In this episode of the Future of Medicine, we welcome Dr. Robert Wachter, physician, author, and chair of medicine at the University of California, San Francisco, for a conversation about artificial intelligence, healthcare, and the future of medicine.

Dr. Wachter reflects on how the digitization of healthcare reshaped modern medicine — and why many of the frustrations clinicians experience today, from burnout to endless documentation, emerged during the first wave of the electronic health record revolution.

The conversation explores why large language models like ChatGPT feel fundamentally different from earlier healthcare technologies, and how AI may finally help medicine move beyond the limitations of current electronic health records. Dr. Wachter discusses the promise of AI scribes, clinical copilots, patient-facing tools, and digital assistants that can help clinicians synthesize information, reduce administrative burden, and improve access to expertise.

At the same time, he examines the risks of overtrusting AI systems, including hallucinations, deskilling, automation bias, and the challenge of keeping humans engaged when machines are right most (but not all) of the time.

Together, Dr. Wachter and Euan Ashley discuss empathy, medical education, patient trust, and what happens when technology begins to act increasingly human.

Looking ahead, Dr. Wachter shares why he remains optimistic about AI in medicine — and why the next few years could fundamentally reshape how healthcare is delivered around the world.

Thank you for listening!

The Transcript

Euan Ashley: Well, Bob, welcome to Stanford.

Robert Wachter: It's a great pleasure.

Euan Ashley: It's so great to have you here. I'm so looking forward to our conversation. We have so many things we could talk about, but I'm particularly excited because you have just published a new book.

Robert Wachter: I have Yeah. Spent, uh, the last two years thinking hard about AI and where, where it leaves us and seems like the most interesting topic in the world. So it was really a lot of fun. Yeah.

Euan Ashley: It's something of course, that we at Stanford spend a lot of time thinking about. It's not your first book.

Robert Wachter: It's my fourth sort of trade book, lay oriented book. And I've done a couple of textbooks when there used to be textbooks. Yeah.

Euan Ashley: <laugh>, we used to have those. Didn't.

Robert Wachter: Yeah. I used to have those. But yeah. This is the fourth book I've written for kind of a crossover audience. I hope to hope it's useful and interesting to health professionals and technology people and all that. Yeah. But also for lay people interested in what happens to the world of healthcare with ai. Yeah.

Euan Ashley: I was really interested, I mean, you're, you're one of your previous bestselling books, A digital doctor, uh, kind of in many ways sets up this current book. And I'd love to dive deeply into that. But it's interesting to me that you've written a book about the digital doctor. You've written a book about AI and medicine, but you, you don't describe yourself as a technologist.

Robert Wachter: Well, I have an iPhone and I use it, so I'm pretty proud of that.

Euan Ashley: <laugh>. That's

Robert Wachter: Pretty good. I really do not know what's happening under the hood. Yeah. Any of these technologies. Uh, I got interested in technology because I'm very interested in the health system, what makes it work, and, and because I've been doing a lot of work in patient safety, uh, for 15 or 20 years. Yeah. Uh, and therefore you couldn't work in patient safety without believing that if we could just digitize, life would get better. And then we digitize. And in some ways, life got worse. And that got me very interested in the topic.

Euan Ashley: Well, that's actually the first thing I wanted to ask you about. 'cause, uh, so many interesting and really resonant parts of the book. I've been listening to it on audio books over the last, uh, couple of weeks. And you read the audio book yourself. I do. So it's been great to have you in my, in my head.

Robert Wachter: Well, for nine hours. Really, <laugh>. Yeah.

