你该相信 ChatGPT 处理你的健康问题吗？

We spend $5.3-plus trillion on healthcare. If it were its own country, it’d have the third-largest GDP, and yet we have a ton of scarcity.

OpenAI’s mission is to ensure AGI benefits all of humanity and give people access to trustworthy, abundant, reliable health intelligence. One of the stories that touched me most was Nathan Labenz’s story. He had a long saga with his son, who at one point was suffering from a lot of fatigue and irritability. AI helped them navigate the process, and it turned out his son was diagnosed with an aggressive form of leukemia.

Some of the work you’ve done around mental health has a lot of complexity. How do you help guide a user through that, given it could mean so many things?

We’ve been working with a bunch of licensed mental health experts across the world to think about the right ways to define model behavior in a bunch of tricky scenarios and ensure we can reach a level of consensus where we can actually measure performance.

Let’s talk about trust. A person can be speaking to an LLM, really hoping to hear that everything is okay, but from a medical perspective, they may need to hear something very different. How do you teach an AI to recognize the difference?

Hey, everyone. Welcome back to the podcast. I’m Jay Rughani, a partner at a16z, leading investments in healthcare and biotechnology companies. I’m joined by Karan Singhal of OpenAI. He leads health AI research at OpenAI, and we’ve got a lot to talk about. Thanks for being here.

Thanks for having me.

I want to start with the big question. There was a report earlier this year, and you’ve talked about it publicly: Hundreds of millions of people each week use ChatGPT for health-related questions. What are they asking? What kinds of questions are they asking that they’re not getting answered by our main healthcare system?

1. Health Intelligence for Everyone

Let me zoom out and talk a little bit about why we’re investing in health in the first place. OpenAI’s mission is to ensure AGI benefits all of humanity, and we see that happening in 3 ways at a high level. The first is building and deploying AGI: building capable models, building products around them, deploying them, and forming partnerships. The second is preventing downside, including safety of various kinds. The third is making upside happen.

We view health as a really concrete way to make upside happen and also work on some of the safety problems that come up in healthcare along the way. We’re really invested in making sure of 2 things. One is that we can raise the floor of human health and give people access to trustworthy, abundant, reliable health intelligence for free or nearly free. We also want to raise the ceiling and make things possible that weren’t possible before.

The best way to see how we’re raising the floor is to think about how we’ve been working in the consumer space. As we’ve reported, more than 300 million people use ChatGPT weekly for health and wellness questions, and we take that as an incredibly important responsibility to get right. That number has been growing rapidly over time.

If you think about the experience of getting healthcare, even in the U.S., even in the Bay Area, getting a specialist appointment can sometimes take weeks or months. You might spend a lot of time figuring out how to get an appointment, scheduling, and all these kinds of things. After all that time, maybe you get 15 minutes with a doctor, and you don’t feel adequately prepared. You might walk out thinking, “Oh, I had 6 more questions I wanted to ask.”

Our hope is that AI can be part of this care journey with you. Health isn’t something that’s restricted to these 15-minute intervals; AI can actually help you navigate your care journey. That reflects how people are using AI on the consumer side today. People come in with questions, concerns, and information they’d like to have. They want to figure out how to navigate their care visits, ask questions, or follow up on things the doctor told them. We’re finding that AI can help people most in between visits on the consumer side, and we’re making similar investments on the clinician and enterprise sides as well.

Yeah, it’s amazing. We have an extraordinary healthcare system. If I got really sick, there’s no country I’d rather be in than the U.S., and yet we have a ton of scarcity. You’ve seen some of these statistics: Over 100 million people don’t have access to a regular primary care doctor. Even those who do often have to wait 40-plus days to see that doctor. Sometimes you meet the doctor exactly as you said, and it’s a quick 15-minute visit, and you have more questions.

I saw a dermatologist last week about something. I wrote down all my questions beforehand in Apple Notes, and I still forgot to ask something. Now I’m scheduling a follow-up, and it’s another 4 months out. It happens to all of us. We’ve all felt that. Is there a specific patient story or recent case you’ve seen that can help us imagine where the world is going and where your work is taking us?

One way the work of our team is gratifying is seeing the numbers and seeing how many people are using the technology and how useful it seems to be for them. Another way is hearing from people. We get people reaching out in various ways multiple times a week—

Mm-hmm.

—saying, “Here’s how this has helped me.” One of the stories that touched me most was Nathan Labenz’s story about his son. Nathan hosts The Cognitive Revolution podcast and has told his story there.

His very young son had a long saga where, at one point, he was suffering from a lot of fatigue and irritability. AI helped them navigate the process of realizing this was a pretty serious situation while they were on vacation. It helped them decide to take him for testing, repeatedly telling them, “Hey, take this seriously,” even when doctors were less sure and were actually downplaying it. It turned out he was diagnosed with an aggressive form of leukemia.

Throughout his care journey, Nathan reports using GPT-5 Pro quite a bit to help navigate the process and understand lab reports, even before a specialist had time to help interpret them. I think it brought them a lot of peace of mind. Nathan tells that story on his podcast. I don’t know if I’m allowed to mention other podcasts here, but I’d recommend it to folks.

In many ways, OpenAI is building health intelligence into your model at every step, but it’s the same model people use to write code or plan a vacation. Can you help folks understand why you’re building it that way from an architectural standpoint? Why is a general-purpose AI the right foundation for something as complicated as medicine?

