I'd be very very hesitant to trust studies like this. It's very easy to mess up these benchmarks.
See for example this recent paper where AI managed to beat radiologists on interpreting x-rays... when the AI didn't even have access to the x-rays: https://arxiv.org/pdf/2603.21687 (on a pre existing "large scale visual question answering benchmark for generalist chest x-ray understanding" that wasn't intentionally messed up).
And in interpreting x-ray's human radiologists actually do just look at the x-rays. In the context the article is discussing the human doctors don't just look at the notes to diagnose the ER patient. You're asking them to perform a task that isn't necessary, that they aren't experienced in, or trained in, and then saying "the AI outperforms them". Even if the notes aren't accidentally giving away the answer through some weird side channel, that's not that surprising.
Which isn't to say that I think the study is either definitely wrong, or intentionally deceptive. Just that I wouldn't draw strong conclusions from a single study here.
I agree with you on this specific study, however, I can't really wrap my head about the fact that doctors will be better than AI models on the long-run. After all, medicine is all about knowledge, experience and intelligence (maybe "pattern recognition"), all those, we must assume that the best AI models (especially ones focusing solely in the medical field) would largely beat large majority of humans (aka doctors), if we already have this assumption for software engineers, we should have it for this field as well, and let's be realistic, each time I've seen a doc the last few months (and ER twice), each time they were using ChatGPT btw (not kidding, it chocked me).
So I’m genuinely curious:
What is the specific capability (or combination of capabilities) that people believe will remain permanently (or at least for decades) where a top medical AI cannot match or exceed the performance of a good human doctor? Let's put liability and ethics aside, let's be purely objective about it.
Medicine is so much more than "knowledge, experience, and pattern matching", as any patient ever can attest to. Why is it so hard for some people to understand that humans need other humans and human problems can't be solved with technology?
How are you defining technology? How are you defining human problems? Inventions are created to solve human problems, not theoretical problems of fictional universe. Do X-rays, refrigerators, phones and even looms solve problems for nonhumans?
Claiming something that sounds deep doesn’t make it an axiom.
"Human problems can't be solved with technology" is just wrong, unless you have narrower definitions of a "human problem" or "technology".
For instance, transportation is a "human problem". It's being successfully solved with such technologies as cars, trains, planes, etc. Growing food at scale is a "human problem" that's being successfully solved by automation. Computing... stuff could be a "human problem" too. It's being successfully solved by computers. If "human problems" are more psychological, then again, you can use the Internet to keep in touch with people, so again technology trying to solve a human problem.
So much of what I know from women in my life is that the human element of medicine is almost a strict negative for them. As a guy it hasn't been much better, but at least doctors listen to me when I say something.
I would personally vastly, vastly to go to a robot doctor, who diagnoses, treats and nurses me. What exactly do I need from a human here? Except of course being the one making the system.
Emotional support. Some human doctors absolutely radiate confidence and a kind of "you're gonna be okay" attitude. For me, this helps a lot. I'm not sure a machine can do this.
> What is the specific capability (or combination of capabilities) that people believe will remain permanently (or at least for decades) where a top medical AI cannot match or exceed the performance of a good human doctor? Let's put liability and ethics aside, let's be purely objective about it.
Being a human when a patient is experiencing what is potentially one of the worst moments of their life. AI could be a tool doctors use, but let’s not dehumanize health care further, it is one of the most human professions that crosses about every division you can think of.
I would not want to receive a cancer diagnosis from a fucking AI doctor.
On the other hand, health care is not scaling to meet the growing demand of societies (look at the growing wait queues for access to basic medical attention in most Western nations). The cause of this is a separate topic and something that deserves more attention than it currently gets, but I digress. If AI can fill the gap by making 24/7/265 instant diagnosis and early intervention a reality, with it then bringing a human into the loop when actually necessary... I think that is something worth pursuing as a force multiplier.
We're clearly not there yet, but it is inevitible that these models will eventually exceed human capability in identifying what an issue is, understanding all of the health conditions the patient has, and recommending a treatment plan that results in the best outcome.
You may not want to receive a cancer diagnosis from an AI doctor... but if an AI doctor could automatically detect cancer (before you even displayed symptoms) and get you treated at a far earlier date than a human doctor, you would probably change your mind.
But liability and ethics cannot be put aside. If treatments were free of cost and perfectly address problems, then a correct diagnosis would always lead to the optimal patient outcome. In that scenario, AI diagnosis will be like code generation and go asymptotic to perfection as models improve.
But a doctor's job in the real world today is to navigate a total mess of uncertainty: about the expected outcome of treatments given a patient's age and other peoblems. About the psychological effect of knowing about a problem that they cannot effectively treat. Even about what the signals in the chart and x-ray mean with any certainty.
We are very far from having unit test suites for medical problems.