Euan Ashley: Yeah. You're in little short segments

Robert Wachter: At a time, you

Euan Ashley: Know, an hour at a time. Um, but you made this point that I've have made many times that, so it really resonated with me, which is that some of the problems that we are trying to solve currently in medicine, maybe even many of the problems were actually created by a prior technology revolution, the one you wrote about in the Digital doctor, which is the digitization of our electronic health records. But I'm really excited and quite optimistic as I think you are mm-hmm <affirmative>. That the current technology revolution can actually solve some of those problems that were in a way created by the prior. But can you talk a little bit about

Robert Wachter: That? Yeah. I mean, I, people, this book is pretty optimistic and people sort of, I've had people ask me, is that who just who you are, Uhhuh? And I say, no, I'm fully capable of writing a very pessimistic book. Yeah. Because the digital doctor is quite grumpy. Yeah. I mean, it was, I've been waiting for this technology to enter our world. Yeah. You know, we really, the book really is about us going from being a paper industry Right. To a digital industry. And I think it made all the sense in the world that this was gonna make care better and safer and more convenient and better for doctors and nurses. And then I found, you know, burnout rates skyrocketed. Doctors were, you know, unhappy with their electronic health record. So it was really my effort to understand why we had gone off the rails. And I think I came to understand first of all that, that any general purpose technology, a technology that just changes everything.

Robert Wachter: Yeah. Uh, they almost never meet their, their, their hype, uh, in the early years. And part of it is because version 1.0 is not very good. Yeah. But part of it is that humans tend to just take that technology and just download it into their existing system, their culture, their governance, their workflow, their workflow. Yeah. And lo and behold, it doesn't work. Right. And when you see these technologies really pay off, and I'm talking about the steam engine or the car or the computer, it really takes sometimes a generation of the industry saying, huh, why are we doing it that way? Yeah. Oh, it's because we always did it that way. Right. The unforeseen consequences of the electronic health record included, now doctors could be prompted to enter data that created a better bill. Mm-hmm <affirmative>. Uh, and so Oh, sure. Let's make the doctor document that, you know, they counsel the patient about seat belts Yeah.

Robert Wachter: Or examine nine body, body parts. And lo and behold, you turn the doctor into an expensive grumpy data entry clerk. Yeah. And then a few years later, we, we turned on this thing called a patient portal, and it gave patients all this data. I could see that my magnesium is low and my EKGs abnormal. Yeah. But absolutely no help in interpreting it. Yeah. And the only bit of help we gave them was a little button at the bottom said, click here to send a message to your doctor. And lo and behold, patients being normal, click there to send a message to their doctor, and the doctors have two hours of pajama time. So those were the sort of things that are obvious that they're going to happen. Yeah. But I can tell you in real time, I didn't predict them happening. And I think the bottom line was that the electronic health record became a useful and definitely positive big filing cabinet. Yeah.

Euan Ashley: Well, you have

Robert Wachter: That, but did not deliver on the kind of clinical intelligence. And, you know, it'd be like if you had a baby, you spent all the time feeding and, and changing the baby, and it never smiled at you. We got so little useful information and we spent so much time putting stuff into it. That's where the frustration came from.

Euan Ashley: Yeah, you're right about it. Very elegantly. Um, I think one of the stories that I, I thought was very funny was that you, you'd admitted that at the beginning you'd spoken to an informatics colleague who, uh, you, you'd sort of believed that their job wouldn't be necessary Yes.

Robert Wachter: In a year or two. Right. Actually, a Stanford trained informatics doc who we hired to be our Chief Health Information Officer. Yeah. It was right after we implemented Epic mm-hmm <affirmative>. And I said, you know, I figured there could be a couple of years of tinkering and improvement. And then I said to him, you know, and what are you gonna do after that's done? What is your career plan? What's the stupidest question? They solved this through the world, solved this. Yeah. He's busier than ever today. Yeah.