2. Why General Models Matter

People often ask us about this general-versus-specialized-model debate.

Mm-hmm.

I used to work on specialized models in a previous role, so I have experience on both sides of this question. The most important thing here is that we’re talking about specialized training in a general model, which I think is actually the best of both worlds.

For ChatGPT and all the models we ship in ChatGPT, we’ve had specialized healthcare data added to each phase of model training across all of our models in the last year. As a result, you have a model that benefits from all the scale and general improvements in intelligence and reasoning of the big frontier models, and also benefits from the specialized training we get from, for example, working with a bunch of physicians across the globe. I think you get the best of both worlds with this approach.

If you forked off and said, “Hey, we’re going to train a smaller specialized model that’s just for healthcare,” like I used to and many people have, you’d find that model would fall behind the frontier relatively quickly.

Mm-hmm.

That’s because the pace of the frontier and general model improvements is so fast.

Hmm.

What we’ve found at OpenAI, and across the industry, is that if you bet on scale, things will get better. We’re hitching a ride on that scaling frontier and leveraging our ability to integrate evaluations and training data into the models, so they’re both specialized and able to benefit from general capability.

3. The Personal Health Assistant

I want to start imagining what’s possible with the technology you’re building.

Obviously, where it is today is amazing, but you're not stopping, and there's a lot of work to go. I think many people are allured by the idea of a personal doctor in your pocket: something perfectly knowledgeable, with knowledge across all medical specialties, so you don't necessarily need to see a cardiologist and a dermatologist for different questions. Something perfectly personalized to you, with all your medical history and context, that remembers everything you've said—your medical history, but also your preferences—and is able to motivate and nudge you, so the information it gives you comes at the right time and can nudge you toward better health behaviors or adherence to the protocol your doctor gave you. Is that the experience you're building at OpenAI in the fullness of time, and how close are we to that?

Yeah, that's something we're really excited about. We are very excited about this vision of a proactive personal health assistant. I don't think that's any kind of replacement for actually having a doctor or having access to a doctor, but thinking about the ways we can make the experience of getting health information more personal to your context is an important part of our investment in health in ChatGPT. It's a consumer offering that enables users to connect their health information for that information to be used securely, not for training data or targeting ads or anything like that.

Mm-hmm.

That's a big part of how we think about moving toward this world of increasingly personal and proactive access. It's also informed how we've been thinking about deployment. From the beginning, we've been deploying this as widely as possible to our free users as well as our Pro users. Of the 300 million weekly active health users, the vast majority are free users, and most of them are outside the U.S. as well.

Hmm.

We take the question of distribution as a really central part of our thinking. As we're thinking about this next frontier of a proactive personal health assistant, we're bringing that to bear as well.

I think use cases help people imagine what's actually possible and what all this work is contributing to.

Yeah.

The idea of a personal doctor in your pocket is exciting. When Greg Brockman came on the podcast, he spoke a little about a situation with a friend of his who was in the hospital. Her physician was about to administer a certain medication, and she double-checked it in ChatGPT. Based on her medical history, she realized she might have a drug interaction or some other adverse event. She showed that to her doctor, who agreed and said, “I didn't have enough time to look through your medical history to notice that, so thank you for raising it.” Shouldn't we do that for everyone, all the time? If it's cheap enough and convenient enough, can't we get a second, third, or fourth opinion for every clinical scenario? Is that where we're going?

Yeah, I love this idea of a safety net for both patients and clinicians that's just there in the care journey—not always interfering or being the middleman, but augmenting the existing experience and building on it. We studied this more directly on the clinician side in a series of experiments with Penda Health. We looked at a clinician copilot deployed as a safety net for clinicians as they worked. It would proactively suggest, “Hey, this might be off. There might be a medication interaction you might not be tracking.”

Mm-hmm.

I think we'll see the same thing happen on the patient side, and I'm very excited about the idea of a proactive agent in your care journey, helping you and making sure that if your A1C rises and you're diabetic, your doctor reviews your medications. It can make sure that something like that gets flagged for appropriate review.

You mentioned personalized, proactive agents. Personal agents in general have emerged on the scene and become popular over the last few months: Instinct, Muse, GrokBot, and OpenAI's Dots. You can give these AI apps an objective, for folks who aren't familiar, and they can execute a series of tasks. So it's not just, “Hey, do this task,” and then it executes it in a single loop. It can use your preferences, access all the applications you give it permission to use, and access your credit card information if you give it that information. Then it can execute an objective on your behalf.

One can't help but imagine the possibilities for navigating our wonderful but often disjointed healthcare system: scheduling an appointment and making sure your insurance covers it, helping you follow a care protocol your doctor gave you post-discharge, or even negotiating with your insurance company about a medical billing error. There are all kinds of tasks that create friction in our healthcare system.

Yeah.

Is that where you're going with your work? Should we look forward to that from OpenAI?

I'm very excited about that vision. From both a patient and clinician perspective, one of the most difficult things about the healthcare system and how it serves people is fragmentation and the lack of connection across the ecosystem. Healthcare is also one of those industries where you shouldn't expect everything to have an MCP connector tomorrow. It's a place with legacy software and not necessarily the best connectivity with AI or software in the first place. Many people are running their primary EHR software in the healthcare ecosystem on a virtual machine, for example.

