read a passage and answer a question about it?

SQuAD made one bounded version of that question measurable. The dataset contained more than 100,000 questions grounded in Wikipedia passages. A shared test gave the field a target. Progress accelerated.

That experience taught me what measurement can do. It also taught me what a benchmark cannot do. Once a number matters, a field becomes very good at improving it. The number can focus effort, but it can also become a substitute for the capability it was meant to represent. Passing a reading comprehension test is not the same as understanding language.

The same distinction became impossible to ignore in medicine.

Work on electrocardiograms and chest radiographs tested whether deep learning could recognize clinically important patterns in medical data. Projects including CheXNet, CheXpert, and arrhythmia classification helped show what was possible on defined retrospective tasks.

But care is not a clean test set. Patients differ from the data used to build a model. Hospitals differ from one another. A visible finding may not be the finding that matters. The right next step may depend on information outside the image.

The early systems were not the endpoint. They exposed the next layer of the problem.

The failures tell us what to build

A generated medical report can sound right while hiding a serious error. In one recent evaluation, more than 90 percent of reports for normal studies were error-free, but fewer than 45 percent of reports for abnormal studies were. The difficult cases were exactly where reliability mattered most. That difference is the result, not an inconvenient detail around it.

The same pattern appears as systems take on longer tasks. In a benchmark that asks agents to solve medical imaging machine-learning problems, the agents performed around the twelfth percentile of human experts. They could make progress, but they could not yet work autonomously. The benchmark makes the limit visible and gives us something concrete to improve.

These results point to a research strategy. Build tests that are honest enough to show where systems fail. Study the mechanism behind the failure. Add the grounding, tools, structure, or interaction that the task requires. Test again under harder conditions.

Scale matters. It is not the whole answer. Progress depends on grounding, honest measurement, and structure around the model, more than on scale alone.

A clinician and an AI form a new system

It is tempting to assume that a capable model will automatically make a clinician more capable. The evidence is more complicated.

People can rely on a model when it is wrong and discount it when it is right. A confident explanation can change a decision even when the underlying answer is weak. The same recommendation can help one clinician and mislead another. A person and a model can perform worse together than either would alone.

Clinician and AI collaboration is therefore not a user-interface detail added after the model is built. It is a scientific problem. The unit of evaluation has to be the person, the model, and the workflow together.

This also changes what responsibility means. A system should not receive credit for being useful because it is accurate in isolation. We should ask what changed in the decision, who benefited, who was harmed, and whether the result holds outside the study.

Four questions organize the work now

Can one system understand the whole medical picture?

A clinician combines images, language, video, laboratory results, physiological signals, and the patient’s history. We are building systems that reason across these sources without losing the context that connects them. The goal is not more modalities. The goal is a better account of the patient.

Can clinical AI reason over time?

Care is a sequence of decisions, not a final answer. We build agents that gather information, revise their view, and work through clinical tasks. We also build environments that test the sequence dynamically, because a static score cannot show whether the reasoning held together.

When does AI actually help a clinician?

We study how people use AI, when the combination improves a decision, and how it fails. The question is not whether a model can be placed in a workflow. It is whether the workflow becomes better.

Can perception lead safely to action?

Interpreting an image is only one part of care. Procedures require systems to understand space, sequence, skill, and consequence. We are beginning to connect medical perception to action in imaging, procedural video, and robotic settings. The evaluation has to become more demanding as the distance between prediction and action shrinks.

The ambition should be large. The claims should be exact.

Several principles follow from the work so far.

Earn autonomy one capability at a time

A system should not receive responsibility because it is generally impressive. It should earn responsibility for a defined task, setting, and population. Greater consequences require stronger evidence, monitoring, and accountability.

Treat failures as evidence

A failure is part of the result. Distribution shift, missing information, confident mistakes, sycophancy, automation bias, and poor clinician and AI combinations show what the system is. They should change the next experiment.

Evaluate the system in the form it will be used

Static benchmarks remain useful. They are not enough for an interactive system. Evaluation should include changing information, sequences of decisions, uncertainty, escalation, the people using the model, and the workflow around it. Claims about clinical benefit should eventually depend on prospective evidence and outcomes.

Keep responsibility visible

AI does not remove accountability. Patients and clinicians should be able to understand what role a system played, what information it used, and where responsibility lies when something goes wrong. More capable systems make this question more important.

Make access part of the engineering

Access is shaped by cost, language, connectivity, privacy, local practice, hardware, and the availability of follow-up care. A strong model behind an expensive interface does not make expert care available to everyone. If access is the aim, these constraints belong in the design from the beginning.

Credit the people who make the work possible

Medical AI is built by teams: students, engineers, clinicians, patients, annotators, researchers, and institutions. The work on this site exists because many people contributed to it. Any honest account should make that visible.

What success would look like

The point is not to hand every patient interaction to a machine. It is to make expert judgment more available, more consistent, and better supported.

For one person, that may mean a second opinion where no specialist is available. For another, it may mean a clinician who can synthesize a long record without missing the change that matters. It may mean earlier recognition of an urgent condition, better preparation for a procedure, or a clearer explanation of the choices ahead.

The form will differ across settings. The standard is the same: the system should improve the care a person can actually receive, not merely produce an impressive answer.

There is no guaranteed path from today’s models to dependable AI doctors. Important capabilities are missing. The evidence is incomplete. Some of our assumptions will be wrong. That uncertainty is a reason to make the tests harder and the claims more precise.

The first decade showed me how quickly a field can move when a difficult capability becomes measurable. The next decade is about making the measurements look more like the world.

I want to help build systems that understand more of the patient, reason through more of the clinical process, collaborate better with people, and know their limits. Then I want the benefits to reach far beyond the places where expertise is concentrated today.

That is what I mean by AI doctors.