Euan Ashley: Yeah. Absolutely. So then that's kind of the setup in a way. And I mean, I think we've all spent a lot of time kind of apologizing to our people or Sure. Or, uh, <laugh>, you know, trying to, to to point out the fact that there were some benefits to the electronic record. I mean, you could for the first time actually get x-rays and lab results and some, uh, you know, local, uh, electronic records mm-hmm <affirmative>. That could be there in front of you. You needed like 50 clicks to, to find them, but at least Yeah. But still, at least they were there. Uh, but I think what we were waiting for was, was a technology that could potentially ingest all that multimodal data and potentially provide some partnership, some wisdom. Uh, and, and I think that, you know, somewhere just in the last two or three years as GPT came to the fore, maybe the, the, the origins of, of the potential for that next revolution mm-hmm <affirmative>. Were there. And that's what you spend most of the time in the book talking about. Yeah. And you talk about really accessing GPT from the, from two, 2.2 0.0, I think

Robert Wachter: I maybe 2.5 or something. Yeah. Two when I first tried to Yeah. Even then, I mean, it was in some ways awful. 'cause it did hallucinate a lot Yeah. And all these kind of crazy things that told you to put glue on your pizza to keep the cheese on, you know, the sort of crazy very caricaturable Yeah. Almost cartoonish problems. But nevertheless, it was obvious to me that here was a technology that was more human-like than anything I'd ever seen. Yeah. And if you think about where we went off the rails with electronic health records, I just came to believe that the first stage was gonna be digitizing the record. The second stage was gonna be connect the parts and interoperability. But we had not hit the, the, the really meaningful stages, which are all right, you have all this data. Yeah. How do we analyze it in a way that's useful Yeah.

Robert Wachter: That helps us understand our patients better. And, and then how do you turn those, that understanding into actions. Yeah. And part of the challenge was until GPT and the, like, a lot of the data in the electronic health record is in narrative form. Yeah. It's in your notes. In my notes. Right. There was no mechanism to, to deal with that. And, you know, even advanced technology tools like, you know, like Google or like up to date. Yeah. If you said to it, you know, I've got a 62-year-old woman who comes in with breast cancer, she's got a calcium of 11 and a creatinine F three, and she's got an infiltrator chest. What, what do you think is going on? Yeah. It would like the, you know, the tools would say, are you kidding me? I have no idea what you're talking about. Yeah. These tools, you could talk to it essentially, like you would talk to a consultant. Right. And as a generalist, I spent a lot of my time when on the, on the wards sort of hoping to run into my favorite consultant, because every patient I take care of, I can name five people in the building who know more about each problem than I do. I know about more about all of it than they do. Yeah. So now in my pocket, I have subspecialty level expertise that I can access and talk to. Like, I would talk to a consultant That to me was a remarkable technological advance. Yeah.

Euan Ashley: Yeah. No, I think it, it was the, the human, human human nature of, of I think the responses. Because at the end of the day, and, and you said this in, in your excellent grand rounds talk that we just came from, you know, we all got used to thinking the Google search was kind of the best answer we could get. Yeah. And, and it was pretty good. But at the end of the day, that's a series of links to other information. Correct. The, the interesting thing about the language models is of course, they respond in prose. Yeah. Uh, and

Robert Wachter: Portraits. And not just, it's not just the output being links, but the input. You know, if you put into Google really more than a single transactional send Yeah. I mean, you could put into the Google what is the right dose of Eliquis for, for, for a patient with pe. Right. You could not put in the pa the complexity of a real patient. Right. Yeah. And so that's, there's a risk of it being so human. One of the things I wrote in the book is we tend to trust who's people more than we trust what's Yeah. Machines. Yeah. And this is the first time we've had a what that acted like a who Right. Which creates a risk that we will give it undue trust. And we've gotta be very careful about that.

Euan Ashley: Yeah. This is something I've actually talked to our, our residency program directors about, because obviously a lot of what you and I do is, is think about education mm-hmm <affirmative>. Uh, and really important in terms of our up and coming medical students and interns and, and really all of us as doctors who are in continuous learning. We, we are built, especially when prose is very well punctuated, when words are, are used in a complex manner when, when a concept is well articulated, part of your brain is trained to sort of switch off a little because of trust.

Robert Wachter: You trust it. Yes. Of course.

Euan Ashley: It's almost your brain says, oh, I'm in sort of encyclopedia,

Robert Wachter: This person's articulate. Yeah. Well, I mean, if they have an accent from the UK, you trust them a little bit more

Euan Ashley: <laugh>. Yeah. Especially from Scotland.

