How can medical AI help doctors make better decisions while preserving clinical judgment and human connection?

Travis Zack, MD, PhD, an oncologist, researcher, and chief medical officer of OpenEvidence, joins The Future of Medicine to discuss the future of AI in clinical care. OpenEvidence is an AI-powered platform that helps clinicians find and understand information from peer-reviewed medical research, clinical guidelines, and other trusted sources.

In this episode, Zack explains why reliable medical AI depends not only on avoiding fabricated information, but also on finding the right evidence for each clinical question. He discusses how OpenEvidence is designed to bring relevant research into physicians’ workflows while leaving the final decision in the clinician’s hands.

The conversation also looks ahead to more proactive forms of care. Zack describes a future in which AI could identify meaningful changes in a patient’s health or in the medical literature and alert physicians before the patient’s next scheduled visit. 

Zack also shares his research using language from medical records to improve cancer care. By analyzing how patients describe symptoms such as chemotherapy-related nerve damage, researchers may be able to better predict when treatment should be adjusted to prevent lasting side effects.

Together, Dr. Zack and host Errol Ozdalga, MD, explore whether reliance on AI could weaken physicians’ skills, how medical education may change, and why human reasoning and empathy will remain essential to medicine. They also consider a central question for health systems: Will AI be used only to increase efficiency, or will some of the time it saves be returned to patients?

The Intro

“We collect probably hundreds of millions of hemoglobins from patients. Why are we not using something that's much more complex?”

Dr. Travis Zack is chief medical officer of OpenEvidence, the fastest growing application for physicians in history. 

“If we're saving any time on operational tasks, we need to at least reserve some of that back for what medicine is supposed to be about. Otherwise, we'll actually deliver worse care to more people.”

We explore his journey from studying the biomechanics of mantis shrimp as an undergrad at UC Berkeley to becoming an oncologist, an AI researcher at UC San Francisco whose work uses artificial intelligence to unlock clinical insights hidden in the medical record. 

“I want to make sure everyone's getting standard of care. And then I want to make sure that every physician has the tools they need to really feel like that's where the human really comes in.”

In this conversation, we discuss how AI is reshaping how doctors access and apply medical knowledge and how the future of medicine as we know it will change. 

Welcome to Stanford Department of Medicine's inside look at the future of medicine.

The Transcript

Dr. Errol Ozdalga: Travis, thank you for being with us today. Tell me more about yourself. I had a lot of fun learning about you. And one of the things you'll know is I learned that you started off with a dual major at UC Berkeley studying the mantis shrimp. But before you answer the question — you brought the shrimp from Hawaii, where you're from. Were you booking them in a bag and checking them on a plane from JetBlue?

Dr. Travis Zack: That is a long time ago. I actually did not start in science. I thought I was going to be a philosophy major. I took a physics class as a breadth requirement. It was the hardest class I'd ever taken — physics for scientists and engineers. It was very difficult. And I decided, well, I actually really enjoy doing very difficult things. I like the sense of accomplishment. I'm like, this is the hardest thing I've ever done, I'm going to be a physics major. I ended up — I was always a naturalist interested in conservation. So I took a lot of organismal biology classes for the field trips and ended up double majoring. I did research in an imaging lab while I was at Berkeley, and then also got into research in organismal biology, in biomechanics. That's how I ended up working on mantis shrimp. I did have a spring break project where I was in Hawaii — I'd done a lot of diving there, so I was a pretty good swimmer. So over spring break, we went specimen collecting, not in Hawaii, but in Catalina. I was paid to go free diving for shrimp and lobsters. That was fun. So I had a fun undergraduate career, wasn't doing anything medicine related. I didn't come from a family of physicians or scientists, and didn't really recognize that there were doctors who did research. I know that sounds silly, but if you're growing up on a small island, it's not something you think about — your doctor is the one who fixes your knee, not necessarily someone who does research. When I was exposed to that during graduate school, I did a turn — both from a research perspective, then went back to medical school, and used a lot of the quantitative skills I'd learned during physics to answer questions in those fields.

Dr. Ozdalga: So you did the dual major undergrad, then a biophysics PhD, but you were doing computational genetics. 