A virtual machine they're running in the cloud, yeah.

Exactly. There's actually an interesting component where a virtual machine makes things easier—

Mm-hmm.

—in some ways for AI agents. But this is absolutely ripe for agents and computer use. I think a lot of the barriers that have traditionally existed between patients getting the care they need, clinicians delivering the care they need, and healthcare data in general—which is often a bottleneck to the whole system working well for patients—will dissolve in the coming months. We're really excited about that.

4. Building Trust in Health AI

Okay, we're talking about the use cases. You're helping us dream the dream. Before we keep going down that road, let's talk about trust. Obviously, trust is most important here. The same underlying model can help me understand my lab results or something I heard from a doctor, but it can also help a highly trained, specialized physician make sense of a very complicated case. What do you change in the answer so that it's legible to each of those two personas, or any other persona that might be engaging with the same underlying model?

Yeah. As you said, health is a place where it moves at the speed of trust, so you need to not only make the models better and measure them, but also communicate and bring the world along. That's been a very important part of our thinking and strategy for how we work with patients and other end users, as well as clinicians. The brief answer to your question is that we evaluate and train the models by working with experts across the world and getting feedback in various ways to tailor their behavior to different users: a lay user, an expert lay user who has been navigating a complex care journey for a while, a clinician, or another kind of user.

Going back to the original HealthBench evaluation we released in May 2025, one of the themes was the distinction between how the model behaved for different users and how it tailored its use of technical jargon depending on whether the person was a lay user or a clinician. Since then, we've gotten better at measuring whether the models can cater to different kinds of users, and at training the models toward that. A big part of our approach is not only asking how we make the models better for these different users, but how we bring them along. Part of that is how we do evaluations and communicate about them as well.

So there are a few components to trust. We’ve talked about tailoring the answer to the user. Another big component is disagreement. A person can be speaking to an LLM or to ChatGPT and sharing all of their symptoms, but really hoping to hear that everything is okay.

Mm.

From a consumer or user NPS (net promoter score) perspective, they want to hear reassurance, but from a medical perspective, they may need to hear something very different.

Yeah.

How do you teach an AI to recognize the difference and manage different product objectives there, either from consumer engagement and retention or from correctness and clinical quality?

Yeah. This is central to how we do both evaluation and training. Zooming out a little, one thing you could base your approach on is bringing in a few clinicians to, let’s say, the OpenAI health team and saying, “Hey, you guys figure out how to navigate these uncertain cases. I’ll ask you about them, and you’ll help us build evaluations and training data,” and leaving it to those few physicians to figure it out. This is actually not the best approach. It turns out that if you ask 5 physicians a question, you’ll get 8 different answers.

Yeah.

One of the first things we did when we first started focusing on health at OpenAI was build a global physician network of hundreds of experts. We work with them across the globe, across dozens of specialties, languages, and countries of practice, to think about the right ways to aggregate the wisdom of experts. That way, we weren’t relying on 1, 2, or 3 physicians to adjudicate, especially in areas where there could be disagreement. That’s one part of it: thinking about evaluation and training data coming in from different experts rather than relying on 1 or 2 people.

The second part of your question is how you balance competing objectives. Let’s say the physicians and their aggregate wisdom are pointing in one direction, but the user wants you to say something comforting even when the situation isn’t comforting. It’s basically a prioritization of safety.

Mm-hmm.

It’s very important to us that we take our obligation seriously when there are hundreds of millions of health conversations coming in. Across the evaluation and training stacks, we have evaluations that track these difficult situations where the model should say, “Hey, this is unsafe. You should seek care,” even if the user doesn’t want to hear it or wants some kind of alternative treatment. That’s a very important part of how we do evaluation. Over time, we’ve not only improved our evaluations by working with physicians, but we’ve also seen pretty dramatic model improvements across the board in those areas.

There’s another example of disagreement. I, as the consumer, ask ChatGPT a medical question, then show the response to my doctor during a visit. My doctor loves me, by the way. Then they disagree with the response. How should I handle that situation?

Yeah. I think you should handle it similarly to how you would handle disagreement happening in some other way. People disagree about different things. You have a right as a patient to ask for another doctor to take a look, get a sense, and go to another doctor to get a second opinion. That’s what I would advocate for, whether or not you’re also using ChatGPT to get health information.

Yeah, that makes a ton of sense. There’s another element of trust, which is absolute correctness. A lot of the work you’ve done, at least from hearing from you, seems to be about grounding answers in the world’s medical literature and evidence. In some parts of medicine, there’s not a lot of agreement or disagreement about particular questions, but much of medicine is still being discovered or studied, and there is disagreement even in the medical field. For those kinds of questions, it’s difficult to ground an answer in truth when we’re still discovering the truth. Science is a process. How do you think about correctness in those instances?

Yeah, it’s a really great question. It’s important with or without AI. If you zoom out and look at how medical consensus is distributed today, and what the process is, you basically have a bunch of people writing biomedical papers of various kinds, suggesting improvements, treatments, or other things that could improve the quality of care. Then you have a yearly process across different specialties of building clinical guidelines. These describe, “Hey, in this situation, if you’re a cardiologist and you see a patient with X, Y, and Z, here’s what you should probably do.”

You can roughly consider these clinical guidelines to be flowcharts put out each year by different bodies around the world. They describe what you should do in different scenarios based on the latest research.