Robert Wachter: Especially from Scotland. Exactly. Yeah. And I think we make judgements about the trustworthiness of the thing from a lot of cues. Yeah. And one of it is the humanness of the language and did it understand my human language? Yeah. And we've never had a technology tool that could do that. Yeah. And so there is a risk that we will grant it undue trust. We've gotta be very thoughtful about how do we test this to be sure it actually merits the trust that we're giving it. Yeah.

Euan Ashley: And that certainly that occurred to me the very first clinic when I used the, our, our copilot system mm-hmm <affirmative>. You talk about the ambient or AI scribe. Um, and, and I found exactly that it was so beautiful, like almost sort of my brain nodded off slightly as I was proofreading. Um, and then I realized that, well, two things. One, it picked up something that I had forgotten from the conversation. So I was like, oh, that's, that's great. Yeah. I was writing this note, I'd have forgotten to include that. Yeah. But the second thing was it, it really did mishear or misinterpret something and put the patient on a drug that the patient was not on. Yeah. And I did catch it, but it was, I was three words past that. Yeah. Before I realized I'd read something that wasn't true.

Robert Wachter: And that's the first time. Think about the 50th time when you've learned to trust the first 49. Yeah. I mean, humans are quite bad at this. Yeah. We, we are not built to be, to remain hypervigilant when we've come to trust a technology that has been Right. The last a hundred times. Yeah. And so, uh, that's a really daunting problem. And one of the ways I framed in the book and in the grand rounds was if the technology was right half the time, it would be worthless. Yeah. It was Right. A hundred percent of the time. That's great. I'm not sure what we're all doing for a living, but Yeah. <laugh>, but the predicament we're gonna find ourselves in for the foreseeable future is it's Right. Often enough to be useful, useful and wrong, often enough that the human has to look over its shoulder. Yeah. That's a task we were not built for.

Euan Ashley: No. Do you have ideas as to how we could program ourselves more?

Robert Wachter: Almost, almost none. Yeah. <laugh>.

Robert Wachter: 'cause I, I really do think it's very much hardwired into the human experience. Yeah. I mean, it may be, um, I mean, there are some ideas. One is this is particularly true for training and preventing de-skilling that maybe we don't ask the tool to weigh in until the human has put its nickel down. Okay. In some ways, analogous to the way we teach young attendings, you know, do not tell the residents and the students what you think. Tell 'em what, ask what they think before you weigh in. Yeah. Um, the analogy to the, to the TSA airport security is they periodically throw in a picture of a bomb Yeah. Or a gun Yeah. To make sure that the agents are staying awake. Should we throw in a the wrong diagnosis or that wrong medicine Yeah. Into the note. And then we gotta figure out a way of being sure it doesn't reach a patient. Yeah. Um, I think some of this is gonna be more sociological than, than technological. It may be just making sure that our trainees understand the risk of, of being asleep at the switch. Yeah. But I'm guessing in the long, uh, over the long game, it's probably gonna be more another AI looking over the shoulder Yeah. Of, of AI one. Right. Than trusting that a human will remain vigilant when they become overly trusting of the technology. But it's a very hard problem. Yeah.

Euan Ashley: It is interesting that, and you described this well in the, in the book as well, that the evolution we've had from the original GPT two to three, where there was hallucination and the, the, the models were mostly using X token prediction. Mm-hmm <affirmative>. We went for the most part going to the, uh, the source of the knowledge and information. But, you know, when we got into the, the rag mode, which is the retrieve, augmented, uh, generation mode where it was parsing kind of real data, going to read, read the, uh, articles if you like, or Yeah. Or read the, the, the digital textbook and then bring back the parsed version of something for real, and, and then provide the link to that. I thought that was an, a really interesting moment because I think everything got better, first of all. Yeah. There was less confabulation. You could also then click through to the link to check yourself