Sure, but my anecdotal experience is that doctors do this regularly in real life, especially when choosing to diagnose or ignore problems that are unlikely to kill an aging patient before some other larger issue does.
> I can't really wrap my head about the fact that doctors will be better than AI models on the long-run.
Nobody said that though?
If the current trajectory continues and if advancements are made regarding automated data collection about patients and if those advancements are adopted in the clinic then presumably specialized medical models will exceed human performance at the task of diagnosis at some point in the future. Clearly that hasn't happened yet.
Until medical models can contrive of unique diagnosis, this will not be true and cannot be true.
Medical models can absolutely get better at recognizing the patterns of diagnosis that doctors have already been diagnosing - which means they will also amplify misdiagnosis that aren't corrected for via cohort average. This is easy to see a large problem with: you end up with a pseudo-eugenics medical system that can't help people who aren't experiencing a "standard" problem.
The pitfall you describe is not inconsistent with exceeding human performance by most metrics.
I'd argue that the current system in the west already exhibits this problem to some extent. Fortunately it's a systemic issue as opposed to a technical one so there's no reason AI necessarily has to make it worse.
Medicine is about knowledge, but acquiring knowledge may in fact require "breaking out of the box" that AI is increasing behind to avoid touching "touchy subjects" or insulting anyone and so on.
This study is based almost entirely on pre-existing "vignettes." In other words, on tests that are already known and have existed for years, the model did well, which is precisely what you should expect.
It provides no information on real world outcomes or expectations of performance in such a setting. A simple question might be "how accurate are patient electronic health records typically?"
Finally, if the Internet somehow goes down at my hospital, the Doctor can still think, while LLM services cannot. If the power goes out at the hospital, the Doctor can still operate, while even local LLMs cannot.
You're going to need to improve the power efficiency of these models by at least two orders of magnitude before they're generally useful replacements of anything. As it is now they're a very expensive, inefficient and fragile toy.
but those kind of x-ray models are already activly used. They are not used though as a only and final diagnosis. Its more like peer review and priorization like check this image first because it seems most critical today.
I think AI can be useful in any kind of context interpretation, but not make a decision.
Could be running in the background on patient data and message the doctor "I see X in the diagnostic, have you ruled out Y, as it fits for reasons a, b, c?"
I like my coding agents the same way, inform me during review on things that I've missed. Instead of having me comb through what it generates on a first pass.
These type of experiments are bound to have biases depending on who is doing it and who is funding it. The experiment is being funded for a particular reason itself to move the narrative in a desired direction. This is probably a good reason to have government funded research in these type of sensitive areas.
I still don't quite understand, after skimming the paper. How does it achieve high scores without access to the images (beating even humans with access to the images)?
I believe in modern medicine but I lost some faith in the American institutions around it when I "diagnosed" my partner with the correct disease that the first rheumatologist dismissed and told them to just stretch. It was officially diagnosed years later, and we lost a lot of time because of it.
It's 50% of the time ER doctors working solely from notes, something they never do, in a situation they know is only for a study, will miss what you have.
I'm even more concerned that current models are not trained to say no, or to even recognize most failure modes.
"Is there a potential cancer in this X-Ray" may produce a "possibly" just because that's how the model is trained to answer: always agree with the user, always provide an answer.
Oh, and don't forget that "Is there a potential cancer in this X-Ray" and "Are there any potential problems in this X-Ray" are two completely different prompts that will lead to wildly different answers.
I'm surprised at both the article and the paper - both seem very hyperbolic. This is LLMs competing against doctors in a way that is heavily weighted in the LLMs favour, which does not represent clinical practice. These reasoning cases are not benchmarks for doctors, they are learning tools.
I think it's important to note that diagnosis also relies on accurate description of the patient in the first place, and the information you gather depends on the differential diagnosis. Part of the skill of being a doctor is gathering information from lots of different sources, and trying to filter out what is important. This may be from the patient, who may not be able to communicate clearly or may be non verbal, carers and next of kin. History-taking is a skill in itself, as well as examination. Here those data are given.
For pattern recognition from plain text, especially on questions that may be in the o1's training data, I'm not surprised at all that it would outperform doctors, but it doesn't seem to be a clinically useful comparison. Deciding which investigations to do, any imaging, and filtering out unnecessary information from the history is a skill in itself, and can't really be separated from forming the diagnosis.
Believable and not shocking. LLMs literally may have saved my sons and potentially her mother too by allowing us to fact check a lot of non sense data and scare tactics by a group of at least 5 different doctors ambushing us to make a life changing decision in minutes. The problem is doctors, at least in the US, prioritize liability exposure over patients long term outcomes. Let’s say you need an intervention where two options A and B are available to you. A carries 1% risk of complications but a great outcome. Option B has 0.1% risk of complications but once you are discharged the short term effects are challenging and long term effects not well understood. Well, 10/10 times doctors will suggest option B and will do anything they can to nudge you into making that choice, like not telling you the absolute numbers and constantly using the word “death”. They also lie about the outcomes, because again, once you accept the procedure, sign and are sent home, they have nothing to do with you.
