Getting to substance, the paper is foundationally weak, even as an opinion piece. There are two denominators for one study in one sentence, the six vignettes that became a load-bearing wall are listed as if they are large in number, sixteen patients per arm, concordance dressed as accuracy: others have done a much more complete analysis of the garbage-masquerading-as-science in here. https://www.fixhealth.ai/p/ai-will-replace-doctors-part-5-of
Look at the observations. On Hager’s 2,400 real patients, model performance collapsed the moment the models had to gather the information themselves instead of receiving a pre-packaged case. The Perspective’s own citation 23, Bean and colleagues in Nature Medicine, found that transmission of information between the model and the user is the particular point of failure. AMIE went from beating doctors in a text simulator to fifty-six percent on a hundred real patients with a physician supervising every interaction. Topol’s (like a broken clock, correct at least twice a day and he is correct here) reaction was that none of the studies were in real-world medicine.
Every study that shows a model outperforming physicians shares a hidden feature: the case arrived complete, consistent, and true, because a human expert assembled it, checked it, and edited it before the model saw it. The clinicopathological conferences, the case-report collections, the standardized patients, the vignettes in the Stanford trial. That assembly is invisible in the results, and it is the whole game. The “autonomous AI” in the evidence base is a model plus an unseen person who manufactured true premises for it, at a cost the study never counts. Take that person away, which is what real healthcare does, and you get Hager. In a separate 2025 experiment in Communications Medicine, when a single false detail was planted in the clinical material, frontier models repeated or elaborated it in fifty to eighty-two percent of outputs. In the real record, nobody has to plant such details. (https://www.nature.com/articles/s43856-025-01021-3) A diagnosis pasted a decade ago repeats with perfect fluency in every note that quotes it. I have personally seen a live commercial record that showed a patient simultaneously pregnant and due for prostate-cancer screening. The software was fine. The premises were not.
This is also why the doctors-make-it-worse finding, which the authors treat as decisive may be pointing somewhere else entirely. Why did fifty physicians add nothing to a model that had the case in front of it? Because in a curated study the case is already true, so there is nothing for the human to correct, and the physician’s only reference is the record. Move the same trial into a real clinic and the physician becomes the only agent in the loop even attempting the job the study quietly outsourced: establishing what is so. Not because human reasoning is superior. Because nobody else in the room is trying.
There is a foundational issue here, generally ignored by everyone because it has proven very difficult to solve. Generative AI succeeds wherever the world supplies a verifier: code compiles or it does not; a cited passage can be opened; a chess move is legal or it is not. Generative AI's failures concentrate wherever no external check exists, wherever premises cannot be verified by anything cheaper than an expert reading everything. Medicine is the largest domain on earth in that second category, and no amount of scale changes it, because the missing thing is not reasoning. It is a complete, reconciled, verifiable account of the one individual in front of the system, formed before the reasoning begins. There is not, and cannot be, a training set for that person. That account is not a generative act, and no model produces it for itself. It has to be manufactured deterministically, with every value traceable to its sources, so that when the reasoning is wrong the error has an address. Build that layer, and a great deal of medicine can safely become far more autonomous than the AMA would like. Skip it, and “AI alone” means a very fluent system reasoning beautifully about a patient who does not exist.
Here is a suggestion that would help everyone interpret this nonsense better. Every study should state where the case came from, who assembled it, and whether the model or a human established the premises. Then run the fair test on records as they exist, assembled by the AI from the actual chart rather than from a vignette, and measure fidelity to the premises alongside fidelity to the diagnosis. Until that column exists, every benchmark in this debate, including the ones that favor physicians, is measuring reasoning over a truth someone else supplied.
Validity requires a foundation of facts, to quote from the article cited above. And the foundation is the part nobody in the paper, and almost nobody in the field, is building. The authors may be right about 2030. But the milestone worth watching is not the day a model outperforms an internist on a case someone else assembled. It is the day the case assembles itself, truthfully, from the wreckage of the real record. That is a harder problem, it is not a language problem, and it is the one that decides everything else.
