Fireside Chat with Nate Gross and Lloyd Minor | Stanford's RAISE Health Symposium 2026

Well, thank you, Dean Miner, and really appreciate everybody having me here today. This is an incredibly important time to be having these conversations, and there's no better place to be having it. I was very lucky. I think there have been a lot of different technology waves that have affected medicine and I was lucky to be entering medical school when the iPhone came out, the newsfeed came out, the app store came out, WhatsApp came out, the like button came out. Each one of those things affected pretty quickly about a billion people on the planet or more and how they communicated or how they distributed knowledge and information. And what a privilege to be able to see that wave, see it affect what your friends were doing, what your patients were doing at the same time as we were still on paper charts at Grady and getting to watch that digitization process and watch how heavy it became. Not necessarily solving every problem, creating as many burdens in some ways as it solved. It was extremely informative and translating some of that energy into a real user-specific product design focus like at Doximity, an online doctor platform, gave me a lot of empathy and respect for the challenges that come with adopting new technology. Sometimes people think the healthcare system is resistant to new tech and I don't think that's the case at all, but rather doctors aren't anti-tech, they just don't have enough time as it is. And there are so many things out there in the news every day, new pieces of technology. Look at OpenAI. We've gone from a 15-month new model release cycle to a six-week new model release cycle. That's a lot for anyone to keep up with, no matter how big of an AI enthusiast you are, and that's just one sliver of technology. So, when we built tools like Doximity Dialer, it was with a relentless zealousness of religion of sorts to make sure that everything that we were doing was deploying a new piece of technology, in this case mobile and mobile telephony, that would protect a doctor's phone number for instance, to a particular problem that the existing healthcare incumbency wasn't solving for individuals. And that's what led to trust and adoption and all the sorts of things that you try to quantify and look for when building a medical product or any product. But in terms of what would make this particular wave different and not necessarily have the same stumbling blocks as some of the previous technology waves, I think there's a few things. One, AI itself is increasingly able to crosswalk many of the problems that exist in healthcare today. Healthcare itself speaks a moon language, and it's an intimidating place to navigate to try to solve problems unless you have decades of experience in it yourself. But what are large language models if not that second L, language, and being able to quickly crosswalk and navigate that? Particularly when you move towards things like generalized agents that have deep access to all of the different silos of information that historically exist in healthcare today. But two, and I think this is what inspires me to come into work each day the most, it's builders. If you look at how you would build just a few years ago, you were gatekept by who had that engineering experience. How do people build, why do people build healthcare startups? I spent a lot of time supporting healthcare entrepreneurs through Rock Health. It's typically born out of pain. People have a tremendously painful experience, whether that's high friction in their day job or a loved one who got cancer, and then they have to go through this filter of essentially luck: were they lucky enough to have that engineering background or whatever it would take to be able to take action on that pain that they're feeling? So much of the problems to be solved in healthcare historically have been that the people identifying the problem are completely distinct from either a skill empowerment basis or a permissions basis in solving those problems. And with AI, for the first time, anyone can become a builder. With tools like Codex or some of these newer models today, you have the ability to prototype and the ability to, sure maybe not deploy it to the patients in the hospital, but get through that first wave of what is possible, what would solve your needs, what would help you learn better, what would help this patient navigate a system. And that is going to do so much for grassroots solving your own problem levels of creativity that can then be doubled down on, whether that's through the venture community or through the IT department, that we can start to affect much more change over a time period that is unfathomably faster than has ever existed before.