LLMs can be a useful second opinion for a highly educated patient with good insight into their health and body, but this is not the average patient I see in an urban emergency department. Many patients can't give a cohesive history without a skilled clinician who can ask the right questions and read between the lines.
I am very skeptical of studies like this that don't adequately reflect real world conditions, but when I was a software engineer I probably wouldn't have understood what "real" medicine is like either.
> "An AI and a pair of human doctors were each given the same standard electronic health record to read"
This is handicapping the human doctors abilities. There is a lot more information a human doctor can gather even with a brief observation of the patient.
Agreed. I think the best use of this sort of tech is to use both to their strengths. Use AI to go over the record and suggest diagnoses which you have the doctor review after observing the patient.
The other thing is that common issues are common. I have to wonder how much that ultimately biases both the doctor and the LLM. If you diagnose someone that comes in with a runny nose and cough as having the flu you will likely be right most of the time.
I don't think AI is a good use case for such critical situations. Maybe in a decade we have AI help out doctors with doing a pre check. What if Ai finds nothing and the doctor does not bother to look into it further? It is this small question which breaks the technology from any angle later down the road from my POV. AI has to stay optional here.
Even if AI is used to sample or summarize a lot of data that a human couldn't do in time: What if it misses something that a human won't? What if a human inversely misses something that AI won't? Would you rather trust the machine or the human? (Especially if the human is held accountable.)
As a 37 year old male with 2 THRs I'm glad the AI was NOT used in my diagnosis. All the models that I used to look at my x-rays said nothing was wrong, even when adding symptoms. When adding age it said the patient was too young.
(I was ~3 months away from wheelchair bound in those x-rays).
The worst one was Gemini. Upload an x-ray of just the right hip, and it started to talk about how good the left hip looked like.
I think with AI taking over it's gonna be harder to get a solution when your problem isn't the run-of-the mill.
All versions and levels of Gemini have terrible spatial reasoning. I don't know why. That kind of task seems to be simply outside of the abilities of the model.
Besides for myself and wife, I've also used LLMs to diagnose my dogs. Convinced there's a huge opportunity for AI based veterinary, especially one which then performs bidding across the local veterinary clinics to perform the care/surgeries. I've noticed that local vets vary in price by more than an order of magnitude. My 80 year old mother and mother inlaw have been regularly scammed by over charging vets, and with their dogs being a major part of their lives, they extremely susceptible to pressure.
All the other points raised in this thread aside, it seems like an odd thing to benchmark because a significant proportion of ER practice is dealing with emergencies, often accidental injuries. There's not a whole of diagnosing going on if you show up to ER with a gash on your forehead or a missing finger.
radiology already had its "AI beats doctors" moment. radiologists are still here. what changed first was the workflow, not the specialty. er is probably next.
5. Private Equity uses this valuable data to stack rank doctors based on how correct / AI-aligned their diagnoses are over time
6. Rankings are used to periodically "trim the fact" thus delivering more optimized cash flows to clinics that have been saddled with toxic debt
7. Sensing an opportunity AI providers start selling a $200 / month Data Leakage as a Service subscription to overworked physicians so that they can avoid the PE guillotine
A more realistic step 7 is that physicians gradually align their diagnoses with the LLM as they sacrifice to Moloch in order to (temporarily) game the metric. Eventually the humans become little more than an imperfect proxy for the LLMs and are eliminated.
I agree with GP's solution but we'd need regulation to prohibit what you describe.
5. Doctors delegate everything to AI assistants because humans are lazy, especially if those AI assistants are correct some significant portion of the time
Then the claim may be that you don't need that many doctors anymore and that one doctor can do the job of X doctors in less time which has the economical effect that there is less demand for/supply of doctors, which then results in a home grown shortage of doctors, since less people are incentivized to become doctors...
It is easy to overinterpret this based on the headline, the doctors were actually at a slight disadvantage. This isn't how they normally work, this is a little more like a med school pop quiz:
An AI and a pair of human doctors were each given the same standard electronic health record to read – typically including vital sign data, demographic information and a few sentences from a nurse about why the patient was there. The AI identified the exact or very close diagnosis in 67% of cases, beating the human doctors, who were right only 50%-55% of the time.... The study only tested humans against AIs looking at patient data that can be communicated via text. The AI’s reading of signals, such as the patient’s level of distress and their visual appearance, were not tested. That means the AI was performing more like a clinician producing a second opinion based on paperwork.