Dr. Zack: Biophysics typically covers a couple of different things — imaging is one big part of it, then crystallography, then neuroscience. I was very much on the imaging route. I thought I was going to do imaging, and the first couple of rotations I did were in that realm. Toward the end of my last rotation, my mom got breast cancer, and I went through a phase where I felt like any papers derived from what I was doing would be very esoteric, very small, and wouldn't answer the big questions I was really excited about. So I had a change of heart — I wanted to do something involving cancer research. I asked to join a lab — I actually just got to meet that person again — and did a rotation in a cancer research lab, very much cell culture work. I had never touched a pipette before, didn't know anything about what I was doing, probably contaminated all sorts of things. But they were very nice, and it was my first exposure, and I decided I really wanted to solve this problem: why does cancer happen?It was a good place to be, because at that time, in the 2010s, the Broad Institute at Harvard was doing a lot of sequencing with large datasets that had a lot of noise, because of all the passenger events that occur in cancer. I got to work with people designing algorithms to identify what's causing cancer — what the driver events are. Got into that, and then, it's a slippery slope, and I applied back to medical school.

Dr. Ozdalga: Even at that early point, you were working with a lot of data, computationally — but that's not really in the AI realm yet, right? The AI comes in more in medical school. How did you transition into the direction you're in now?

Dr. Zack: For most of those genomic projects back then, it was mostly statistics — fancy statistics, biostatistics — how do we generate p-values from large datasets. That's most of what I did. In medical school, I started to get interested in machine learning because of exposure to what clinicians and data scientists were seeing. I was a third-year student, and we were on rounds, and the moment for me where I felt like we had to do something different was — we were arguing on rounds about whether a 0.5 drop in hemoglobin in an inpatient was relevant, and I felt like we only had two data points here, and we weren't using any of the information we really needed to determine if this was relevant. We collect hundreds of millions of hemoglobin values from patients throughout an entire hospital's history — why aren't we using something much more complex to help guide that decision-making? That led into the realm of machine learning — how do you take data points from disparate sets and come up with the right answer, the right prediction? I got interested and taught myself a lot of the machine learning models. Again, a great place to be. I'll say at that time, in the HMS system, it wasn't very easy to get your hands on real-world data — it was very hard to do as a researcher, despite being in the genomics realm. What I was very interested in was how we use real-world outcomes paired with genomics to create predictive models, and it was very hard to do that at MGH and Brigham. Stanford and UCSF were doing a very good job of that — with the people who were really visionaries in that realm. So I sent an email when I was applying to residency saying I'm really excited about what you're doing, and I came out here, both for the training and for the ability to work on these large datasets to develop machine learning models.

Dr. Ozdalga: You come here, do your residency and fellowship in oncology, now you have a research lab, and of course now you're chief medical officer for a very important company, OpenEvidence, that all of us are using. I love AI, I use it non-stop. I was one of the first people to sign up for OpenEvidence, but then I hadn't been using it until more recently, preparing for this — and I got back into it over the last couple of weeks. It is amazing, absolutely amazing, and I promise you, moving forward, it's going to be one thing I use every single day. The ability to ask questions — I want to chat with you more about the deep consult feature. I tried different questions, compared them side by side, and there's a lot going on there. The most wonderful thing I want to hear from you today is: where do you think we're going with all this? What's the future of medicine in relation to all this?Congratulations — you had 100 million consults just last year, Americans alone.

Dr. Zack: I often think, as a clinician, that a lot of burnout comes not from the paperwork, which is real, but from the angst about decision-making — did you make the right decision at the end of the day, what are the things you could have done that you didn't do. That's exhausting. I think it's one of the privileges but also the onus of being a physician. Now I go to bed thinking about the fact that there are 950,000 questions being asked by doctors every day that affect patient care, and did we do everything we could today to make sure every single one of those was as good as we could make it. Still a lot of angst. So it's very exciting, but it also feels like a lot of pressure to make sure we continue to be as good as we can be. I'll push back on one thing: I'm chief medical officer, but I'm very lucky — I kind of fell into this, right place, right time. I was doing machine learning work in fellowship, on real-world data. I had a separate project, for medical education, using pre-GPT tools — I was interested in how we find the right case, so that when I'm a resident and teaching, I can teach based on a case related to a patient we're seeing. It was a project using old NLP tools, a total side project out of passion. Through that, I got connected to the group at MIT doing what eventually became OpenEvidence, because there weren't very many clinicians doing NLP at that time. Zack Ziegler and Daniel Nadler were the co-founders. I basically luckily fell into all this, and feel so privileged to have been able to help build it. But I want to give credit where it's due.