Sure. So ChatGPT is a popular app. It's not the only way that we reach users. For instance, we have an API that powers a large number of the healthcare startups out there. Our models and our mission are achieved through our partners as much as it is through our own products. But our main product was very big when it launched and has only continued to grow. There are now over 900 million people per week who are using ChatGPT. When you look at how much of that is health, it's a decent number. Around a quarter of those weekly active users will be making at least one health-related query. So it's 230 million, I think was the last public number that we put out there, people per week are asking a health-related query. It's about 40 million people per day. So to your point, that's a responsibility as much as it is an opportunity. But I think in order to meet that responsibility, we have to understand what's different, not only in how it's being used, but how it might differ and why it's being used versus the existing tools that are available, whether that's a search engine or a local urgent care or talking to a friend or talking to a stranger on a Facebook group. Some of that is obvious. For instance, AI is conversational and healthcare topics are intimate and it makes more sense sometimes to be able to have that dialogue compared to a single search. But also single searches, if you look at the search engine generation of use, it puts a lot of work on the patient or on the searcher. You might nail those seven keywords that you're putting in, but now you have to go navigate the 10 results and you pretty quickly end up on the bad side of Las Vegas. There are banner ads over here and there are vague ads over here and you have to be tremendously careful and take on a tremendous amount of effort to get to the answer. These are skills that we want people to still have in the age of AI. I think scrutiny, skepticism, logic are all going to be more important in the age of AI rather than less. But AI, when it's the middle of the night and you're looking for a particular question, you want to double-check your thinking or know if this is something you should be concerned about, it's a little easier for that kind of task than has previously existed. The other thing to compare it to is not just apples to apples digital. It's not just chat versus search. It's also chat versus the existing healthcare system. People today often ask questions like, is ChatGPT the new front door to healthcare? I don't think there is a single front door to healthcare. I think people have to enter the healthcare system today through many different paths. Sometimes that's asking a family member about symptoms they've been having a long time. Sometimes that's ending up in the emergency department when they should have gone in earlier through a different pathway. Sometimes that's the one primary care visit they have per year and then the other 364 days they're on their own. A lot of our work in training our models and building our products are to address those kind of gaps. For instance, in those 364-day gaps, there could be how do I prepare for my next visit? It could be post-visit, what was that big Latin word that the doctor said? What is a low sodium diet? How does that reconcile with my interest in fishing? Really personalized following of care plans. Other times it's having increased context. You've probably learned over time as you've used ChatGPT that it gets better the more that it knows about you. Imagine having to interface with someone whose only exposure to you are those seven words that you put into a search engine. Search engines have amnesia, or even an untrained version of ChatGPT where it isn't grounded in your own circumstances. A lot of the investment that we're making today in ChatGPT is to let people who want to ground it in their own context securely. So syncing their medical records, syncing their biosensor and wearable data, being able to treat uploads securely, safely, encrypted separately, only called in in the context that the patient wants. And then being able to leverage those in the conversation because of course the accuracy that the models can provide will be so different and so much more personalized if it knows that you're 20 and running a marathon or you're 70 and have a particular chronic condition or two that you're managing.

It's a great question. I think the early use case that we have of AI probably signifies where the need exists and where we'll see more. For instance, when we look at the 40 million people per day that are asking health-related questions to ChatGPT, that is disproportionately happening after clinic hours. That is disproportionately happening in rural areas, the farther that you would have to drive to a healthcare institution. So they really are aligning with where the friction points currently exist. On the professional side, I think physicians very rightfully were skeptical of AI and its earliest iterations, and that's because AI needs to earn trust with them but also the early iterations were interesting but they weren't useful. They hadn't quite reached a level of expertise that could contribute in every single scenario. They may have still had too high of a percentage of hallucinations or they may not have been able to actually do meaningful work besides just answering questions. But over the past two years, we've seen physician use of AI essentially double year-on-year and has quickly become an extremely popular use case to the point where I think a year or two from now, the questions are going to be, is it problematic if you're not using AI at some point in your care to make sure that it's meeting the standard of care? So I will say, it's a trend. Will everybody ever have access to AI? We certainly hope from a mission standpoint. Our mission at OpenAI is to make sure that AI and all of its benefits affect all of humanity. That's both direct and indirect. In order to do that, you have to make not just your most capable models like GPT-5 Pro that can run for half an hour or goal mode and Codex that can run for a full day. Those are a little expensive. But make sure that the free models, the models that everyone has access to, are performant for the needs of those users too. That's constantly a balance that we're working on in our design process, but I think it's probably one of the most important things to be focused on as a research lab over the next year.