"I don't know, let's run more tests" is also a very important ability of doctors that was apparently not tested here. In addition to all the normal methodological problems with overinterpreting results in AI/LLMs/ML/etc. Sadly I do think part of the problem here is cynical (even maniacal) careerist doctors who really shouldn't be working at hospitals. This means that even though I am generally quite anti-LLM, and really don't like the idea of patients interacting with them directly, I am a little optimistic about these being sanity/laziness checkers for health professionals.
I’ve had much better luck with diagnosis of my own family’s issues than with doctors. Usually now, I’m feeding them more information to begin with, so that their 30 minute office visits are not wasted, requiring another expensive follow up appointment.
While I’m sure there can be ways in which such studies are wrong, it’s very obvious that AI can accelerate work in many of these areas where we seek out professional help - doctors, lawyers, etc.
It can speed up some aspects of work, but please don't trust some llm with variable quality of output more than professional. If you don't like current doctor try another, most are in the business of helping other people.
If you have string of issues with 10 last doctors though, then issue is, most probably, you...
My wife is a GP, and easily 1/3 of her patients have also some minor-but-visible mental issue. 1-2 out of 10 scale. Makes them still functional in society but... often very hard to be around with.
That doesn't mean I don't trust your words, there are tons of people with either rare issues or even fairly common ones but manifesting in non-standard way (or mixed with some other issue). These folks suffer a lot to find a doctor who doesn't bunch them up in some general state with generic treatment. There are those, but not that often.
It helps both sides tremendously if patient is not above or arrogant know-it-all waving with chatgpt into doctor's face and basically just coming for prescription after self-diagnosis. Then, help is sometimes proportional to situation and lawful obligations.
Not only should AI misdiagnose to save lives, but a human should too. You walk in with symptoms that most likely is a harmless virus that clears up on its own or 5% of the time is a deadly bacteria. The correct course of action is to try to test if it is the 5% case (most often the wrong diagnosis), not send people home because they are most likely fine. Many cases have a similar low but not 0 risky diagnosis.
Unfortunately, from my understanding Doctors don't necessarily diagnose for accuracy, they often diagnose to limit liability.
They aren't going to take a stab at an uncommon diagnosis even if it occurs to them, if they might get sued if they're wrong.
Edit: I'm not trying to say Doctors deliberately diagnose wrong. Just that if there are two possible diagnoses, one common that matches some of the symptoms and one rare that matches all symptoms, doctors are still much more likely to diagnose the common one. Hoofbeats, horses, zebras, etc
The Guardian needs to raise their bar on what to report and how to give readers full context on the ongoing NFT AI trust me bro crypto scam and that context would be that it is a mathematical model of human language and not medical expert or replacement for one.
Fair enough. But there's lot of faulty and wrong peer reviewed research as well. One such paper comes to mind which is probably cited some 7000+ times in other papers but itself is wrong.
Humans could not diagnose and treat me correctly. They almost killed me. Curious where I could feed my symptoms and the same data I gave to an ER to an AI to test it.
I’d love to see a follow to that radiologist evaluation, where it failed so miserably on the thing it was supposed to be the best at that now there’s a shortage of radiologists.
Not an expert but what I’ve heard is that AI-based radiology analysis has brought down prices so much that there’s been a huge increase in demand, which has led to employee shortages.
I'd be very very hesitant to trust studies like this. It's very easy to mess up these benchmarks.
See for example this recent paper where AI managed to beat radiologists on interpreting x-rays... when the AI didn't even have access to the x-rays: https://arxiv.org/pdf/2603.21687 (on a pre existing "large scale visual question answering benchmark for generalist chest x-ray understanding" that wasn't intentionally messed up).
And in interpreting x-ray's human radiologists actually do just look at the x-rays. In the context the article is discussing the human doctors don't just look at the notes to diagnose the ER patient. You're asking them to perform a task that isn't necessary, that they aren't experienced in, or trained in, and then saying "the AI outperforms them". Even if the notes aren't accidentally giving away the answer through some weird side channel, that's not that surprising.
Which isn't to say that I think the study is either definitely wrong, or intentionally deceptive. Just that I wouldn't draw strong conclusions from a single study here.
I agree with you on this specific study, however, I can't really wrap my head about the fact that doctors will be better than AI models on the long-run. After all, medicine is all about knowledge, experience and intelligence (maybe "pattern recognition"), all those, we must assume that the best AI models (especially ones focusing solely in the medical field) would largely beat large majority of humans (aka doctors), if we already have this assumption for software engineers, we should have it for this field as well, and let's be realistic, each time I've seen a doc the last few months (and ER twice), each time they were using ChatGPT btw (not kidding, it chocked me).
So I’m genuinely curious:
What is the specific capability (or combination of capabilities) that people believe will remain permanently (or at least for decades) where a top medical AI cannot match or exceed the performance of a good human doctor? Let's put liability and ethics aside, let's be purely objective about it.