Robert Wachter: To check check the source. Yeah. Yeah. It's interesting because, I mean, the advance there, for example, in healthcare is what do you get when you put in a clinical case into Gemini or into Claude, or into GBT versus what do you get when you put into open evidence? Yeah. A tool that was built for medicine and uses as its main data source, the medical literature and guidelines from the American College of Cardiology, et cetera. Yeah. And the answer is, I think a more trustworthy source using that, you know, uh, uh, called down literature. Yeah. Um, what is surprising though, I have to say is GPT and Gemini are actually pretty good. Yeah. And they are using the entire internet, right. Including PubMed, but also including the Onion, <laugh> and Reddit and everything. And it is, it, it's a little bit surprising how good they are. They're, you know, I, I tend to trust the tools built for this purpose better than I trust the general purpose tools, but the general purpose tools aren't bad.

Robert Wachter: I think one of the things that is fascinating and we don't emphasize enough, we spent a lot of time thinking about these tools as they come into medicine and health systems and doctors using them. One of the things I came away understanding better after writing the book is we have to think about patient facing tools in a very different way. Yeah. That the, the cognitive act of taking a data set of might be 50 items from a patient symptoms, past history, medications, and culling it down to the thing that you put into the prompt box, or the thing that the resident presents to his or her attending Yeah. Is really complex and requires expert level knowledge. Patients have no ability to do that. Yeah. So a patient doesn't know of the 50 fa, you know, I woke up this morning, I have a headache and my foot hurts and I have a fever and I have a past history of this, and I'm on this medication, patient has no idea which ones are relevant.

Robert Wachter: Right. So having them use, and then on the output side, if it gives you an output and I look at it, or you look at it and say, you know, number one and two diagnoses, that makes sense. And I hadn't even thought of number one, but number four, that's ridiculous. I'm gonna pay no attention. Patients also have no ability to do that. Yeah. So the tools for experts and the tools for non-experts really need to be quite different. And I think that right now you have millions and millions of patients using generic tools like GPT or Gemini putting in their symptoms, trusting the results. Sometimes it's great, sometimes it's a little hazardous. I think we are gonna have to develop better tools for patients that probably are gonna act more doctor. Yeah. Meaning the patient says, I have a headache, and then the tool says, tell me about it. How long has it last? Do your eyes hurt in the bright light? Do your, does your neck hurt back and forth the way a physician would as opposed to just a generic chat bot where the patient has to be trusted to put in all the right information in the tool. I don't think that's quite ready for prime time.

Euan Ashley: Yeah. Yeah. And, and I think, I mean, that's, if we think about the process of medical education, that that's maybe the first thing we teach our students when they

Robert Wachter: Start. Absolutely. I watch my daughter go through from being a, a novice to an expert. Yeah. And it's very subtle, but really in some ways what you're learning is how do I take this massive, uh, you know, number of facts and then call it down to the thing I'm gonna present to my attending on rounds. Yeah. And that is, there's a lot that goes into going from novice expert and patients have not done that. Yeah.

Euan Ashley: Well, and I think about, you know, a chest pain history, you know, being similar, like we teach very early, uh, the questions to ask and, and, but we do that in the context of limited time. Mm-hmm <affirmative>. I mean, ultimately those questions are, are you, you want to leave space for the patient to talk openly, but then you need to drill down. Yes. And that's the skill we teach. The one thing these models have. And I think we could, we could talk a little bit more about some of the, uh, data you presented on empathy and some of it came from Stanford. Yeah. Um, but the one thing they have is essentially infinite time. Yes. Right. <laugh> <laugh>, which we do not. Yeah. Um, so how do you think that shakes out in the end? What is the balance there? Where we have is, I mean, essentially infinite compute, infinite time to sort of talk, ask questions. How do we use those tools then and, and the benefits of that model along with the still, there are a few things we both hope, I think, that are uniquely human. I hope so.

Robert Wachter: Yeah. We'll see. <laugh>

Euan Ashley: <laugh>. Yeah.