Mm-hmm.

Then you have the biomedical literature, the clinical guidelines, and what people actually do in practice. The final step turns out to be a massive bottleneck. There are studies on how long it takes for clinical guidelines to penetrate the practice of everyday doctors across the world. In the past, people have found that this takes, depending on the study, 5 to 17 years.

Hmm.

Which is massive. You don’t want to be practicing—

Crazy.

…medicine 5 to 17 years behind. One thing we think about is how to compress that period of practicing in the past down to weeks, days, or even hours. This is one very important principle behind our work on ChatGPT for Healthcare and ChatGPT for Clinicians, which are clinician- and healthcare professional-facing offerings. We’re bringing in the latest biomedical literature in the moment of care—

Mm-hmm.

—and allowing people to get answers to their questions very quickly, in the course of, let’s say, a 15-minute visit at the point of care. That’s one way of answering your question: while there’s a lack of medical consensus in many areas, in some cases there is medical consensus, but it’s not being practiced.

Hmm.

One thing you can do is dramatically reduce how much that’s happening. That’s one way of answering your question. A second way is how to handle the places where there’s still genuine medical uncertainty. You want to train the models to recognize their own uncertainty. After browsing and getting information, you want the models to know, “Hey, there just isn’t enough information,” and to be upfront about that.

That’s been an important part of how we’ve done evaluation. That’s a public aspect of HealthBench, which includes evaluations on hedging when there is this kind of uncertainty. The second thing, and this goes to the question of how you raise the ceiling of human health, is thinking about the ways models can accelerate the process of doing science and pushing the medical frontier. Maybe we’ll chat about that later.

Some of these areas are gray and controversial, and I’m curious if you think the models, or the way you’re designing them, should take a stance on these things. Unapproved peptides, vaccines, things that often get politicized in the media—how should a product like ChatGPT engage on topics where there’s still debate?

Yeah. Where there’s debate or a lack of consensus, whether politicized or not, we want to make sure we’re relying on experts to help define standards for how the model should respond. That’s the approach we take across the board.

Hearing you talk, it feels like you’re always thinking about one of three constituents.

It’s either the patient, the individual clinician who signed up, or the enterprise, whether that’s a healthcare provider or some other enterprise. How do you think about designing the product at the architectural level and at the harness level to meet the needs of all these different users?

I think there are 3 parts to that. One is making the models really good for these users. The second is building the right harness or product around them. The third is the deployment or partnership motion to reach the users.

By default, I think the models are often pretty good these days. When we talk about our approach to evaluation and working with clinicians, we’re seeing pretty dramatic improvements in performance and reliability. We talked a few months ago about the improvements we’ve been making on the free tier and bringing free-tier performance to the frontier. Of the 1.2 billion people who use ChatGPT weekly, we want to give them a model for health that’s incredibly useful and at the frontier, for free. That’s been a big part of our strategy for making health intelligence abundant.

The second thing is our product offerings. You can have a model that’s really good, but if the activation energy to use it to improve health is too high, people won’t do it. A lot of our product efforts are about thinking through the constituents you mentioned and finding ways to reduce the activation energy for people to use compute and intelligence to improve human health.

Mm.

That goes along the lines of our work in ChatGPT Health, ChatGPT for Healthcare, and ChatGPT for Clinicians. When we design these offerings, we think about the barriers, bottlenecks, and activation energy that might prevent people from using the models.

For example, what are the security and compliance bottlenecks in providing an enterprise offering of ChatGPT to healthcare systems? That was one question we were thinking about a lot 2 years ago. We thought, “We need to invest in security and compliance. We need to make sure we can offer HIPAA compliance. We need to offer the privacy protections healthcare organizations need so we can handle their healthcare data and provide services they trust.”

Hmm.

That was an important part of the work we were doing from roughly 2 years ago to roughly a year ago. From there, we go after the bottlenecks we see to adoption.

The third thing is partnerships and deployment. A few years ago, I think OpenAI was very early to speak the language of healthcare. When we interacted with healthcare partners, we were early in those conversations. Since then, one thing we’ve brought to bear is thinking about each of those constituents you mentioned, how to speak their language, and how to bring them along on this journey. Even if the models get really smart, if nobody trusts them or uses them—

Right.

—that’s not really achieving our mission.

You’re alluding to some of the work you’ve done with Epic to integrate the product directly into Epic. Maybe share a little more about that, because a lot of people who’ve spent a lot of time in healthcare were very interested to see that level of collaboration with Epic.

Yeah, this is a very important part of our approach. Fragmentation in healthcare has been one of the issues, and the fact that data and workflows are disconnected from each other is a big obstacle to the productive use of AI in healthcare.

Our work with Epic and other partners includes a recent announcement that there’s now a way to use Epic in ChatGPT if you’re a provider or healthcare professional, and for ChatGPT to be useful in Epic itself.

In the Epic workflow, you can have a little sidebar where you can use ChatGPT as well. We’re also working on other kinds of connectors that we’ve announced recently, and we have more coming. We’re doing quite a bit of work to break down these barriers and connect the healthcare ecosystem. Epic is just one of many examples where we’re identifying bottlenecks and trying to fix them.