To answer your question: talking to a human.
Medicine is so much more than "knowledge, experience, and pattern matching", as any patient ever can attest to. Why is it so hard for some people to understand that humans need other humans and human problems can't be solved with technology?
> human problems can't be solved with technology
How are you defining technology? How are you defining human problems? Inventions are created to solve human problems, not theoretical problems of fictional universe. Do X-rays, refrigerators, phones and even looms solve problems for nonhumans?
Claiming something that sounds deep doesn’t make it an axiom.
One doctor didn't want to give me ritalin, so i went to another one.
One was against it, the other one saw it as a good idea.
I would love to have real data, real statistics etc.
"Human problems can't be solved with technology" is just wrong, unless you have narrower definitions of a "human problem" or "technology".
For instance, transportation is a "human problem". It's being successfully solved with such technologies as cars, trains, planes, etc. Growing food at scale is a "human problem" that's being successfully solved by automation. Computing... stuff could be a "human problem" too. It's being successfully solved by computers. If "human problems" are more psychological, then again, you can use the Internet to keep in touch with people, so again technology trying to solve a human problem.
So much of what I know from women in my life is that the human element of medicine is almost a strict negative for them. As a guy it hasn't been much better, but at least doctors listen to me when I say something.
The human doesn't need to be as highly trained and paid as a doctor if the human is not performing tasks concordant with that training.
I would personally vastly, vastly to go to a robot doctor, who diagnoses, treats and nurses me. What exactly do I need from a human here? Except of course being the one making the system.
Emotional support. Some human doctors absolutely radiate confidence and a kind of "you're gonna be okay" attitude. For me, this helps a lot. I'm not sure a machine can do this.
Technology is on a generational 10,000 year run of non-stop successfully solving human problems.
> What is the specific capability (or combination of capabilities) that people believe will remain permanently (or at least for decades) where a top medical AI cannot match or exceed the performance of a good human doctor? Let's put liability and ethics aside, let's be purely objective about it.
Being a human when a patient is experiencing what is potentially one of the worst moments of their life. AI could be a tool doctors use, but let’s not dehumanize health care further, it is one of the most human professions that crosses about every division you can think of.
I would not want to receive a cancer diagnosis from a fucking AI doctor.
On the other hand, health care is not scaling to meet the growing demand of societies (look at the growing wait queues for access to basic medical attention in most Western nations). The cause of this is a separate topic and something that deserves more attention than it currently gets, but I digress. If AI can fill the gap by making 24/7/265 instant diagnosis and early intervention a reality, with it then bringing a human into the loop when actually necessary... I think that is something worth pursuing as a force multiplier.
We're clearly not there yet, but it is inevitible that these models will eventually exceed human capability in identifying what an issue is, understanding all of the health conditions the patient has, and recommending a treatment plan that results in the best outcome.
You may not want to receive a cancer diagnosis from an AI doctor... but if an AI doctor could automatically detect cancer (before you even displayed symptoms) and get you treated at a far earlier date than a human doctor, you would probably change your mind.
You commonly receive very close proxies for diagnoses through MyChart already when results come back from the lab.
But liability and ethics cannot be put aside. If treatments were free of cost and perfectly address problems, then a correct diagnosis would always lead to the optimal patient outcome. In that scenario, AI diagnosis will be like code generation and go asymptotic to perfection as models improve.
But a doctor's job in the real world today is to navigate a total mess of uncertainty: about the expected outcome of treatments given a patient's age and other peoblems. About the psychological effect of knowing about a problem that they cannot effectively treat. Even about what the signals in the chart and x-ray mean with any certainty.
We are very far from having unit test suites for medical problems.
Isn't that conflating diagnosis and treatment plan?
Sure, but my anecdotal experience is that doctors do this regularly in real life, especially when choosing to diagnose or ignore problems that are unlikely to kill an aging patient before some other larger issue does.
Gotcha, I was thinking more about radiologists than patient-facing doctors.
> I can't really wrap my head about the fact that doctors will be better than AI models on the long-run.
Nobody said that though?
If the current trajectory continues and if advancements are made regarding automated data collection about patients and if those advancements are adopted in the clinic then presumably specialized medical models will exceed human performance at the task of diagnosis at some point in the future. Clearly that hasn't happened yet.
Until medical models can contrive of unique diagnosis, this will not be true and cannot be true.
Medical models can absolutely get better at recognizing the patterns of diagnosis that doctors have already been diagnosing - which means they will also amplify misdiagnosis that aren't corrected for via cohort average. This is easy to see a large problem with: you end up with a pseudo-eugenics medical system that can't help people who aren't experiencing a "standard" problem.
The pitfall you describe is not inconsistent with exceeding human performance by most metrics.