I can't really say how many individuals there are whose very name in the list of authors would alert me to not waste my time reading the article. But Dr. Emanuel would be very high on the list.
Dr. Emanuel has spent his entire career in academia and public health---much of the latter as a member of the public health bureaucracy. His articles and opinions are based on a collectivistic mindset with which I profoundly disagree. He is an advocate for medical care policies that, in my opinion, would further degrade an already degraded system. He may have taken some positions with which I might agree, but I can't remember any offhand. This is my personal opinion. You will have to read some of his articles and decide for yourself.
One thing which has to change is medical training. Not sure how but many years in youth are wasted upon knowledge that is not directed to how one should think as a doctor - and you see AI doing it better.
I feel much more comfortable discussing mathematics and AI than medicine and AI, but perhaps similar points hold. AI has proven some very important mathematical results, but usually there is a fair amount of human 'hand holding'- for example pointing out incorrect arguments. It is not a deus et machina. More importantly, AI does not seem to have any ability to ask interesting questions or come up with frameworks for a theory (the Langlands program would be a great example of the kind of thing that AI seems to have no ability to do). Perhaps the same sort of issues will be true in medicine. That is AI will be very good at certain kind of diagnostics, but not good at more subtle issues and unable to understand when it has made an egregious mistake. Perhaps the future is a partnership between human skills and AI skills.
Taking the human element out of medicine will end up a tragedy. The human connection between my team and I is - imo - WHY I have thrived despite myelomeningocele and the multiple issues I’ve worked through despite it.
I would point out that the human element was lacking during the COVID fiasco. People, including my friend, were forced to die alone. My father was denied an advocate during medical care. The human element was lacking in the medical industry as a whole during COVID.
This one ranks up there with your turmeric benefits video (better for curries than as a nutrient). More seriously, this little perspective was so full of holes that it did JAMA, Zeke, and the Khoslas no good. The case for autonomous AI-based care could have been made so much more strongly; this was a step backward.
Getting to substance, the paper is foundationally weak, even as an opinion piece. There are two denominators for one study in one sentence, the six vignettes that became a load-bearing wall are listed as if they are large in number, sixteen patients per arm, concordance dressed as accuracy: others have done a much more complete analysis of the garbage-masquerading-as-science in here. https://www.fixhealth.ai/p/ai-will-replace-doctors-part-5-of
Look at the observations. On Hager’s 2,400 real patients, model performance collapsed the moment the models had to gather the information themselves instead of receiving a pre-packaged case. The Perspective’s own citation 23, Bean and colleagues in Nature Medicine, found that transmission of information between the model and the user is the particular point of failure. AMIE went from beating doctors in a text simulator to fifty-six percent on a hundred real patients with a physician supervising every interaction. Topol’s (like a broken clock, correct at least twice a day and he is correct here) reaction was that none of the studies were in real-world medicine.
Every study that shows a model outperforming physicians shares a hidden feature: the case arrived complete, consistent, and true, because a human expert assembled it, checked it, and edited it before the model saw it. The clinicopathological conferences, the case-report collections, the standardized patients, the vignettes in the Stanford trial. That assembly is invisible in the results, and it is the whole game. The “autonomous AI” in the evidence base is a model plus an unseen person who manufactured true premises for it, at a cost the study never counts. Take that person away, which is what real healthcare does, and you get Hager. In a separate 2025 experiment in Communications Medicine, when a single false detail was planted in the clinical material, frontier models repeated or elaborated it in fifty to eighty-two percent of outputs. In the real record, nobody has to plant such details. (https://www.nature.com/articles/s43856-025-01021-3) A diagnosis pasted a decade ago repeats with perfect fluency in every note that quotes it. I have personally seen a live commercial record that showed a patient simultaneously pregnant and due for prostate-cancer screening. The software was fine. The premises were not.
This is also why the doctors-make-it-worse finding, which the authors treat as decisive may be pointing somewhere else entirely. Why did fifty physicians add nothing to a model that had the case in front of it? Because in a curated study the case is already true, so there is nothing for the human to correct, and the physician’s only reference is the record. Move the same trial into a real clinic and the physician becomes the only agent in the loop even attempting the job the study quietly outsourced: establishing what is so. Not because human reasoning is superior. Because nobody else in the room is trying.
