It's hard to draw any conclusion from this study precisely because of this. Since 2024, we went from AI being able to do a few minutes of coding work to now a few weeks autonomously. That's like going from an intern to staff engineer level.
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.
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.
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
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.
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.
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'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 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.
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 it's plausible since doctors tend to have human cognitive biases and miss things. People tend to fixate on patterns they're most familiar with.
When you read through the article it shows that the gap between doctors and LLMs actually disappeared (in terms of statistical significance) once both were allowed to read the full case notes.
The headline is quoting a number based on guessed diagnoses from nurse's notes. The LLM was happier to take guesses from the selected case studies than the doctors is my guess.
I haven't finished reading the linked paper, but I'm intrigued by the assumption that the results show illusion or mirage results when not giving access to the x-rays.
It seems like a very reasonable take away, but it skips the other one. Do x-rays make results less accurate?
In a study like this, there’s also a difference in motivation. An AI will mechanically “take the study seriously.” I’m not convinced the doctors will.
But when making decisions about a real patient’s care, a doctor will be operating under different motivations.
They can also refer patients to a specialist, defer a diagnosis until they have more information, use external resources, consult with other doctors.
Doctors aren’t chatbots. They are clinical care directors.
Presuming there are no issues with information leakage, it’s genuinely impressive AI can perform this level of success at a specific doctoring skill. That doesn’t make it a replacement for a doctor. It does make it a useful tool for a doctor or a patient, which is exactly what we’re seeing in practice.
Why? Simply because there is a plethora of "studies" from the AI industry benchmaxing? Or that every single time the outcome is in favor of the tools then when actually checking the methodology they are comparing apple and oranges? Truly I don't get your skepticism. /s obviously.
Jokes aside whenever I read about such a study from a field that is NOT mine I try to get the opinion of an actual expert. They actually know the realistic context that typically make the study crumble under proper scrutiny.
Yup, there's a reason while ROC is a thing in data science. You can build a 99% accurate cancer detector that's just a slip of paper saying 'you don't have cancer', but everybody understands its worthless intuitively. With more complex setups, that intuition goes away.
Ultimatly you'd want humans and AI to study separately cases separately and independtly, and flag cases that have been found by only one analysis so that a separate analysis is done by a second pair of eyes.
Definitely not a "fair" test... which would probably include say a 5-10 minute conversation with a doctor or an AI agent (maybe a nurse operator to obfuscate the use of AI).
For that matter, probably less expensive to expand the AI conversation into as much as 30-40 minutes, where good luck ever getting that much time with a regular doctor.
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.
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.
Also, this is not how ER doctors work? They are not trained for this, nor does it reflect their day-to-day performance. If they would work like this, perhaps they would know a bit more about the nurse writing down those notes, and the kinds of things that particular nurse is likely to miss or overemphasize - just as an example.
The article gives a neat example: In one case in the Harvard study, a patient presented with a blood clot to the lungs and worsening symptoms. Human doctors thought the anti-coagulants were failing, but the AI noticed something the humans did not: the patient’s history of lupus meant this might be causing the inflammation of the lungs. The AI was proved correct.
Which is nice and all, but in the presence of a blood clot, I can understand that treating inflammation instead is not the first thing on a doctor's mind, what with blood clots being potentially life threatening and all. It raises the question; was this a real-life case, and what happened to that patient? Since this is a case for which the correct diagnosis is known, it was eventually correctly diagnosed - presumably then the patient did not die of a blood clot, nor of an uncontrollable fever.
Also, how representative is a patient with Lupus? According to House, MD, it's never Lupus.
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.
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.
> "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.
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.
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.
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.)
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.
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.
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.
97 comments
[ 2.8 ms ] story [ 79.3 ms ] threadEvery sniffed out systematic service overcharge can be aggressively undercut by competition.
"Your margin is my opportunity", etc.
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
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.
"In the most extreme case, our model achieved the top rank on a standard chest Xray question-answering benchmark without access to any images."
"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.
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.
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.
The headline is quoting a number based on guessed diagnoses from nurse's notes. The LLM was happier to take guesses from the selected case studies than the doctors is my guess.
It seems like a very reasonable take away, but it skips the other one. Do x-rays make results less accurate?
But when making decisions about a real patient’s care, a doctor will be operating under different motivations.
They can also refer patients to a specialist, defer a diagnosis until they have more information, use external resources, consult with other doctors.
Doctors aren’t chatbots. They are clinical care directors.
Presuming there are no issues with information leakage, it’s genuinely impressive AI can perform this level of success at a specific doctoring skill. That doesn’t make it a replacement for a doctor. It does make it a useful tool for a doctor or a patient, which is exactly what we’re seeing in practice.
Why? Simply because there is a plethora of "studies" from the AI industry benchmaxing? Or that every single time the outcome is in favor of the tools then when actually checking the methodology they are comparing apple and oranges? Truly I don't get your skepticism. /s obviously.
Jokes aside whenever I read about such a study from a field that is NOT mine I try to get the opinion of an actual expert. They actually know the realistic context that typically make the study crumble under proper scrutiny.
From my limited experience hanging on ER hallways for other people, they don't look at the notes, they look at the damn patient.
I know it might look like a loss for radiologists, but I don't see it that way. More like you can't trust these studies.
1. https://www.npr.org/sections/health-shots/2013/02/11/1714096...
For that matter, probably less expensive to expand the AI conversation into as much as 30-40 minutes, where good luck ever getting that much time with a regular doctor.
The article gives a neat example: In one case in the Harvard study, a patient presented with a blood clot to the lungs and worsening symptoms. Human doctors thought the anti-coagulants were failing, but the AI noticed something the humans did not: the patient’s history of lupus meant this might be causing the inflammation of the lungs. The AI was proved correct.
Which is nice and all, but in the presence of a blood clot, I can understand that treating inflammation instead is not the first thing on a doctor's mind, what with blood clots being potentially life threatening and all. It raises the question; was this a real-life case, and what happened to that patient? Since this is a case for which the correct diagnosis is known, it was eventually correctly diagnosed - presumably then the patient did not die of a blood clot, nor of an uncontrollable fever.
Also, how representative is a patient with Lupus? According to House, MD, it's never Lupus.
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.
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.
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.
(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.
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.
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.)