I'd argue that the current system in the west already exhibits this problem to some extent. Fortunately it's a systemic issue as opposed to a technical one so there's no reason AI necessarily has to make it worse.
Medicine is about knowledge, but acquiring knowledge may in fact require "breaking out of the box" that AI is increasing behind to avoid touching "touchy subjects" or insulting anyone and so on.
This study is based almost entirely on pre-existing "vignettes." In other words, on tests that are already known and have existed for years, the model did well, which is precisely what you should expect.
It provides no information on real world outcomes or expectations of performance in such a setting. A simple question might be "how accurate are patient electronic health records typically?"
Finally, if the Internet somehow goes down at my hospital, the Doctor can still think, while LLM services cannot. If the power goes out at the hospital, the Doctor can still operate, while even local LLMs cannot.
You're going to need to improve the power efficiency of these models by at least two orders of magnitude before they're generally useful replacements of anything. As it is now they're a very expensive, inefficient and fragile toy.
Weird that this is the case and a new study.
but those kind of x-ray models are already activly used. They are not used though as a only and final diagnosis. Its more like peer review and priorization like check this image first because it seems most critical today.
I think AI can be useful in any kind of context interpretation, but not make a decision.
Could be running in the background on patient data and message the doctor "I see X in the diagnostic, have you ruled out Y, as it fits for reasons a, b, c?"
I like my coding agents the same way, inform me during review on things that I've missed. Instead of having me comb through what it generates on a first pass.
These type of experiments are bound to have biases depending on who is doing it and who is funding it. The experiment is being funded for a particular reason itself to move the narrative in a desired direction. This is probably a good reason to have government funded research in these type of sensitive areas.
hallucination on steroids, wow. I had to read through the abstract to believe it:
"In the most extreme case, our model achieved the top rank on a standard chest Xray question-answering benchmark without access to any images."
I still don't quite understand, after skimming the paper. How does it achieve high scores without access to the images (beating even humans with access to the images)?
I think the bigger takeaway here is that 50% of the time doctors will miss what you have.
I believe in modern medicine but I lost some faith in the American institutions around it when I "diagnosed" my partner with the correct disease that the first rheumatologist dismissed and told them to just stretch. It was officially diagnosed years later, and we lost a lot of time because of it.
That's not a takeaway here at all.
It's 50% of the time ER doctors working solely from notes, something they never do, in a situation they know is only for a study, will miss what you have.
I'm even more concerned that current models are not trained to say no, or to even recognize most failure modes.
"Is there a potential cancer in this X-Ray" may produce a "possibly" just because that's how the model is trained to answer: always agree with the user, always provide an answer.
Oh, and don't forget that "Is there a potential cancer in this X-Ray" and "Are there any potential problems in this X-Ray" are two completely different prompts that will lead to wildly different answers.
I'm surprised at both the article and the paper - both seem very hyperbolic. This is LLMs competing against doctors in a way that is heavily weighted in the LLMs favour, which does not represent clinical practice. These reasoning cases are not benchmarks for doctors, they are learning tools.
I think it's important to note that diagnosis also relies on accurate description of the patient in the first place, and the information you gather depends on the differential diagnosis. Part of the skill of being a doctor is gathering information from lots of different sources, and trying to filter out what is important. This may be from the patient, who may not be able to communicate clearly or may be non verbal, carers and next of kin. History-taking is a skill in itself, as well as examination. Here those data are given.
For pattern recognition from plain text, especially on questions that may be in the o1's training data, I'm not surprised at all that it would outperform doctors, but it doesn't seem to be a clinically useful comparison. Deciding which investigations to do, any imaging, and filtering out unnecessary information from the history is a skill in itself, and can't really be separated from forming the diagnosis.
Believable and not shocking. LLMs literally may have saved my sons and potentially her mother too by allowing us to fact check a lot of non sense data and scare tactics by a group of at least 5 different doctors ambushing us to make a life changing decision in minutes. The problem is doctors, at least in the US, prioritize liability exposure over patients long term outcomes. Let’s say you need an intervention where two options A and B are available to you. A carries 1% risk of complications but a great outcome. Option B has 0.1% risk of complications but once you are discharged the short term effects are challenging and long term effects not well understood. Well, 10/10 times doctors will suggest option B and will do anything they can to nudge you into making that choice, like not telling you the absolute numbers and constantly using the word “death”. They also lie about the outcomes, because again, once you accept the procedure, sign and are sent home, they have nothing to do with you.
LLMs can be a useful second opinion for a highly educated patient with good insight into their health and body, but this is not the average patient I see in an urban emergency department. Many patients can't give a cohesive history without a skilled clinician who can ask the right questions and read between the lines.
I am very skeptical of studies like this that don't adequately reflect real world conditions, but when I was a software engineer I probably wouldn't have understood what "real" medicine is like either.