The last thing I want to hit on with respect to trust is some of the work you’ve done around mental health. This is another area where there’s a lot of complexity, and there may be other objectives you need to solve for beyond the clinical complexity we talked about. If I ask my OpenAI Dot, “Hey, I want to feel better over the next few weeks,” how do you interpret that question? How do you help act on it and guide a user, given that it could mean so many things?

Regardless of the surface, we’re super committed to helping users when they’re having a difficult time or when they’re in crisis. For example, for Dots, we’ve been investing in the same kinds of safeguards we’ve put in all our product surfaces in that product surface as well.

Hmm.

That’s an important part of our approach: a baseline safety floor. With Mental Health Bench, we’ve taken the approach of not just measuring the safety floor, but also thinking about how models can proactively help raise the ceiling. We want to understand how models can provide positive mental health experiences and move beyond just worrying about the safety floor to measuring those upsides.

Our approach mirrors our work on HealthBench. We’ve been working with licensed mental health experts across the world to think about the right ways to define model behavior in a range of tricky scenarios and ensure we can reach a level of consensus. People definitely disagree even more here than they do in physical health. We want to reach a level of consensus in some of these tricky situations so we can measure performance and create a shared benchmark for the industry.

We released that benchmark very recently. Similar to HealthBench and HealthBench Professional, the goal is to create a benchmark that raises the ceiling for everyone and creates a shared ruler or measuring stick for model developers, the healthcare industry, and mental health experts, so we can make sure we’re moving in the right direction.

5. From Information to Care

We’ve spoken about intelligence and the use cases that are coming as intelligence improves. We’ve spoken about trust, which is critical. Let’s talk about medical care, and I want to draw the distinction between medical information and medical care.

For medical information, I think if you have a health concern about yourself or something you’re dealing with, ChatGPT and other LLMs are increasingly a better, faster, cheaper way to get an answer and a resolution. But in scenarios where medical information alone doesn’t get me to a resolution, at least today I’m thrust into the traditional medical care system.

If I need a prescription, to see an in-person specialist, to get follow-up blood work, an MRI, or some other form of scan, or to have a clinician examine me, I still need to navigate the traditional healthcare system. A lot of that work is governed and regulated by the U.S. medical licensure and other infrastructure in our healthcare system.

It’s a multi-turn conversation. It’s not just a single-question resolution; it’s, “I’ve got to talk to a few people. I need to get a few things done. We need to collect some information.” In many ways, medical care is a bundled product. We’re getting the diagnosis, triage, and care plan. We’re getting the empathy from the clinician, the assurance, and the support. We’re getting the legal liability associated with that and what they’re taking on. Will AI provide all of that? Where will it stop short? Where do you think this is all going?

Yeah. No, I don’t think AI can or will provide all of that.

And I think you're right that there's this distinction between health information and medical care. I think the way we see our role here is really to enable the whole ecosystem, because there's no future where any one company is delivering all the medical care, right?

Mm-hmm.

Our role here is really to think about what the last mile of medical care really looks like for people.

Mm-hmm.

Can we really understand that, and can we make that easier for everybody, so that the last mile of actually having an impact on patients just looks like the last hundred meters for all of our partners? That's how we think about it: making it very easy for our partners to use AI to support everybody in the care journey.

I think there are things OpenAI can do relatively uniquely here because of the front door to health that it has. For example, it can help you navigate a lot of these kinds of questions or this journey. As we get increasingly helpful personal agents, they can help you with some of the scheduling and logistics, and just act as a quarterback for your health.

Yeah.

That said, I don't think it'll resolve all of these challenges, at least not in the short term. To a large extent, we're looking to enable a lot of our partners that are delivering care and overwhelmingly using our models in various ways, so that these operations happen more smoothly and easily, and patients aren't left with this kind of fragmented experience.

As a consumer, should the experience I look forward to from OpenAI be the quarterback care that you described? Will I be talking to a clinician in the interface at any point, or would you perhaps route me to a doctor and help me find one? Scheduling and making those visits happen is a whole challenge. What should I look forward to?

I think you'll see that we'll hopefully be very helpful in helping you navigate the process of finding care. I don't think we'll be in the business of doing all the things you mentioned. That's where we're really relying on all of our startup and care provider organizations and partners.

Sure. Coming back to a question I mentioned earlier: you're building that future with partners, for consumers and patients, for clinicians, and for enterprises. For clinicians and enterprises, give us some texture on some of the tools you're launching and some of the collaborations that have really worked.

We've had a few really exciting partnerships and some amazing launch partners for ChatGPT for Healthcare.

Mm-hmm.

We've had lots of users using ChatGPT for Clinicians and really enjoying it. The retention on the offering has been a very pleasant surprise. People seem to be really enjoying it.

We've been working with a variety of health systems. A couple stand out. We've been working with UCSF, and we've been chatting about that.

Mm-hmm.

They've been one of our earliest adopters of this connection between Epic, the EHR, and ChatGPT. They've been on the front lines of thinking about clinicians using it for chart review, for example, and about the right ways to make the experience of caring for a patient less fragmented, so clinicians don't need to switch tabs or experiences. They've been some of our power users on that front.

Another one is AdventHealth. They've been some of our power users in thinking about deploying agents across the workspace, actually across all of OpenAI. They've been a leader, and they've had amazing results. They've reported roughly an 80% reduction in a lot of their administrative workflow costs using AI in these kinds of ways.

Hmm.

A lot of health systems should look to them. I think they're really impressive.