Dr. Ozdalga: That makes it sound like you were at the right place at the right time, but you were also the right person.

Dr. Zack: I did see that natural language processing in 2018, 2019, 2020 was super frustrating with real-world data, but you could see things were starting to change, and there was potential if you were in the field. So it was exciting to be there at the right time.

Dr. Ozdalga: You gave a talk earlier this year — there were multiple "aha" moments for me around the amount of power in the different data and how you can use it. You talked about how, even as a generalist before becoming an oncologist, you thought there was an algorithm — patient gets the algorithm, goes through it, done. But life is life. They get symptoms, they have issues. No one actually matches exactly what the studies tell you they're supposed to get. Tell me more about how OpenEvidence, how AI, can address that.

Dr. Zack: The way it addresses it now is by helping physicians identify where other people have tried to find a solution, piecing together breadcrumbs within the evidence. I'm hoping that once we identify the closest thing, we can extrapolate from that. To your point — in oncology, for example — here's how far the evidence can take you, and here's what's missing based on where you're at; here's where you're going to have to make a decision. I'm really excited for the future, because now we have a strong understanding, at the AI-system level, of where the limitations are. How do we push the people who have the power to actually fill those gaps? That can mean different things. The simple one is: we'll run the trial, or we found what could be a new drug interaction based on how often people are asking about a side effect — we should go study this prospectively. There's another gap: things that will never get a clinical trial. How do we find the right expert to answer that question? Maybe it's not evidence-based, but it's better than what we have now, which is nothing. And how do we do that responsibly, so it's very clear this is expert opinion or consensus, as opposed to an evidence-backed decision — but it's still better for a community oncologist to have that versus nothing. Because we have such a good picture of where questions show up and where the evidence is lacking, we can be very targeted about trying to fill those gaps. We're working on that this year — how do we identify where those gaps are and then find the right people to fill them.

Dr. Ozdalga: Your own research is about taking the medical record — where a potassium value is a number, but you take the text and provide meaning, real proxy data out of the text. Tell me more about that.

Dr. Zack: I'll tell you about a project I'm really excited about, working with a couple of graduate students at Berkeley, around chemotherapy management. Despite everything we've done with targeted therapy and immunotherapy in cancer over the last 20 years, most patients still get some form of traditional chemotherapy at some point in their cancer journey, at least metastatic patients. Those chemotherapies have a lot of toxicities that are poorly understood in terms of how to manage. The art of medicine in oncology is largely a balance between efficacy and toxicity, where both are unknowns when you make the decision. You don't know whether dropping the drug or decreasing the dose affects efficacy, because the clinical trial wasn't run that way — nobody ran a 30%-dose or 70%-dose arm. So we can't tell the patient it won't be 100% as effective; it's not a linear relationship. Similarly, with toxicity, we don't know how much of a dose reduction is needed to avoid it. And that includes both short-term or catastrophic toxicities, like hospitalizations, but also things like peripheral neuropathy that can last a lifetime and actually be more debilitating than the cancer itself in the end. So rather than what we used to do for predictive models — say, "this person had neuropathy, let's train on neuropathy and see if we can predict who will get it" — we now look at what symptoms the patient was actually describing, in natural language, during each cycle of chemotherapy, and how that progressed. We train a model so that when a patient describes symptoms moving from their toes to their feet, and we drop the cycle and decrease the dose, we can avoid toxicity that persists past six months, which is really debilitating. So we use a combination of the language the patient describes to the physician, along with the traditional features, to try to find a better predictive model to balance efficacy and toxicity. Nothing to do with OpenEvidence, but that's the kind of project I'm trying to work on in the lab.