> "An AI and a pair of human doctors were each given the same standard electronic health record to read"
This is handicapping the human doctors abilities. There is a lot more information a human doctor can gather even with a brief observation of the patient.
On the other hand,
> there are few things as dangerous as an expert with access to open-ended data that can be interpreted wildly, like a clinical interview.
https://entropicthoughts.com/arithmetic-models-better-than-y...
Agreed. I think the best use of this sort of tech is to use both to their strengths. Use AI to go over the record and suggest diagnoses which you have the doctor review after observing the patient.
The other thing is that common issues are common. I have to wonder how much that ultimately biases both the doctor and the LLM. If you diagnose someone that comes in with a runny nose and cough as having the flu you will likely be right most of the time.
This feels like a deeply important observation. Now also, would be interesting to include e.g. a short video or photograph for the AI to use as well.
I don't think AI is a good use case for such critical situations. Maybe in a decade we have AI help out doctors with doing a pre check. What if Ai finds nothing and the doctor does not bother to look into it further? It is this small question which breaks the technology from any angle later down the road from my POV. AI has to stay optional here.
Even if AI is used to sample or summarize a lot of data that a human couldn't do in time: What if it misses something that a human won't? What if a human inversely misses something that AI won't? Would you rather trust the machine or the human? (Especially if the human is held accountable.)
As a 37 year old male with 2 THRs I'm glad the AI was NOT used in my diagnosis. All the models that I used to look at my x-rays said nothing was wrong, even when adding symptoms. When adding age it said the patient was too young.
(I was ~3 months away from wheelchair bound in those x-rays).
The worst one was Gemini. Upload an x-ray of just the right hip, and it started to talk about how good the left hip looked like.
I think with AI taking over it's gonna be harder to get a solution when your problem isn't the run-of-the mill.
The general AI models are useless if you need precision. They are designed to create/analyze pretty pictures.
But specialized models can be inhumanly good. I know, our main product is a model that does _precise_ analysis :)
I'd love to see the output of your system for my x-rays!
All versions and levels of Gemini have terrible spatial reasoning. I don't know why. That kind of task seems to be simply outside of the abilities of the model.
Besides for myself and wife, I've also used LLMs to diagnose my dogs. Convinced there's a huge opportunity for AI based veterinary, especially one which then performs bidding across the local veterinary clinics to perform the care/surgeries. I've noticed that local vets vary in price by more than an order of magnitude. My 80 year old mother and mother inlaw have been regularly scammed by over charging vets, and with their dogs being a major part of their lives, they extremely susceptible to pressure.
The Pitt third season leak? All of the ER is fired and Robbie is fighting schizophrenia with 15 agents and Dana?
The paper: https://www.science.org/doi/10.1126/science.adz4433 (April 30, 2026)
All the other points raised in this thread aside, it seems like an odd thing to benchmark because a significant proportion of ER practice is dealing with emergencies, often accidental injuries. There's not a whole of diagnosing going on if you show up to ER with a gash on your forehead or a missing finger.
radiology already had its "AI beats doctors" moment. radiologists are still here. what changed first was the workflow, not the specialty. er is probably next.
Yes, but what was the overlap
I’ve some family in medicine and it scares me how much they now rely on AI. Some even quote it like Bible.
I'll repeat my idea on how this MUST be done:
1. AI gets data about the patient and makes a diagnosis. This is NOT shown to doctor yet.
2. Doctor does their stuff, writes down their diagnosis. This diagnosis is locked down and versioned.
3. Doctor sees AI's diagnosis
4. Doctor can adjust their diagnosis, BUT the original stays in the system.
This way the AI stays as the assistant and won't affect the doctor's decision, but they can change their mind after getting the extra data.
5. Private Equity uses this valuable data to stack rank doctors based on how correct / AI-aligned their diagnoses are over time
6. Rankings are used to periodically "trim the fact" thus delivering more optimized cash flows to clinics that have been saddled with toxic debt
7. Sensing an opportunity AI providers start selling a $200 / month Data Leakage as a Service subscription to overworked physicians so that they can avoid the PE guillotine
A more realistic step 7 is that physicians gradually align their diagnoses with the LLM as they sacrifice to Moloch in order to (temporarily) game the metric. Eventually the humans become little more than an imperfect proxy for the LLMs and are eliminated.
I agree with GP's solution but we'd need regulation to prohibit what you describe.
This still promotes metacognitive laziness later down the road as the doctor can hand in something quickly and rely on AI to close that gap.
5. Doctors delegate everything to AI assistants because humans are lazy, especially if those AI assistants are correct some significant portion of the time
Then the claim may be that you don't need that many doctors anymore and that one doctor can do the job of X doctors in less time which has the economical effect that there is less demand for/supply of doctors, which then results in a home grown shortage of doctors, since less people are incentivized to become doctors...
o1 is several generations old and was released in 2024. Is this some quite old research that took a long time to get published?