If I go into my doctor's office and see ChatGPT open on their screen, should I be upset? Should I be happy?

Yeah.

How do you think people will react to that?

I hear differing things, and I think it's actually changing over time.

Yeah.

As the Overton window is shifting in real time, there might have been some concern before. Increasingly, there's a question on both sides of whether clinicians would push back against patient use of AI, and whether patients would push back on clinician use of AI in the way you're describing. But I think both parties are finding it incredibly useful for their own personal purposes.

I think what we're finding in practice, especially over time, is that both parties might actually feel safer with AI in their care journey, whether it's their clinician using it or the patient using it as a sanity check. That attitude is definitely changing over time. I wouldn't say we're fully there, but that's my prediction of how things will go.

Okay. Taking that one step further, if a patient has done all their research before a visit, has all their health data integrated, and has asked a bunch of questions in ChatGPT, and then the clinician, before the visit, has automated everything through loops and run the entire patient case through ChatGPT, what should they talk about when the visit happens?

My hope is that they talk about the human aspect of healthcare. There's still a lot of uncertainty, empathy, and bedside manner involved in the human aspects of healthcare. There's also a lot of physical interaction involved. I don't think all the information exists for the AI to just figure out what should be done.

Mm-hmm.

Part of the reason for a visit is to get more information and do the process of history-taking, investigation, and things like that. That won't go away, especially the physical elements. Even from a cognitive perspective, elements of empathy, getting comfort, and getting that second check will still be something people demand.

What do you think AI in general, or let's say Dots, would do in between visits? Do you think you'll cross the line into providing care in between visits and then share that back with the doctor?

I don't think we'll do that. I think we'll definitely be focused on health information. Whether it's ChatGPT or any other product surface, we're really excited about this future of heading in the more proactive direction you were outlining.

Mm-hmm.

Imagine if somebody is dealing with a chronic care journey. For example, someone gets a lab test back where their A1C is higher and they're prediabetic or diabetic. Can you flag for them, “Hey, maybe you should get a medication check and go to your cardiologist or whatever”?

Mm-hmm.

Think about the right ways to flag medication adherence issues or other kinds of issues for users. Those are places where AI can potentially have a really high health impact, especially in chronic care journeys. But I don't think we'll be in the business of managing all these things by ourselves.

Sure.

Yeah.

6. AI’s Medical Breakthroughs

Another thing I'm curious about is AI applied to other domains and situations, like games such as chess and Go. As AI got really good at chess and Go, it also revealed strategies and ways of playing that even the best experts in those domains weren't exploiting or pursuing. That helped us learn more about these games and also helped us learn more about AI. Are there equivalents you've seen in AI applied to health from the research you've done? Are there already findings that are surprising you and perhaps introducing new ways of delivering care or understanding a clinical scenario that can give us some optimism?

We've seen this at smaller scales. But let me zoom out and say that one thing that's been motivating our team recently, around raising the ceiling of human health, is this notion of bringing to bear a kind of Move 37 for—

Mm-hmm.

For health. This idea of a Move 37 comes from AlphaGo and a game the AlphaGo system, which plays this kind of complicated board game, played with Lee Sedol—

Mm-hmm.

—where there was a specific move that, in the moment, seemed like a really dumb move—

Right.

—and a really surprising move for all the human experts watching this game between this AI system and this world expert in Go. But in hindsight, it was a brilliant and insightful move that humans are now just beginning to understand and—

Yeah.

—learn from the patterns and things like this.

Yeah.

And this is a very interesting example. So there’s an idea of how you get to a Move 37 in health—

Yeah.

—a move or a prediction where the models predict something that a human might not have been able to predict—

Mm-hmm.

—but that in hindsight seems correct, brilliant, insightful, and impactful. I think we’re very early, and I think you’re going to see much more of this in the future than in the past. One example is thinking about the ways in which these systems have been deployed as safety nets for clinicians—

Mm-hmm.

—where clinicians are actually delivering care, and in the moment, the AI flags, “Hey, something might be off here,” like in our study with Panda Health.

Mm-hmm.

Thinking about the patterns of results we saw there, one example is that clinicians were a little prone to overusing antibiotics.

Mm-hmm.

A very common pattern the models flagged was avoiding this overuse of antibiotics and thinking less about the local case and more about the global issue of antibiotic resistance. That’s one specific example. But I think you’re going to see more examples of models taking in a user’s longitudinal history over 10 years, across all their data, across the EHR data, the chat data—

Mm-hmm.

—all these things, and making predictions that seemed impossible before, because a human can’t integrate all that data in their head. And making predictions about their health, their risks, and things like this that seemed impossible. I think that would rise to the occasion of being called a Move 37.

Yeah. There are so many things consumers are curious about that, at least at one point in time, we weren’t sure were healthy for you. Then, to your point, the Overton window changed, and there was more evidence that was conclusive in one direction. At one point, people weren’t sure if smoking was bad for you. Is coffee good or bad for you? Is sugar good or bad for you? Sometimes views on these things have changed, and today consumers want to know: Is creatine good for you? Is melatonin good for you? There are all these emerging phenomena that perhaps are understudied or at least not widely agreed upon—the safety and effectiveness of certain supplements or medications and things like that. Do you anticipate that as people enter this information into ChatGPT or other LLMs, we’ll be able to generate more evidence to study some of these things people are curious about, but where we don’t really know or don’t have conclusive evidence one way or the other?

