11 ms·
HealthBench – An evaluation for AI systems and human health
- Zaheer 1y agoImpressive how well Grok performs in these tests. Grok feels 'underrated' in terms of how much other models (gemini, llama, etc) are in the news.
- tough 1y agoyou can't download grok's weights to run locally
- simianwords 1y agohow is that relevant here?
- tough 1y agoit helps explain why theres' less people talking about them than gemini or llama? less people using them.
- Insanity 1y agoI can guarantee you none of my friends (not in tech) use “downloading weights” as an input to select an LLM application.
- simianwords 1y agoisn't chatgpt the most used or most popular model?
- tough 1y agoYes OpenAI has a first-mover advantage and Claude seems to be close as a second player with their closed models too, open weights is not a requirement for success but in an already crowded market (grok's prospect) their preposition isn't competing neither with top tier closed models nor the maybe lesser-capable but more-available battle-tested freely available to run locally open ones
- reissbaker 1y agoYou can't download Gemini's weights either, so it's not relevant as a comparison against Gemini. I think the actually-relevant issue here is that until last month there wasn't API access for Grok 3, so no one could test or benchmark it, and you couldn't integrate it into tools that you might want to use it with. They only allowed Grok 2 in their API, and Grok 2 was a pretty bad model.
- tough 1y agolol sorry mixed them up w gemma3 which feels like the open lesser cousin to gemini 2.5/2.0 models
- moralestapia 1y agoIt's not. Also, only one out of the ten models benchmarked have open weights, so I'm not sure what GP is arguing for.
- tough 1y ago> in terms of how much other models (gemini, llama, etc) are in the news. not talking about TFA or benchmarks but the news coverage/user sentiment ...
- bearjaws 1y agoI am amazed Gemini did as well as it appears. Gemini frequently avoids discussing health problems, which likely hurt its scores. My guess is any censorship was considered a fail.
- ramon156 1y agoI don't want to be a conspiracy theorist, but could this be in preparation for Amazon's (to be) health branch?
- srameshc 1y agoIs the Med-PaLM model that Google's has been working on meant to be considered for comparison ? If I'm not mistaken, it isn't publicly available. > https://sites.research.google/med-palm/ https://sites.research.google/med-palm/
- aix1 1y agoMed-PaLM is old and has been superseded by (multiple generations of) Gemini.
- GuinansEyebrows 1y agoi have zero trust in openai's ability to do anything impartially. why should we leave the judgement of a private tool up to the makers of the tool especially when human lives are at stake?
- simianwords 1y agoI agree - we should exercise a bit of caution here. There is no way they would release a benchmark which makes their model look bad. But then again we know that their models are one of the best for other uses so its not a big leap to accept this benchmark.
- beezlebroxxxxxx 1y agoI can already see the pharma salesmen drooling at the idea of how various symptoms can be marketed to. "As an AI agent, a possible diagnosis is [xxx]. Ask your doctor about / look into [yyy™] for a possible solution!"
- ceejayoz 1y agoAnd OpenAI is definitely thinking about this on their end: https://www.axios.com/2024/12/03/openai-ads-chatgpt https://www.axios.com/2024/12/03/openai-ads-chatgpt > OpenAI CFO Sarah Friar told the Financial Times that OpenAI is weighing the inclusion of ads in its products but wants to be "thoughtful about when and where we implement them."
- barnas2 1y agoAd spots inside chatgpt are going to be worth an obscene amount of money.
- amarcheschi 1y agoI think that the damage of "chatgpt misdiagnose X as Y, person dies of Z" would be quite bad for PR
- dcreater 1y agoIsn't there an obvious conflict of interest when the model maker is also the creator of a benchmark? I think at the very least it should be from a separate business entity under the non profit or from the non profit holding entity itself
- gwd 1y agoI don't think it's necessarily bad to have the benchmark, but the graphs of Gemini and Claude doing worse than o3 did kind of leave a bad taste in my mouth. "Oh look, your models are worse than ours at this very important metric that we just made up! How terrible!"
- progbits 1y agoIt's also barely better, within the error bar on some metrics. So if this is the best they can do on their own benchmark that's pretty pathetic and shows they have no edge anymore.
- pizzathyme 1y agoNon-clinicians are using ChatGPT every day now to try to find assistance (right or wrong) to real-life medical problems. This is a great evaluation set that could prevent a lot of harm
- unsupp0rted 1y agoRecently I uploaded a lab report to chatGPT and asked it to summarize it. It hallucinated serious cancer, along with all the associated details you’d normally find on a lab report. It had an answer to every question I had pre-asked about the report. The report said the opposite: no cancer detected.