It's also important to note that it beat doctors in diagnosing in a way doctors do not diagnose.
Yes, the preprint of the same paper (https://arxiv.org/abs/2412.10849) was first written in December 2024.
This is a rather new article about an old model...
Study design, data collection, analysis, and peer review take time. O1 came out a little over 1.5 years ago
I think this is more a commentary on how bad ER diagnosis is.
It is easy to overinterpret this based on the headline, the doctors were actually at a slight disadvantage. This isn't how they normally work, this is a little more like a med school pop quiz:
"I don't know, let's run more tests" is also a very important ability of doctors that was apparently not tested here. In addition to all the normal methodological problems with overinterpreting results in AI/LLMs/ML/etc. Sadly I do think part of the problem here is cynical (even maniacal) careerist doctors who really shouldn't be working at hospitals. This means that even though I am generally quite anti-LLM, and really don't like the idea of patients interacting with them directly, I am a little optimistic about these being sanity/laziness checkers for health professionals.I’ve had much better luck with diagnosis of my own family’s issues than with doctors. Usually now, I’m feeding them more information to begin with, so that their 30 minute office visits are not wasted, requiring another expensive follow up appointment.
While I’m sure there can be ways in which such studies are wrong, it’s very obvious that AI can accelerate work in many of these areas where we seek out professional help - doctors, lawyers, etc.
It can speed up some aspects of work, but please don't trust some llm with variable quality of output more than professional. If you don't like current doctor try another, most are in the business of helping other people.
If you have string of issues with 10 last doctors though, then issue is, most probably, you...
My wife is a GP, and easily 1/3 of her patients have also some minor-but-visible mental issue. 1-2 out of 10 scale. Makes them still functional in society but... often very hard to be around with.
That doesn't mean I don't trust your words, there are tons of people with either rare issues or even fairly common ones but manifesting in non-standard way (or mixed with some other issue). These folks suffer a lot to find a doctor who doesn't bunch them up in some general state with generic treatment. There are those, but not that often.
It helps both sides tremendously if patient is not above or arrogant know-it-all waving with chatgpt into doctor's face and basically just coming for prescription after self-diagnosis. Then, help is sometimes proportional to situation and lawful obligations.
would it ever diagnose incorrectly to save more lives? kinda weird an ai would decide who die so others may survive, but i guess whatever.
Not only should AI misdiagnose to save lives, but a human should too. You walk in with symptoms that most likely is a harmless virus that clears up on its own or 5% of the time is a deadly bacteria. The correct course of action is to try to test if it is the 5% case (most often the wrong diagnosis), not send people home because they are most likely fine. Many cases have a similar low but not 0 risky diagnosis.
Now show me the result of Triage Doctors with aided AI help
Unfortunately, from my understanding Doctors don't necessarily diagnose for accuracy, they often diagnose to limit liability.
They aren't going to take a stab at an uncommon diagnosis even if it occurs to them, if they might get sued if they're wrong.
Edit: I'm not trying to say Doctors deliberately diagnose wrong. Just that if there are two possible diagnoses, one common that matches some of the symptoms and one rare that matches all symptoms, doctors are still much more likely to diagnose the common one. Hoofbeats, horses, zebras, etc
The Guardian needs to raise their bar on what to report and how to give readers full context on the ongoing NFT AI trust me bro crypto scam and that context would be that it is a mathematical model of human language and not medical expert or replacement for one.
>The Guardian needs to raise their bar on what to report and how to give readers full context
Should they not report on peer reviewed articles published in Science? or only report published articles that fit your priors?
Fair enough. But there's lot of faulty and wrong peer reviewed research as well. One such paper comes to mind which is probably cited some 7000+ times in other papers but itself is wrong.
So we can eventually classify AI models as Software experts, but not as Medical experts, why so?
I don't classify them as software experts either. Anyone doing so is probably not an expert themselves.
I take them as those code generation command line tools like create react app and such.
We can't. It's just that everyone and their dog has an interest in selling you that lie because money.
Stochastic parrots can code yes, but that does not make them experts. Don't trust them with your life.
It’s a peer reviewed study in one of the world’s top science journals. It’s not some random person on a podcast.
Humans could not diagnose and treat me correctly. They almost killed me. Curious where I could feed my symptoms and the same data I gave to an ER to an AI to test it.
https://aistudio.google.com/
Chatgpt.com?
I’d love to see a follow to that radiologist evaluation, where it failed so miserably on the thing it was supposed to be the best at that now there’s a shortage of radiologists.
Not an expert but what I’ve heard is that AI-based radiology analysis has brought down prices so much that there’s been a huge increase in demand, which has led to employee shortages.
Did you hear this in the US or Europe?