There is a foundational issue here, generally ignored by everyone because it has proven very difficult to solve. Generative AI succeeds wherever the world supplies a verifier: code compiles or it does not; a cited passage can be opened; a chess move is legal or it is not. Generative AI's failures concentrate wherever no external check exists, wherever premises cannot be verified by anything cheaper than an expert reading everything. Medicine is the largest domain on earth in that second category, and no amount of scale changes it, because the missing thing is not reasoning. It is a complete, reconciled, verifiable account of the one individual in front of the system, formed before the reasoning begins. There is not, and cannot be, a training set for that person. That account is not a generative act, and no model produces it for itself. It has to be manufactured deterministically, with every value traceable to its sources, so that when the reasoning is wrong the error has an address. Build that layer, and a great deal of medicine can safely become far more autonomous than the AMA would like. Skip it, and “AI alone” means a very fluent system reasoning beautifully about a patient who does not exist.
Here is a suggestion that would help everyone interpret this nonsense better. Every study should state where the case came from, who assembled it, and whether the model or a human established the premises. Then run the fair test on records as they exist, assembled by the AI from the actual chart rather than from a vignette, and measure fidelity to the premises alongside fidelity to the diagnosis. Until that column exists, every benchmark in this debate, including the ones that favor physicians, is measuring reasoning over a truth someone else supplied.
Validity requires a foundation of facts, to quote from the article cited above. And the foundation is the part nobody in the paper, and almost nobody in the field, is building. The authors may be right about 2030. But the milestone worth watching is not the day a model outperforms an internist on a case someone else assembled. It is the day the case assembles itself, truthfully, from the wreckage of the real record. That is a harder problem, it is not a language problem, and it is the one that decides everything else.
I can't really say how many individuals there are whose very name in the list of authors would alert me to not waste my time reading the article. But Dr. Emanuel would be very high on the list.
Thank you.
Could you elaborate on why Dr Emanuel is not trustworthy? I never heard of him or else forgot his name and controversy
Dr. Emanuel has spent his entire career in academia and public health---much of the latter as a member of the public health bureaucracy. His articles and opinions are based on a collectivistic mindset with which I profoundly disagree. He is an advocate for medical care policies that, in my opinion, would further degrade an already degraded system. He may have taken some positions with which I might agree, but I can't remember any offhand. This is my personal opinion. You will have to read some of his articles and decide for yourself.
"The art of medicine consists of amusing the patient while nature cues the disease"...Moliere
One thing which has to change is medical training. Not sure how but many years in youth are wasted upon knowledge that is not directed to how one should think as a doctor - and you see AI doing it better.
I feel much more comfortable discussing mathematics and AI than medicine and AI, but perhaps similar points hold. AI has proven some very important mathematical results, but usually there is a fair amount of human 'hand holding'- for example pointing out incorrect arguments. It is not a deus et machina. More importantly, AI does not seem to have any ability to ask interesting questions or come up with frameworks for a theory (the Langlands program would be a great example of the kind of thing that AI seems to have no ability to do). Perhaps the same sort of issues will be true in medicine. That is AI will be very good at certain kind of diagnostics, but not good at more subtle issues and unable to understand when it has made an egregious mistake. Perhaps the future is a partnership between human skills and AI skills.
Taking the human element out of medicine will end up a tragedy. The human connection between my team and I is - imo - WHY I have thrived despite myelomeningocele and the multiple issues I’ve worked through despite it.
I would point out that the human element was lacking during the COVID fiasco. People, including my friend, were forced to die alone. My father was denied an advocate during medical care. The human element was lacking in the medical industry as a whole during COVID.
This one ranks up there with your turmeric benefits video (better for curries than as a nutrient). More seriously, this little perspective was so full of holes that it did JAMA, Zeke, and the Khoslas no good. The case for autonomous AI-based care could have been made so much more strongly; this was a step backward.