What Foundation Models Mean for Medicine, Explained by a Physician-Scientist

A foundation model is a very large AI system trained on broad data so it can be adapted to many tasks, and in medicine that generality is both its promise and its problem. These models can draft notes, summarize a chart, answer a factual question, and sketch a differential, often impressively. What they do not do on their own is know your patient, understand cause and effect, or guarantee that a fluent answer is a correct one. The honest position is that a foundation model is a capable general-purpose assistant whose medical trustworthiness has to be earned task by task, not assumed from a benchmark score.

That gap between fluency and reliability is where the responsible-AI questions live. To use these systems well, a clinician needs to understand what they are, where their competence actually holds, and what to verify before one is allowed near a decision that affects a person.

What a foundation model actually is

The older generation of medical AI was narrow: a model trained to read one kind of scan or predict one outcome, tested against that one job. A foundation model works differently. It learns general patterns from enormous, mixed datasets of text, images, and sometimes code and structured records, then is adapted, or fine-tuned, to specific uses. Large language models are the most familiar example, though the same idea now spans imaging, pathology, and multi-modal systems that combine several data types.

The appeal is obvious. One base model can be pointed at many problems, which is why a single system can summarize a discharge note in the morning and answer a drug-interaction question in the afternoon. Generality, though, is not the same as competence in any particular case. A model can be broadly capable and still wrong in the specific, narrow situation that happens to be the one in front of you.

Where these models genuinely help

The strongest, least controversial uses reduce work without owning a clinical decision. Drafting documentation, summarizing long records, turning a conversation into a structured note, retrieving and phrasing reference information: in each, a human reads the output before anything happens, so an error is caught rather than acted upon. Used this way, a foundation model is closer to a fast, tireless clerk than to a diagnostician.

Breadth helps too. A generalist model has seen far more written medicine than any individual could, which makes it useful for surfacing a possibility a busy clinician had not considered, or for explaining a mechanism in plain language for a patient handout. The value there is prompting human thought, not replacing it.

Where fluency is not reliability

The central risk is that these systems are optimized to produce plausible language, and plausibility is not truth. A model can generate a confident, well-formatted answer that is subtly or completely wrong, complete with a citation to a study that does not exist. In consumer settings this is annoying. In medicine it is dangerous, because a fluent wrong answer is exactly the kind a hurried reader is most likely to accept.

Three failure modes deserve naming. The first is fabrication, often called hallucination, where the model states something false with full confidence. Another is silent drift from intended use, the same trap that breaks narrower tools, where a system built to condense notes starts being trusted to decide. A third is unreliable uncertainty: these models are frequently just as confident when wrong as when right, so their tone cannot serve as a safety signal. A tool that cannot reliably say it is unsure has to be double-checked every time.

Equity belongs on the list as well. A model reflects the data it learned from, and medical data underrepresents whole populations. My own research on ethnic differences in the relationship between insulin sensitivity and insulin response is a small reminder that a pattern true in one group does not automatically hold in another. A generalist model trained mostly on data from some populations can be quietly worse for others, and the failure will not announce itself.

The questions to ask before trusting one

The discipline that governs any clinical tool applies here, with a few additions specific to generality. Start with intended use: what exact task is this model evaluated and approved for, and is that the task you are about to hand it? A model validated to draft notes has not been validated to triage symptoms, even if it will happily attempt both.

Then ask about evidence in your setting. Benchmark scores on medical exam questions are the weakest proof; they show the model can pattern-match on test items, not that it improves care for real patients. Stronger is evaluation on data like yours, and stronger still is a prospective study measuring what changed when the tool entered real workflow. In the EASY-1 randomized controlled trial (NCT03258268) we evaluated a diabetes decision-support system I helped develop against standard of care, with prespecified endpoints and a real comparator. That is the evidence that lets you say a tool helped, rather than that it sounded helpful.

The human role matters just as much. A responsible deployment keeps a clinician as the decision-maker, with the model drafting, retrieving, or summarizing under review, never acting unsupervised on a consequential call. Confirm that the system can flag uncertainty and decline rather than guess. Check how it will be monitored after go-live, since these models drift as language, guidelines, and data change; a good answer names what is measured, how often, and what triggers a pause. Confirm, too, who is accountable when the output is wrong, and get that written down before deployment rather than after a case review.

Where the standards sit

None of this lowers the bar medicine already sets. Regulators are converging on treating clinical AI as software that needs a stated intended use, real clinical evidence, lifecycle monitoring, and a named responsible party, principles reflected in good machine-learning practice guidance and, for anything touching a trial, in the modernized Good Clinical Practice framework of ICH E6(R3), finalized in early 2025 and adopted across major regions through that year, with its emphasis on quality by design and risk-based oversight. How a model was built does not change what it must prove. A foundation model earns clinical trust the way anything else does, by showing, in the patients you actually treat, that it helps and does not harm.

This article is educational and reflects my own view as a researcher; it is not medical advice, and patients should always talk to their own clinician about their care.