- maliker 1y agoInteresting. What LLM model? 4o, o3, 3.5? I had horrible performance with earlier models, but o3 has helped me with health stuff (hearing issues).
- unsupp0rted 1y agoWhichever the default free model is right now- I stopped paying for it when Gemini 2.5 came out in Google's AI lab. 4o, o4? I'm certain it wasn't 3.5 Edit: while logged in
- pants2 1y agoIf you're logged in, 4o, if you're not logged int, 4o-mini. Both don't score well on the benchmark!
- askafriend 1y agoThis gets at the UX issue with AI right now. How's a normie supposed to know and understand this nuance?
- unsupp0rted 1y agoOr a non-normie. Even while logged in, I had no idea what ChatGPT model it was using, since it doesn't label it. All the label says is "great for everyday tasks". And as a non-normie, I obviously didn't take its analysis seriously, and compared it to Grok and Gemini 2.5. The latter was the best.
- 1y ago
- iNic 1y agoI like that they include the "worst case score at k samples". This is a much more realistic view of what will happen, because someone will get that 1/100 response.
- pants2 1y agoThis appears to be a very thoughtful and helpful study. It's also impressive to see the improvement in performance in just the last year of model development - almost double. I've found o3 & deep research to be very effective in guiding my health plan. One interesting anecdote - I got hit in the chest (right over the heart) quite hard a month or so ago. I prompted o3 with my ensuing symptoms and heart rate / oxygenation data from my Apple watch, and it already knew my health history from previous conversations. It gave very good advice and properly diagnosed me with a costochondral sprain. It gave me a timeline to expect (which ended up being 100% accurate) and treatments / ointments to help. IMO - it's a good idea to have a detailed prompt ready to go with your health history, height/weight, medications and supplements, etc. if anything's happening to you you've got it handy to give to o3 to help in a diagnosis.
- quantumwoke 1y agoHow confident was it in that diagnosis? Did you trust its confidence if so? I asked this to my medico wife and apparently the other possibility is pericardial effusion from injury which can be life threatening.
- pants2 1y agoDefinitely - o3 proposed a few things initially: Costochondral separation, Intercostal strain, rib fracture, pericardial effusion and a few other things... After answering questions about symptoms, it gave me a few movements/actions to do and to rate the pain of each. Based on that it diagnosed me and gave me a timeline of what to expect and things to watch out for. Really it was as helpful as a doctor could have been in that scenario, IMO.
- nradov 1y agoSpO2 data from an Apple Watch (or any wrist device) is highly suspect and not reliable as a diagnostic tool. Under ideal conditions it can work reasonably well but it can also be thrown way off by multiple factors like wristband tightness, skin color, patient movements, dirt, etc. A cheap and simple dedicated pulse oximeter that clips on to your finger will generally be much more accurate and consistent.
- simianwords 1y agoI would really rather like a benchmark purely focusing on diagnosis. Symptoms, patient history vs the real diagnosis. Maybe name this model House M.D 1.0 or something. The other stuff is good to have but ultimately a model that focuses on diagnosing medical conditions is going to be the most useful. Look - we aren't going to replace doctors anytime soon but it is good to have a second opinion from an LLM purely for diagnosis. I would hope it captures patterns that weren't observed before. This is exactly the sort of thing game that AI can beat a human at - large scale pattern recognition.
- mrcwinn 1y agoHappy to see this. I've struggled with an injury for the past five years. I've been to multiple sports-focused physicians, had various scans. Responses from doctors have ranged from "everything seems fine, can't really figure this out" to [completely wrong hypothesis]. Tried acupuncture. Tried a chiropractor. I remember one doctor, though, had an interesting thought that seemed to make sense - but I've been so discouraged from so many false starts or misplaced hope, I didn't bother following up. Finally I typed in my entire history into o3-deep-research and let it rip for a while. It came back with a theory for the injury that matched that one doctor, diagrams of muscle groups and even illustrations of proposed exercises. I'm not out of the woods yet, but I am cautiously optimistic for the first time in a long time.
- Noumenon72 1y agoI hope recent cuts to government science have managed to hit enough of the safetyists and industry captures who keep us from just trying out new healthcare approaches like this and learning. They'd like nothing better than to replace the help you got with "As a large language model, I am unable to offer medical advice."