Well, it's certainly true that the AI technologies available today are non-deterministic. And that creates a new set of challenges to solve as a lab that is working on making AI performant in all of these healthcare scenarios, but also not just performant, but safe in those rare scenarios and those one-off events and those potential end of the histogram type moments. There is a difference though. I think responsibility and reliability are two different things. If we compare to the status quo today, which is a world-class Stanford trained doctor who's been through fellowship and has two board certifications, they're non-deterministic. If you're catching them after a 16-hour shift and you're only able to talk with them for 30 seconds after paging them when they're in the ICU, that might not be the same as if they're at their best, well-caffeinated moment while they have an hour to talk to you. So sometimes the comparison bars that we give aren't always apples to apples, but there are clear distinctions between what you get with LLMs and some of these older algorithms that you can go through and measure and verify every step of the way. But I think medicine for a long time is probably the profession that is most grounded in proxies for trust than any other industry in the world. I think every industry is going to need to start thinking through proxies for trust. Just look at the past decade or so with fake news and the price of generating unlimited content going to zero. Everyone is going to have to come up with their own frameworks to evaluate information that they're seeing and decide what is reliable and trustworthy. And I think medicine has been doing that to varying degrees of success for a very long time. You had standardizations of the medical education system a century ago. You had brands that correlate with trust like the New England Journal of Medicine or JAMA or groups like that. You have board certifications. You have different rankings and proxies that people give to the steps of their medical education. You have citation count and H-index and things like that. So we're all able to look at a piece of information in medicine and decide how much additional scrutiny versus immediate trust and movement is warranted. And I think the question is, will that be extensible to non-deterministic things like AI? Many of those early steps are already being built. For instance, citations. Rather than just take the word of the AI, it would be much better if the answers that you're seeing from the AI are full of citations to the latest guidelines or the medical literature or to your own institutional care pathways. That's for example one of the things that we build into ChatGPT when it's deployed in a hospital like it is here at Stanford Children's. In other cases, it's making sure that you have physician-level reliability on the model outputs to another doctor but also the model outputs to a patient. For instance, all of our model development steps go through review by over 250 doctors across all of the different specialties and care settings that you might expect. Looking at not only optimizing for bread-and-butter scenarios, but the toughest possible clinical scenarios and red teaming, like deliberately trying to break the experience and make sure that it'll still work even in those one in a million scenarios. That allows us to not just look at did the AI get the diagnosis right or whatever you might be trying to measure in a more basic like how did it fare on an exam type score, but instead apply rubric criteria across all of the different things that a doctor, if they had unlimited time, would ideally want to be checking. For instance, one of the big themes right now is escalation. Being able to make sure that a patient in a moment of need is being routed appropriately back into the care system, whether that's emergent care or guideline recommended care. Another would be adaptive literacy, making sure that you can explain a particular concept at the 12th grade literacy level or the third grade literacy level or to an oncologist or to a pharmacist in Kenya and making sure that's reviewed by not just experts in that field but regional expertise that understands what care pathways and resources may be available. And then finally, managing uncertainty. Everyone remembers the hallucination scenarios that can exist in AI and how one in a thousand times there could be an answer that is as confident as an overconfident intern with no ability to differentiate from the other 999 answers that were given in terms of confidence. That's something that actually can be worked on. Not only grounding your answers in citations helps with that, but making sure that the model, particularly in scenarios with patients, is able to recognize that not all the information might be there. It's much better to ask follow-up questions or recommend a next set of steps that might be run or information that could be gathered rather than immediately jumping to a particular question. Just like all of that was trained into many of the folks in this room through all the hard work of their education, many of those same characteristics can be distilled with physician guidance and professional guidance into these models to make sure that they're a good teammate for doctors on the wards or for patients when they're in those 364 days a year that they're not in the system.