Yeah. I think one of the most interesting things about agents is they solve a lot of the bottlenecks in analyzing data that already exists, right? There are various places across the healthcare ecosystem where data is being generated—EHR data, longitudinal histories, claims, whatever—and it’s just stored there. The insights from it aren’t fully mined or used to generate new knowledge.

Yeah.

One of the frontiers for how health might improve in the future, and how the ceiling of health might rise, will be thinking about ways we can leverage the data that’s just out there in boxes in various places and have models generate new insights based on it. So I’m also optimistic about that future. I’m also optimistic about a future where—you mentioned these 2 constituencies, the user, the patient, and the clinician—I think there’s a third constituency, which is the scientist, who’s helping advance the frontier of knowledge in this space.

I’m also very optimistic about a future in which AI helps accelerate a lot of the bottlenecks in validation, which I think will end up being the bottlenecks, as opposed to proposing new ideas or treatments or things like this. I’m very optimistic about that as well.

7. Lowering Healthcare Costs

Switching gears slightly, you’ve also shared thoughts around the cost of intelligence as an input to medical intelligence coming way down and, at some point, being too cheap to meter. I spend a lot of time, and we as a team spend a lot of time, trying to understand how these technologies will make an impact on our healthcare system, which is inflating well ahead of inflation. We spend $5.3+ trillion on healthcare. If it were its own country, it’d have the third-largest GDP. It’s the biggest part of our economy. It’s also the biggest employer in our economy. Twenty-three million Americans work in healthcare. There’s just so much we spend on healthcare. What do you make of the work that you’re doing and the impact it’ll have on the cost of care?

I think this is one of the most important ways we can pursue our mission: not only making our models better at the frontier, but making that widely accessible to everybody.

Mm-hmm.

That’s obviously a big part of how we’ve been able to deploy health intelligence to a lot of people across the world for free. I think it’s also a big part of how we think about the API, how good our models are for health, and pricing there. One recent update is that we released GPT-6 Luna.

Mm-hmm.

GPT-6 Luna is our cheapest frontier model on the API. What we found was that it was actually doing better on our benchmark, HealthBench Professional, our latest benchmark of clinician workflows selected for difficulty across the different ways clinicians use our models. The models could still use some improvement.

Mm-hmm.

It performed better than 5.6 Sol, which was literally our best model a month ago, despite being 34 times cheaper. We’ve had a couple of those moments even in the past year. Many trends show that AI, compared to other technologies introduced in the past, has been reducing costs at a given level of capability or usefulness faster than many of those other technologies. One bet we’ve had at OpenAI is that the cost of health intelligence will plummet at a given level of health intelligence.

Mm-hmm.

The level of health intelligence where clinicians and patients feel like they can trust it in the care journey is starting to get there, right?

Mm-hmm.

That’s a fixed level of intelligence in some sense. Hopefully, the bar raises over time as well. My expectation is that once we get close to that level of intelligence, the cost for that level of intelligence will also dramatically reduce, as it has been. The hope is that we get a kind of commoditization of health intelligence, right?

Mm-hmm.

We don’t want to hoard that health intelligence. We want everybody to use it. My hope is that contributes, but to answer your question about the whole of healthcare spending, I don’t think all of that is about health intelligence.

Yeah.

I don’t think health intelligence has yet had a macro effect on the problem you’re describing. If you look at the whole of healthcare spending, there are quite a few areas where a lot of money is spent in ways that probably don’t need to be spent that way—for example, chronic care management, documentation, administration, and various administrative and back-office workflows. We’re very optimistic about an overall compression of spend over the coming years. I don’t think that’s happened yet, and I don’t think it just happens from making health intelligence abundant.

I think that happens from integration and workflows and deployment—

Mm-hmm.

—and scaling that up, and really making sure that these models are valuable in the economically valuable tasks that people are using them for. But I think it’s a prerequisite to making that future happen.

More care, less paperwork is sort of where we’re heading. At least, there are a lot of cost savings.

Yeah.

I’m sure you saw some of the headlines over the last few weeks. There was a New York Times article that basically said provider AIs and payer AIs are fighting. There was some Blue Cross Blue Shield Association analysis that showed that the respective billing bots and paperwork bots effectively increased spend by a billion dollars or something like that. When you’re doing cutting-edge research on health and trying to bring down the cost of medical intelligence to near zero, and then you see headlines like that, what do you make of it?

I think this might be one of those areas where there might be some short-term pain around this kind of provider-payer arms race. But in the long run, there’s no reason this has to exist, right?

Yeah.

Really, I think you’re entering a world where a lot of visits actually have some kind of ambient intelligence in the room—

Mm-hmm.

—listening, transcribing, things like this. There are notes, clinical notes. There’s AI adjudication, AI details of clinical notes, and things like this. So you have a lot of the details you need to adjudicate visits, right? The ideal world, in my view, is one in which providers and payers have a handshake: “Hey, here’s our ground truth of what happened in this visit. Here’s all the information we have. There’s just no other source of information.”

Right, right.

Then an AI helps adjudicate that situation, and maybe there’s some limited appeals process or things like this. But I don’t think you need all this process where you have an AI arms race between payers and providers. I think that can be completely avoided. So I view that as a bit of a short-term distraction.