Dr. Ozdalga: Do you foresee — bouncing back a bit to OpenEvidence — that as it develops relationships with health systems, integrated with the medical record so you don't have to copy and paste, will it also take that evidence-based, PubMed-sourced data to the next level and say, your patient is starting to have increasing neuropathy, make these changes?

Dr. Zack: I'll talk about a middle step first, which I'm also excited about and which we're already doing, before getting to that final, proactive step. The middle step is making sure that when a physician makes a decision, they understand the evidence behind it — not making decisions proactively, but when a clinician has made a decision, putting the evidence front and center in that process. We've done that with our AI scribe. I know there are a lot of AI scribes, but we built ours with a specific goal: when you write your note, with an assessment and plan with bullet points, for every single one of those bullet points, we ask OpenEvidence what it thinks about that decision based on the patient's context. We're not going to countermand the decision, but we'll say, here's what we found the evidence supports. We put it in the note as a small chip — a one-sentence summary you can click to read the full research behind, as if we'd asked twenty questions for you. When you sign the note, it all goes away — nobody wants that in their documentation — but it's a way of bringing the evidence and reasoning directly into the workflow, so we know why we're doing what we're doing, and if something's changed, we bring it front and center so the physician doesn't have to ask.

Dr. Ozdalga: So with the AI scribe — I'm the patient, you're the physician, we have a regular conversation, and you have the power of OpenEvidence coming in and telling you, based on the evidence, here's what's recommended — and then it gives you your note?

Dr. Zack: We easily could provide recommendations, but that felt a little paternalistic. So what we're trying to do is: you write the note, or say what you're going to do, and then we show you the evidence behind what you did — including something like zinc for a cold, where maybe there isn't a lot of evidence for that decision. That was a product decision we made. The idea was, we're not going to tell someone how to practice, but given that you're practicing this way, here's what the evidence says supports that practice.

Dr. Ozdalga: When you have these relationships with health systems, do they share their data with you?

Dr. Zack: We don't train on patient health information coming from health systems. Even in direct integrations with a health system, data is stored for a very limited period, just so we can write the note. What we're excited about is the next phase — running OpenEvidence within a health system, not externally, so that whenever there's a change in the evidence, or a change in the patient's note or chart, we can proactively surface that and identify what changes we might recommend based on a change in the literature or in the chart. Pairing those together in a way that's very sensitive and very specific is important, because in any patient's documentation there are things that aren't evidence-based, but for good reason — a conscious decision by the physician based on knowledge that may not be present in the record. So we have to be careful not to create alarm fatigue with these recommendations. That's research we're doing with health systems right now, not an active product yet. But that's what I'm excited about for the future, because physicians don't have time to go through every single change in a patient's chart, or in the medical literature, every time it happens, much less see that patient every time something changes. Why should we, in 2026, wait six months to make a change just because that's when we happen to see them? Is there some way we can be more proactive when changes happen, to extend patient health? That's the vision — not a reality yet, but what we're aiming at.

Dr. Ozdalga: One of the beauties of OpenEvidence is trust but verify — the risk of hallucination goes down compared to other LLMs because you're dating back to source reference material, peer-reviewed journals from the best journals that exist. OpenEvidence has an option called Deep Consult, where you give information and it asks you more questions. I have a history of atrial fibrillation from my 20s, and I don't take omega-3s because I've read it can increase AFib risk, even though I've been ablated. I asked OpenEvidence whether people with this history should take it, then went to Deep Consult and got a different, more custom answer — it said stay under this dose and you'll be okay.

Dr. Zack: We're actively improving Deep Consult, turning it into two separate models. We created a single model the first time, which was a conscious decision, but now we can do it better with two: one is a true deep consult about deep patient history, and the other is more like a deep research version. We were trying to solve two problems at once, which wasn't ideal. When we built and trained the original OpenEvidence model, there was a specific form factor that, as succinct as possible, about four paragraphs, so a clinician could get an answer and go about their day. But if a clinician asks what the treatment is for stage three lung cancer, there's no four-paragraph answer for that — the NCCN guideline is sixty pages. It's impossible to summarize sixty pages into four paragraphs. So the point of Deep Consult is to go two ways. One, we throw off those limitations — it's a reasoning model that can take as long as it needs, and we'll email you when it's done. It has no length restriction — if stage three lung cancer takes six pages to answer with fifty references, we'll give you that. The other issue was long context — the original model didn't do a great job if you pasted in four pages of medical records, say from a health system we have a BAA with. It struggled to find the right dozen references among everything going on. So Deep Consult was a way to parse through that in an agentic way. Those are actually two very separate problems that we answered with one model, and very soon we'll have separate models — one for a deep dive into a patient record, to find the breadcrumbs you might miss, and one more for research, to comprehensively identify everything relevant to a specific question, no matter how long it takes.