Robert Wachter: I, I, I try to come into that question with some equipoise. Yeah. You know, recognizing I have a bias toward the humans. 'cause I am one, my daughter and son-in-law are physicians. Yeah. And, and I kind of, I mean, it's, it's, it's, it's too depressing to believe that everything we learned and all of the humanness about what we bring to healthcare has no value. Yeah. Yeah. That's depressing. On the other hand, I think we have to have a two-tailed test, right. That says maybe there's some value there, but what if the machines get as good at it as we are and are more scalable or less expensive and are more convenient, you know, then they should probably win. I, you know, I I I, you can imagine a world where, for example, think about the, the life of a primary care doctor. Yeah. I've got 15 minutes to see a patient.

Robert Wachter: Um, and I've got a hundred things to do that the patient has interacted with a GPT like thing Yeah. Before the visit. Ha. It's, it's had, it has infinite time. Maybe the patient does or doesn't, but a lot of the questions have been asked. And then it's called, it's down to then delivering to me as the primary care doc, sort of a relatively complete data set. Yeah. And maybe even a, a tentative diagnosis that I then test. And again, am I gonna be able to sleep at the switch or am I gonna be, you know Yeah. Awake and, and, and really probe that. I don't know, I I, we've got a lot of work to do to figure this out. The tests of, of AI empathy versus humans, you could argue are unfair. Yeah. Because when you look at why the AI does better than the humans on a measure of empathy Yeah.

Robert Wachter: It was largely because the responses of the AR are longer. Longer. Right. But that's not, in some ways not unfair because the AI has infinite time and we don't, and so that's, that's, that is the real world. That is the test. And so don't know. I think one of the real interesting things that's gonna happen, I think about a lot about primary care. 'cause I think the predicament of primary care is awful. Yeah. I think, you know, very few people go into it. It's underpaid, it's underappreciated, and the job is essentially impossible. Yeah. You know, the patients come in with 10 problems on 15 medicines, got 15 minutes to deal with it. A study a few years ago, it said, if you did only spent your time on prevention, yeah. That's 27 hours a day. That's if no patient had the tamari to have anything wrong with <laugh>, that's 27.

Robert Wachter: So how do you deal with that? Well, some of it might be the AI does some pre-work and then hands it to you. Some of it though might be, you know, for the management of your blood pressure or your cholesterol, it's pretty algorithmic. Yeah. Maybe that is all done by ai. Right. Right. Now there's a challenge there, which is if I'm a primary care doctor and I look at my schedule and the only patients I'm seeing are the wildly complex patients with social needs and, you know, massive, complex, all the easy stuff has been taken off my plate by ai. Yeah. Yeah. I'm not sure that's made my life any better, unless you give me enough time to, to deal with that and reimburse me in the right way. But I, I think that's the way this is gonna shake out that yeah, we're not gonna fully replace primary care doctors, but we're gonna say, what is the things that humans can do uniquely?

Robert Wachter: Yeah. And one of the things that the AI can do well and safely and more conveniently and probably at a lower cost and figure out how to divide up the test that way. Yeah. And I think we probably will be happier in that, in that version of the world. Think, I think, I mean, patients will be happier patients. I think so. Yeah. I mean, I am happier as I mentioned, you know, I will take a waymo over an Uber. Yeah. I prefer to be in this thing that I know is incredibly safe. Yeah. And I can sing Springsteen at the top of my, my lungs in the backseat or take a nap. Is that

Euan Ashley: Safe for everyone else?

Robert Wachter: I, as long as I'm alone, it's, it's, it's great. <laugh>. Uh, so yeah. I think, you know, I think we as, as human doctors may overvalue how much patients get out of seeing us. Yeah. And I think for the things that can safely be done by technology, if it can be done safely and conveniently and they don't have to take half a day off to see the doctor and sit in the waiting room, I'm not sure patients are gonna be longing for the Yeah. Human touch. Yeah. But we've gotta be thoughtful about that. 'cause there are some things where, you know, we don't want ai, I don't think telling a patient they have cancer or diabetes or heart or, or, or, or kidney failure. Um, and then so how do you divide up those categories in a way that that actually works? Yeah.