- candiddevmike 1y agoWhy would you trust a LLM over a battery of human experts? I find it hard to believe that the doctors never proposed exercises or some kind of physical therapy for you, at least in the US.
- BeetleB 1y agoI can't speak to the OP's condition, but having seen plenty of doctors and physical therapists in the US for over a decade: Yes, they propose exercises. No, they don't work. For certain (common) conditions, PT seems to have it nailed - the exercises really help. For the others, it's just snake oil. Not backed by much research. The current state of the art is just not good when it comes to chronic pain. So while I don't know if an LLM can be better than a battery of human experts, I do know that those human experts do not perform well. I'm guessing with the OP's case, that battery of human experts does not lead to a consensus - you just end up with 10 different treatments/diagnoses (and occasionally, one is a lot more common than the other, but it's still wrong).
- ray_ 1y ago[dead]
- kypro 1y agoWhy are all the label colours for the "Worst-case HealthBench score at k samples" chart the same colour and the same shape? Completely unreadable.
- brap 1y agoThe colors are slightly different. I think the person why made this was more concerned with appearance and not providing information.
- andy99 1y agoMy sense is that these benchmarks are not realistic in terms of the way the model is used. People building specialized AI systems are not, in my experience, letting users just chat with a base model, they would have some variant of RAG plus some guardrails plus other stuff (like routing to pre-written answers for common question). So what use case does this test setup reflect? Is there a relevant commercial use case here?
- dimitri-vs 1y agoYou could say the same for code benchmarks, no? For general medical Q&A I can't see how a specialized system would be better than base o3 with web search and a good prompt. If anything RAG and guardrail prompts would degrade performance.
- dontreact 1y agoIt tests people chatting to ChatGPT! That's a pretty big and important use case.
- programmertote 1y agoI have no doubt that a lot of garden-variety diagnoses and treatments can be done by an AI system that is fine-tuned and vetted to accomplish the task. I recently had to pay $93 to have a virtual session with a physician to get prescription for a cough syrup, which I already knew what to take before talking to her because I did some research/reading. Some may argue, "Doctors studied years in med school and you shouldn't trust Google more than them", but knowing human's fallibility and knowing that a lot of doctors do look things up on places like https://www.wolterskluwer.com/en/solutions/uptodate https://www.wolterskluwer.com/en/solutions/uptodate to refresh/reaffirm their knowledge, I'd argue that if we are willing to take the risk, why shouldn't we be allowed to take that risk on our own? Why do I have to pay $93 (on top of the cough syrup that costed ~$44) just so that the doctor can see me on Zoom for less than 5 mins and submit an order for the med? With the healthcare prices increasing at the breakneck speed, I am sure AI will take more and more role in diagnosing and treating people's common illnesses, and hopefully (doubt it), the some of that savings will be transferred to the patients. P.S. In contrast to the US system, in my home city (Rangoon, Burma/Myanmar), I have multiple clinics near my home and a couple of pharmacy within two bus stops distance. I can either go buy most of the medications I need from the pharmacy (without prescription) and take them on my own (why am I not allowed to take that risk?) OR I can go see a doctor at one of these clinics to confirm my diagnosis, pay him/her $10-$20 for the visit, and then head down to the pharmacy to buy the medication. Of course, some of the medications that include opioids will only be sold to me with the doctor's prescription, but a good number of other meds are available as long as I can afford them.
- imiric 1y agoGood lord. The idea that a system that produces pseudo-random output without any semantic understanding can be relied on to give accurate health-related information is deeply flawed and troubling. It's one thing to use these things for finding patterns in data, for entertainment purposes, and for producing nonsensical code a human has to fix, but entirely different to rely on them for health diagnosis or advice. I shudder at the thought that a medical practitioner I go to will parrot whatever an LLM told them. This insanity needs to be regulated yesterday.
- FlamingMoe 1y agoGood lord. The idea that a sleep‑deprived, bias‑prone mammal can be relied on to give accurate health‑related information is deeply flawed and troubling. It’s one thing to seek a human for empathy, small talk, or a routine vaccine, but entirely different to assume an overworked doctor will consistently out‑diagnose a modern data tool. I shudder at the thought that the practitioner I visit might simply repeat whatever a drug‑rep lunch pitch or outdated guideline left in their memory.
- stavros 1y agoThis insanity needs to be regulated yesterday.