Well, I would say the spirit of our work again is not to intercept care whatsoever. It's to make sure that we can support the existing care that's being delivered by the healthcare system which is spread very thin today. If there's one tight commodity in healthcare, it's time. I think the trickiest part about the word adherence, not to nitpick on words, but it sometimes puts the burden too much on the patient in some ways just in how it's interpreted. It was the patient's reliability or something that somehow caused adherence to fail when in reality, you can have plans that are medically very sound but operationally hard to execute in the realities that that patient is experiencing, which could be for a variety of different factors ranging from just how hard it is to stick to a particular care plan or medication side effect profile to how easy it is to apply a care plan to your own personal context to many social determinants of health factors. So I think if we do our job right, the way that I describe it to the team internally sometimes is imagine if you had two patients that were going through a tough diagnosis or juggling multiple chronic conditions or whatever it might be. And one patient had to navigate that system by themselves, which is really hard. And then the other patient, they have a spouse and the spouse used to work in healthcare and is neurotic as heck and carries a clipboard and comes to every visit and takes notes and is with you and when you get home you're not watching TV, you're going to go to the pharmacy and pick that up. You're going to eat that healthy dinner. You're going to go on the walk. I think there are elements of that, some of the better elements, that actually could be brought into AI as it becomes more of a super assistant over time for people. That means helping people ground care plans in their own personal context. We touched earlier on the classic patient handout example of making sure that if you have a diabetic foot ulcer and your hobby is fishing, did you have a chance to have that conversation on if you can go fishing with that condition and how should you care for that? What does that mean? To getting rid of all of the micro frictions in life that lead to adherence drop-offs. Whether that's affordability and navigating a really complex payment and insurance system to decide where the best way to affordably get access to a medication or treatment is, to adjusting your calendar to make sure that you can swing by the pharmacy to pick something up on the way home, to being able to stay in communication every day with a modality that isn't like filling out a multiple-choice survey or a snail mail how was your visit type form, but instead is adaptive to how you want to communicate on your own pace the way you might with your family members that can find ways to take something that you might be resistant to. Maybe that's a colonoscopy. Colonoscopy is something where in many states half of patients never get one. If you dig into the underlying factors, many of those factors actually could be solved with more of a conversation. But the system is not equipped to necessarily have that conversation when visits are only 7 minutes long. Being able to think through where there may be resistance, how to get moving, how to reflect your own personal circumstances. In order to do this, we have to build the architecture for memory and personal context and model reliability and overarching infrastructure and connectivity and tool calling and connections into hundreds of different systems that folks in this room are building. I think that's what's going to make the next year or so the most interesting. We can move from chat as a question response, question response, to a true life empowerment tool that ideally can play these roles, or at least for those who want to, can be an equalizing factor for many where they will then be able to have, healthcare is still going to be hard, but they'll then be able to have a fair shot at executing their own care plans with essentially a really smart virtual care team behind you.

I would say the most interesting thing folks could do right now is try to go really deep with an AI tool. I certainly hope it's ChatGPT and Codex, but it doesn't have to be. There are other good ones out there. But I think there's a difference between trying five or six models and saying, oh Google's a little better for this and Claude's a little better for this and ChatGPT's a little better for this. That's all still kind of superficial at the end of the day. Now is the right moment to see if you can actually get the AI to build a dashboard for you end to end of something that you're tracking, or to see if you can have an agent run automatically for you once or twice a day to do something that you're interested in taking action on in your own life, or to build some sort of multi-agent system that can evaluate or complete a task from multiple different perspectives at once. You have the ability today to download tools like Codex that can not only take action on those things immediately, but they'll put you in the 0.1% of AI users that will have the perspective, it's certainly very early, perspective into where this is headed next and where there might be really great opportunities to partner or build.

Thank you.