Dr. Ozdalga: When you get more into the agentic field of Deep Consult, and it feels like deep research, asking you questions and coming back — does that increase the risk of hallucination, since it's not just retrieving data but reasoning through it in more complex ways?

Dr. Zack: There are a lot of different failure modes in language models, and it's important to talk about them. The risk of hallucination doesn't go up with Deep Consult. What there is a risk of is inaccurate wayfinding — what if we ask the wrong questions, ones that don't help us get to the right answer, and instead ask distracting questions? You answer them, we find an answer, but it's the right answer to the wrong question, so to speak. Doing that well depends on the first step — searching the medical literature well — because the follow-up questions are based on the papers we found and how we parse them to find a targeted answer. Hallucinations don't go up, but the wayfinding that guides someone to give us the right information does create more forks in the tree, and making sure it goes down the right path is a challenge. Since we're talking about failure modes — hallucination really isn't a problem for OpenEvidence, given how it's built. What is a problem is making sure those references are the right references. Every day, we have dozens of engineers thinking about how to make sure we start with the right references, because if we don't, it's not a hallucination, it's writing about the wrong material. That's the real failure mode, and it's two pipes: making sure the data is good — that we're getting the right references from journals, societies, the FDA — and that we're creating ML algorithms to find the right references when someone asks a question. 

Dr. Ozdalga: This is why clinicians love it so much — you really can't trust other tools the same way. We've all had LLMs make up articles — not in OpenEvidence, but in others, and that makes a big difference. In preparation for this interview, I polled our house staff and faculty about what they'd like you to hear. The most common theme was: what if, six years from now, the internet goes out — what do we do? Are we deskilling — taking the scalpel away from the surgeon, having someone else do surgery? Is that a concern, and how do we avoid it?

Dr. Zack: I talked about it a bit during my lecture. There's a part of me, as an oncologist, that thinks if I lost access to the EHR, it would be very difficult for me to practice — I don't know how much I'd trust my skills as a "wilderness medicine" oncologist. We're continually dependent on the skills — or the tools — we trained with. I think with active utilization of these tools, just like with active utilization of PubMed or other resources — where to understand why it's the answer and ingesting it a little — when the lights go out, you're actually going to have a better understanding of the primary literature and guidelines than if you'd done the reverse, which is often, "I do that because that's how my attending does it." A lot of medicine has been done this way — we do what we were taught to do, and there's a lot of incestuous learning, because it takes so much effort to go figure out why we're doing what we're doing. If we lower the activation barrier to learning why we're doing what we're doing, and spend that ten , especially as trainees, when the lights go out, you're actually going to have better knowledge about what to do than you would have otherwise. But that active utilization is incredibly important.

Dr. Ozdalga: On the flip side, I learned from your talk today that you're playing a much bigger role than I realized in education — you have the data to see what a hospitalist actually needs. How many times have I taken a maintenance-of-certification exam and asked myself why I needed to know something? You could actually guide that.

Dr. Zack: We were talking to the American Board of Pediatrics yesterday about the responsible way to do this. If we have all this data — forget the individual user — on what's actually relevant to practice at different stages of training, and compare that to core competencies, how do we pair that to create longitudinal assessment, or maintenance of certification, that actually helps people fill deficiencies? We have very strict rules around privacy, so we'd never give raw individual query data. But it's interesting — if you're a PCP in Oklahoma suddenly doing much more OB than you ever have, and at the end of the year we see you're doing a lot of OB and not much pediatrics, for a board exam, do you want to spend time teaching OB to someone who may already know more about it from practice, or teaching pediatrics because they aren't getting it in practice? I don't know if there's a right answer, but it's interesting to think about the fingerprint we could use to guide that — how do we make longitudinal knowledge assessments useful to helping me do my job better, rather than a waste of time I have to click through?