Euan Ashley: It's interesting. I mean, Stanford is, is and UCSF, you know, probably, you know, not globally representative of of medicine, but, uh, it's now more than a year ago that someone called to get an appointment with me. And I think it was a few months before they could get it. Mm-hmm <affirmative>. So their follow up question was, when can I see Dr. Ashley's ai? Yeah. And I mean, that's a genuine, like, patient call. And my nurse at the time was like, I don't think we have that yet. Uh, but the, but the idea that we would each have an avatar that actually is built like us Yes. Maybe it's actually fine train, fine tuned at the minimum or, or trained on our records Oh, will notes it

Robert Wachter: Will be or will be. No, the book starts off in a scene where I'm sitting with the CEO of the Mayo Clinic. Yeah. And he had on his computer at something that he hadn't even shown to his board yet. Yeah. And I could see, should I show it to him or not? And he said, I, the look on his face was, this is just so cool. And he turned around and it was a Mayo radiologist speaking to a patient about the results of her x-ray. She had rheumatoid arthritis, and it showed it was the worst. And he was wonderful and compassionate, and the patient seemed to value the conversation. And then behind him in walked the real, the real doctor, Dr. So-and-so who waved awkwardly to the camera. And Jane Ricker Faria said to me, you know, 30 years ago when we wanted to scale the Mayo Clinic expertise Yeah.

Robert Wachter: What did we do? We built a campus in Scottsdale and a campus in Jacksonville. He said, we wouldn't do that today. Yeah. We would say, how do we take, you know, Dr. Ashley avatar Yeah. And make that that con level of competence and humanity and empathy available to more people. And it's not trained on the entire internet. It's trained on you in the last 10,000 patients you've seen, or the last 10,000 patients have been seen by great doctors at Stanford. I think that's inevitable. Yeah. And, and probably net good for patients. You think about, you know, not only not being able to see A-U-C-S-F or Stanford doctor, but not having a cardiologist within a hundred miles. Yeah. And you live in rural California. Absolutely. And now you have access to that level of expertise. That's pretty, pretty terrific. Yeah.

Euan Ashley: And I think that even, you know, uh, the USA has more cardiologists than I think any other country. So if you think about on a global scale Yeah. That the number of people who don't have access to subspecialty care.

Robert Wachter: I remember I was in Haiti a few years ago, they had, I think two oncologists in the entire country. Right,

Euan Ashley: Right. Yeah. I think Alaska had one cardiologist. Yeah. Um, but I, you know, thinking about that and, and scaling of subspecialty care, there's also the question we talked a little bit earlier about scaling empathy mm-hmm <affirmative>. If, if, if we understand that people actually do treat, even though they know the, the model is a model as a computer, but they can still feel the benefit of that empathy even although they understand at some level that this machine isn't feeling like a, like a human. It's, it's acting.

Robert Wachter: I've had, I've had discussions with my palliative care people about this, and they're kind of mixed about it. Yeah. But in some ways they, they say, you know, when we teach our students about empathy and an end of life conversation Yeah. Or in some ways we're teaching 'em a formula. Yeah. It's like we're teaching. They, you know, you, you, you identify with the patient, you reflect back what they said. You, you give them a plan, you don't take away their hope. Yeah. It's kind of pretty formulaic. Now you're starting with a human. And so they bring to the table whatever humans bring to the table in terms of empathy. But, you know, these machines have learned to simulate it. And again, they will begin taking on more human form. And you can imagine a world where people kind of forget that, you know, I'm not actually talking to a real human.

