Future medicine · AI · earned trust

Forty Years Before I Sign My Name

A seasoned physician's answer to whether AI is ready for medicine: not yet — and here is exactly what "not yet" has always meant.

I have spent four decades earning the right to sign my name to a clinical judgment — bedside cardiology, a pharmacy doctorate, a career inside managed care where I now direct physician education. I have watched residents cross that same threshold, one supervised decision at a time, until the day their attending stops looking over their shoulder. I know exactly what that transition looks like, because I have graded it in other people for most of my career. So when I'm asked whether AI is ready to practice medicine, I don't answer as a novice dazzled by a demo. I answer as someone who has spent forty years deciding, again and again, whether a given mind — human or otherwise — has earned the authority to be trusted unsupervised. My answer is: not yet.

I came to that answer twice in the same week, from two very different directions. The first was a Perspective in the New England Journal of Medicine titled "Can AI Say 'I Don't Know'?" The authors open with a resident asked to explain a patient's rising creatinine. She pauses. "I don't know," she says — and the team responds the way a well-trained team should: they check levels, they call nephrology, they narrow the uncertainty instead of pretending it isn't there. Then the authors hand the same case to a clinical AI system. It doesn't pause. Despite conflicting evidence, it answers with total confidence, and it never once flags what it doesn't know.

The second came from an unlikely source: my astronomy professor, Dr. Bos, grading a discussion assignment in my freshman course. She'd asked us to hand a scientific statement to an AI tool and see what came back. When she closed the thread, she wrote the class a note with three lessons. AI can be talked into things — one student kept prompting until the tool caved and partially agreed with a statement it should have rejected outright. Different tools give different answers — run the same question through several, and completeness and reasoning vary widely. And the one that stuck with me: don't overlook the textbook — the answer was sitting on pages 32 and 33 all along, plainly explained, with a diagram. A freshman astronomy class and the pages of the New England Journal of Medicine arrived at the identical verdict, from opposite ends of a forty-year career.

Here is what those forty years actually taught me about trust, and why I recognize its absence so quickly in a machine. Trust in medicine is never granted for a right answer. It is granted for a track record of the right response to being wrong — for pausing, for saying "I don't know," for reaching for the textbook instead of bluffing past it, for tolerating the discomfort of an unresolved question long enough to actually resolve it. I didn't earn my patients' trust, or my residents' respect, by having every answer. I earned it by being reliably honest about which answers I didn't have yet, and by knowing where to look next. That is the entire curriculum, disguised as a career.

Confabulation isn't a system telling you a story. It's a system that has never once had an attending stop it and ask, "How sure are you?"

The NEJM authors cite a study in which large language models were fed vignettes containing a single fabricated detail; the models accepted and amplified it 50 to 82 percent of the time. They cite another in which a Pokémon name, slipped into a medication list, got confidently dosed 90 percent of the time. I would never let a second-year resident sign off on either of those charts unsupervised — not because they lack intelligence, but because they haven't yet built the reflex that stops a hand before it moves. That reflex is not a feature you can bolt onto a model. It is grown, slowly, the way I grew mine: by being wrong in front of someone senior enough to catch it, often enough that caution eventually became instinct.

So this goes in the "Future" cell of the lounge, not the "Present," because I think the question worth asking isn't whether AI can pass a medical exam — several already have. It's whether it can be trusted the way I finally trust myself: not because I am infallible, but because I have a well-worn habit of saying "not yet" exactly when I mean it. Until a clinical AI system can do the same, reliably, under pressure, on a patient it has never seen before, it isn't ready to sign its own name. Forty years taught me that earning that signature takes longer than anyone in a hurry wants to hear. I see no reason a machine gets to skip the line.

— V.W.P.