Dr. Ozdalga: This is a general question, since you're in this space so much. Have you heard of the song "Papaoutai"? It's a famous French song, very famous everywhere, normally in French. There was an English version that came out this year and went viral — I loved it, went back to listen to it repeatedly. It's not available for streaming, so I researched it, and the song was made by AI. Somebody had recreated it, made a dance video, a faster version, in English, and I had no idea. I even looked at the video and thought, "this is her" — and didn't realize she was AI-generated. That was an "aha" moment for me, because for the longest time I didn't think AI would infiltrate certain things, and it makes me think about my interest in humanism in medicine, the bedside — I'd always taught that if anything, this would give us more time at the bedside, for humans and connection. But I'm also wondering, will AI actually take that role too? If I have a companion who answers all my questions, anytime, smarter than anybody else, who can help reduce my anxiety if I'm worried about something — isn't AI maybe going to fill the gap of humanism too?

Dr. Zack: Maybe — but I don't feel that way, for a couple of reasons, based on my experience with patients and what they choose to make decisions based on. I see it as a foreshadowing of what they'll want to help them decide in the future. A lot of patients come in with all these resources for information, and they know the right resources — the ones likely to give the right answer — but the one that really helps them decide is the uncle who went through it, a family member or friend they trust. That can be the bane of a physician's existence, and it's not about whether the answer is right — it's about where the source is coming from and why they're choosing it over other, possibly more accurate, sources. I think they're choosing it because there's a true element of empathy — a shared experience, someone who went through something, and the patient can see themselves in that trajectory. I think there will always be that same element for human physicians. A patient can get all the information they want from an AI, but what they really want to know is what a physician who's been there, who's seen twenty other patients just like them, would do in their situation. It's very hard to believe an AI model really knows what it's saying when it says, "this is what I would do if it were my mother" — which is often what patients ask. So I think there's that human element. Beyond that, clinicians have a much harder job than they often give themselves credit for, both from a humanism perspective and a decision-making one. A lot of the decisions physicians make aren't driven by evidence, for good reason — they're driven by the patient in the room, and it will be very hard for an ML model to find the data to train on to make those kinds of decisions. That's one person's perspective, and just because I work at OpenEvidence, which has had 150 million questions asked by physicians, doesn't mean I'm in a better position to speak to that than any given physician. But that's my perspective.

Dr. Ozdalga: It means a lot, and your perspective is very valuable. Our house staff love OpenEvidence. What advice would you give a medical student, a resident, early faculty — how should they make sure they stay relevant and have the right skills for the next ten-plus years of their career?

Dr. Zack: Let's talk about clinicians specifically. I think a lot of our identity over the last fifty-plus years as clinicians has revolved around this really inhuman task: when you enter medical school, your job is to be Encyclopedia Britannica for everything related to medicine — know it all, and then know even more for your specialty and subspecialty. If you think about the pace of knowledge generation, that's an inhuman task, but it's baked into our identity that this is what it means to be a good doctor. There's another part of our identity — the reasoning, a much more human task — how do I take that encyclopedic knowledge and apply it to the patient in front of me. I'm actually excited that, in the next five years, there'll be a lot less emphasis on the first part in training, recognizing that I don't have to be Encyclopedia Britannica. What I do have to be is really good at reasoning, which involves fundamentally understanding the system I'm working in — basic pathophysiology, the hospital system, patient interactions, all of it — and applying the knowledge I can access within that system. So I'm hoping there will be a lot more focus, in training and practice, on that second part, which I think humans are much better at than being encyclopedias. From a career perspective, it's really hard to determine what that trajectory will look like. My hope is there will be a lot less of the mundane operational tasks we've been burdened with for the last twenty years, and we'll get to spend more time on the harder things in medicine. My concern is it'll just mean we see more patients, and I think it's on us to fight that to some degree — if we're saving time on operational tasks, we need to reserve some of that back for what medicine is supposed to be about, because otherwise, I think we'll deliver worse care to more people instead of better care to more people.