- imiric 1y agoThankfully, it is. :)
- hiatus 1y agoAre there regulations regarding doctor's working hours similar to say, truck drivers? I found this [1] article which states doctors are limited to 80 hours a week whereas truck drivers are limited to 70 hours in a 7 day period. [2] [1] https://psnet.ahrq.gov/primer/duty-hours-and-patient-safety https://psnet.ahrq.gov/primer/duty-hours-and-patient-safety [2] https://www.fmcsa.dot.gov/sites/fmcsa.dot.gov/files/docs/Drivers%20Guide%20to%20HOS%202015_508.pdf https://www.fmcsa.dot.gov/sites/fmcsa.dot.gov/files/docs/Dri...
- yapyap 1y agoSam Altman does not care about “improving human health”
- NKosmatos 1y agoMost probably I’m going to get downvoted, but I’m gonna say it… It’s a pity they don’t support Greek language, keeping in mind that almost all medical terminology has Greek origins. Anyhow, this is a step in the good direction and for sure it will aid many people looking for medical assistance via ChatGPT.
- Sammyboys 1y ago[flagged]
- Quenby 1y agoI walked away from this with a feeling I can't quite put into words. I'm not a doctor, and I don’t really understand medical AI, but a friend of mine who is a doctor has been relying more and more on ChatGPT lately—to look up guidelines, organize his thoughts. He says it’s not that it knows more than he does, but that it’s fast, clear, and saves him time. That got me thinking. I used to assume AI in healthcare was about replacing people. Now, it feels more like an extension. Like doctors are getting an extra pair of hands—or a second consultation room that's always online. Maybe that’s what progress looks like: human judgment is still there, but increasingly shaped by structured input. I don’t know if that’s good or bad. It just feels complicated.
- tmpz22 1y agoThe problem is the healthcare industry will leverage this to place more responsibilities in the hands of increasingly less qualified individuals - as they're already doing with nurses. Costs will never go down, but the salary of healthcare professionals and quality of care will. AI is a tool to transfer wealth from the middle class to the ultra wealthy. If there is a brief intermediary period where quality of life improves for the poor, it's a temporary coincidence that hasn't been optimized away yet. Your. Costs. Will. Never. Go. Down. While. Maintaining. Equal. Service. In. The. Long. Term. That's. Not. What. Capitalists. Want.
- Quenby 1y ago[dead]
- mallardgryph 1y agothe amount of obvious astroturfing by people with investments in LLMs on this website is insane. you only need to look at this thread alone to prove your point. this is a race to the bottom that we're being accelerated towards by some of the worst people on earth.
- trollbridge 1y agoExactly correct. So eventually we’ll pay a nurse’s aide to type prompts into GPT the same price for a visit where we used to see an MD (and then an NP, and then a PA, and then a BSN, and then an RN…), and get charged the same.
- ziofill 1y agoIf your condition can easily be resolved by waiting a little and letting your body recover, an honest doctor will tell you so. I wonder if an AI will ever risk not recommending you to see a doctor.
- trollbridge 1y agoAn AI would absolutely tell you not to go to a doctor unless it’s got a prompt trying to keep it from saying that.
- tmpz22 1y agoWhich begs an important question. To what extent is American healthcare built around liability? There are a lot of innocuous things built into the healthcare system, like gating the cough syrup that gets you really high easily, behind a visit to a provider, that many people are just ignorant of as a virtue of the current setup. When we disrupt fields that we are not experts in we often miss oceans of nuance.
- AStonesThrow 1y agoNo, they will not, actually. At least in the lower-middle, Medicaid spectrum of care. Physicians at this level serve the labor market, and everything they do is ordered to keeping you able-bodied and working and on the job and not taking sick days. They have always been reluctant to dispense excuses from work, to document any sort of disability related trouble, or to give advice that does not involve more exercise or more medication. I have honestly approached multiple doctors in multiple health systems to complain about fatigue, syncope, and exhaustion, and they all go to extraordinary lengths to deny that anything is wrong or that anything should change.
- trollbridge 1y agoI use Medicaid and have never found that to be the case. In fact, I’d prefer to be back on full work duties, but I’m still on light duty until my next appointment, assuming he restores me to full duty. Providers who are reimbursed via Medicaid have zero incentive specific to Medicaid to not give out work “excuses”, document disabilities and so on. Generally speaking the same providers serve both Medicaid and non-Medicaid patients. Perhaps there are some states where the Medicaid providers are really awful (California?).
- Klaus_ 1y ago[dead]
- kristel100 1y ago[dead]