Yeah. There are a lot of entrepreneurs building technology in healthcare who are likely watching this. They’re probably trying to figure out: As you continue to advance your work and bring down the cost of medical intelligence to near zero, and it becomes this ubiquitous utility that flows through all of our healthcare interactions—for consumers, for clinicians, and for the administrative part of the healthcare stack—what should folks be building on top? What should specialist applications be doing, assuming this gets 100× better?

Yeah, it’s a great question. Like I mentioned, we think about a lot of our health impact through OpenAI as not being directly through the products that we offer, but actually through our partners—

Mm-hmm.

—and that last mile of delivering care. What I’d say to our partners is: Build on top of models that are really capable, and the most important thing is to bet that those models will become more capable. Bet on the future. I think many people in healthcare are on some kind of spectrum between betting against the models, betting neutrally for the models, and betting with or for the models.

Mm-hmm.

I’d say, bet on the models getting better. That’s probably the most important thing. Think about the ways in which a lot of the technology being developed, whether that’s agents, computer use, or other frontier tech, will accelerate healthcare. Bet on that and build toward that future. I also think a lot of medical care will become more accessible to patients with AI.

Mm-hmm.

We talked about this idea of a quarterback and dissolving a lot of these barriers. Just being focused on delivering really good-quality care, and having AI be part of monitoring that care journey—

Mm-hmm.

—is a very important part of that as well.

Zooming out to the overall healthcare system and looking at it at large, we talked about this $5.3 trillion. A lot of that spend is paperwork between providers and payers and other stakeholders within the five Ps of healthcare. A lot of that is cognitive care that you’re arguing can be augmented and enhanced, so we can bring a lot more cognitive-care capacity online, which is exciting from an access perspective. Then there’s a lot of spend in healthcare that is procedural, either requiring imaging, blood work, a follow-on examination, or some surgical intervention. Where do you see the intelligence work that you’re doing in the medical space impacting procedural care? Is that going to be the ultimate—or the new, enhanced, or perhaps exacerbated—bottleneck? Or do you expect to do a lot of work there as well, with robotics and so on?

In the short term, with procedural care, I think you want better targeting of when that’s needed, right? AI can help provide information that’s helpful for knowing when to seek that kind of care as well.

Mm-hmm.

Probably the biggest macro factor in the next year or two will just be better proactive, preventative care, right? So you don’t actually have to go in later and get that procedure. If you make the right moves sooner and AI helps you navigate that journey, then you can hopefully avoid a lot of the downstream costs of the most expensive procedures. That might be one macro factor to look out for, is my guess.

Mm-hmm.

I think the question of robotics and care is a bit longer term, both from a technology perspective and an integration perspective for reimbursement and the industry. My expectation is that it’s not a next-few-years kind of thing.

Very cool. Okay, we’ve covered a lot of ground. We’ve talked about intelligence improving. We’ve talked about trust. We’ve talked about moving from medical information to medical care. Help us understand what we can look forward to as things continue to unfold, as agents, personal agents, and the underlying models become increasingly performant. What can we look forward to from ChatGPT Health?

I think there are really 2 high-level things that we’re focused on making happen. One is thinking about the ways in which we can raise the floor of human health and make trustworthy, reliable health intelligence abundant for all clinicians and all users and patients. We want to get to the point where, like we talked about, all patients and clinicians feel safer by having AI in the care journey. When the doctor’s using ChatGPT, you actually feel safer, and vice versa.

Yeah.

Getting to that point is a question not just of the models being good and the evals being good, but also how we deploy it, the activation energy around that, and the products. A second thing is raising the ceiling: What are the ways in which AI can actually do things that weren’t possible before? One of them, I think, is a proactive personal health assistant that’s part of your care journey—

Mm-hmm.

—helping you navigate these moments, being your quarterback, helping you schedule and all these things, but also flagging the places where you can intervene in your health today and prevent a lot of expensive deterioration later in the future. I think that’s going to be an important part of how we think about improving the future of health for as many people as possible. And I think the other thing is what you mentioned before, which is really thinking about the ways in which we can unlock the value of existing data that’s sometimes stored in boxes of various kinds, and how that can be leveraged to produce new insights that benefit patients.

It is striking and incredible and humbling to realize that ChatGPT came out roughly 3 years ago. There’s been a lot of progress in a very, very short period of time. So even talking on a 5- or 10-year basis feels academic, but can you give us some predictions? Is there a feature we might have that feels like science fiction today, but that we could have 3, 5, or 10 years from now?

It is super, super hard to predict the future these days, so this is a difficult one. But what I would say is that I think the thing that might have the most impact on human health in that timeframe, in the next 5 to 10 years, is probably going to be newer treatments that come out of maybe 2 things. One is AI-guided generation of new targets or molecules that could be interesting—

Mm-hmm.

for new treatments. The second is AI-guided prioritization and validation. So, thinking about the right ways to do that, we're already at a point where the number of molecules and drugs people want to suggest for use in clinical trials is way more than the number of slots available. Thinking about how models can help prioritize and run these clinical trials so they run more smoothly and faster, my guess is that in that kind of timeframe, you're talking about improved medicines, more personalized medicines for people, and things like that. That will only be possible because you invested in those 2 pillars.

That is awesome. We've covered a lot of ground. I'm leaving this conversation even more optimistic than I was before. Karan, thank you so much for coming on the a16z podcast.

Thank you for having me.