Dr. Ozdalga: Perfect segue — between us we have four children, two each, both young. My oldest is eight. If my daughter, or any of our children, were to ask you what they should be when they grow up, or said "I want to be a doctor," what advice would you give them?

Dr. Zack: Those are two different questions. I was raised by two great parents who believed you should grow up to do whatever you love doing, because if you love it, you never have to work a day in your life. I believe that — I've done that, and it's been great for me. So if they asked what they should do, I'd say, "do whatever you love," and that can change — what a four-year-old loves might be different from what an eight- or twelve-year-old loves. As parents, you have the responsibility to make sure they can achieve whatever that is. Most recently, my daughter wanted to be a doctor for marine animals — we took her to the marine mammal exhibit in Marin recently, and I can't wait for her to be a doctor for marine animals. That answers part one. For part two — if you really love understanding the human body, understanding what can go wrong, and helping other people walk through that journey, then you should be a doctor, and I think there will always be a place for that in human society. I don't know what that will look like — whether it becomes a minimum-wage position compared to now, I have no idea — but if you feel fulfilled doing it, I see no reason not to aim for it.

Dr. Ozdalga: You mentioned a concern about seeing more patients. Is there anything else that concerns you with all the changes that are coming?

Dr. Zack: Deskilling is a concern for me. I think physicians were already being pushed more and more toward being technicians rather than the full scope of what medicine is, and if that push for volume and efficiency continues, and we see AI only as a way to improve efficiency rather than quality, we'll lose a lot of what makes medicine fun — and we'll lose quality of care for people who fall outside the mean. When there's a focus on efficiency, the people who live in the gaps — rare diseases, people who don't respond the same way to medication — may get left out, because everyone's focused on what's within the line of evidence-based medicine and efficiency. Those are my two biggest concerns.

Dr. Ozdalga: It sounds like you're working to make sure all patients are captured in one way or another.

Dr. Zack: I think about it constantly — it's a very tough problem, especially in the AI space. What I like to think about is raising the floor: making sure everyone's getting standard of care, and then making sure every physician has the tools to take care of the edge cases, because that's where the human really comes in. We're excited about things like finding the expert — figuring out what the expert would say about an edge case, or, if you're a primary care doctor seeing IgG4 disease for the first time, making sure you have the information you need for the first step, and telling you where the nearest expert is so we can connect you to the right place — because we have such a network of doctors now facilitating edge-case identification and treatment.

Dr. Ozdalga: Last question — you've talked about so many things you're excited about, and your face lights up. What are you most excited about for the future of medicine?

Dr. Zack: Two things, and I think we've touched on both. One is proactive monitoring — a human can't be there every day for every patient. A lot of times a patient leaves a doctor's visit frustrated, feeling like they knew more about what was going on than the doctor did, and the flip side is the doctor only had three minutes before the visit and twelve minutes after. I'm excited about the idea of an AI system thinking about that patient 24/7, so that during the doctor’s three minutes, they have a summary of exactly what flags have come up and where to look to verify that information – so those twelve minutes with the patient leave the patient feeling like the doctor understood everything and already had a plan. The other thing I'm excited about is working with experts who've been practicing for many years, to fill the gaps in knowledge we're identifying, in a way that the moment we fill a gap, 500,000 doctors have access to it. We both trained in very privileged places, and one thing I love reflecting on is that I'm not an expert clinician — I spent 80% of my time in the lab — but I trained somewhere that if I had a clinical question, the expert was a phone call away. What a privilege that is for my patients — not because it's me, but because I happen to know the right person. How do we expand and disseminate that kind of knowledge and access to all the doctors in America, so all their patients have that same privilege, in a way that's responsible with the experts' time? That's where we're looking at the gaps in knowledge we can take to experts — whether that's working with a society to identify gaps and create panels, or finding the expert in an ultra-rare disease and helping them answer the questions a lot of people are asking. That's the other example: creating new knowledge.

Dr. Ozdalga: My favorite line was always, "I may not know the answer, but I know someone who does." Exactly — and thanks to you, now everybody can say, "I may not know the answer, but I know OpenEvidence does." Travis, thank you so much for your time today. That was awesome.

Dr. Zack: Thanks so much.