Robert Wachter: Yeah. And you're seeing that with some of the mental health chatbots where there have been some egregious examples of horrible things that they've said that went off the rails and are, are, cannot be defended. But there are millions of people today getting what they see as useful mental health counseling and coaching and advice. And you might say, that's awful, but try to find a mental health professional in Palo Alto or in San Francisco, and if you can find one, try to find one for less than 300 bucks an hour. Right. You know, so I, you know, the Biden quote of, don't compare me to the almighty compare to the alternative. Yeah. And the alternative in lots of parts of medicine is we have these great institutions and great people, but try to get in and see one. Yeah.

Euan Ashley: I think we raise the bar too high sometimes for technology. I think the quote encapsulates it beautifully. Uh, if, if you don't, if your alternative is nothing Yes. You know, which it often is. It often

Robert Wachter: Is. Yeah.

Euan Ashley: Then, uh, the bar for something Yeah.

Robert Wachter: <laugh> is, is

Euan Ashley: Lore. <laugh>

Robert Wachter: It should be. But, but then we do something which is predictable and natural, which is the bar should be relatively low if your alternative is nothing. And yet for a new technology, particularly one that people find weird and are scared about, particularly about their jobs, the a single bad accident will become a cause celeb. Yeah. And we saw this in San Francisco. It used to be, there were two driverless car companies. There was Waymo and there was Cruise. And two years ago, a cruise car didn't actually run over a, a a one was hit by a regular car. Ah. But the cruise then came along and she was under the wheels and it dragged her about 20 feet. And that company was outta business a year later. Yeah. And even though you could, the data were quite clear that it still is safer than a car with a driver. But our tolerance for a bad outcome of a technology is very different than our tolerance for a bad outcome when a human's at the switch involved. Yeah.

Euan Ashley: I think that it just takes us to, to another, another sort of decision making kind of regime in, in our heads. Yeah. One of the, the things I really liked as, as an aviation enthusiast, you, you often make comparisons to medicine and aviation and often in those comparisons. And you're not the first doctor to have made those comparisons. But medicine doesn't come out all that well, <laugh>. Um, but the, the reason I bring it up here is the de-skilling issue. Yeah. And I, I wondered, talk a little a bit about that. Yeah.

Robert Wachter: Well, medicine doesn't come out that well because I think aviation is a field that has really been incredibly thoughtful about safety and about integrating technology. And for the digital doctor, I actually went to Boeing and spent a couple days with seeing how they think about cockpit design. And you know, when I told them that at UCSF, there was a study several years ago that looked at the number of alerts that got fired by our bedside monitors in the ICU. Yeah. So this is when something is out of parameters for your heart rate, your pulse or O2 SAD and the EKG tracing in a single month in our 80 ICU beds, there were 2.5 million alerts fired in a month. And a nurse I was interviewing for this said, you know, said, uh, you know, the alerts are going off all the time. How do you, and I said, what would get you scared about something going wrong?

Robert Wachter: She said, if there was no alerts going off, I'd be really worried. <laugh> silent. That's, that's, that's like wacko world. It's upside down. Yeah. So aviation's been very thoughtful about the integration of technology and it's demonstrably safer than it used to be. But there had been a few high profile aviation crashes where the technology, uh, went on the fritz. Yeah. And the pilots now had to fly a plane without their technological wingman. Right. And in a couple of very vivid cases, the, uh, most prominent being the crash of air friends flight off the coast of Brazil several years ago, the technology went wrong. They're flying through a squall, and the pilot had to make a decision about what to do and did precisely the wrong thing. Yeah. And when I interviewed Captain Sully about this, and he said the pilots were flying a plane that they were unfamiliar with, which is basically a plane without the technology.

Robert Wachter: So de-skilling happens very quickly. I think in medicine, it's quite something we need to worry about quite a bit as the technology, if it's giving you draft diagnoses or draft therapies, you know, how do we or, or a technologically enabled colonoscopy or surgery. Yeah. Right. Now you might say, I don't care about de-skilling. 'cause the tool is better than the humans. Yeah. Right. But what if you don't have access to it? What if it's not working? Or what if it gives you every now and then bad results? Yeah. Um, we had a, one of the radiologists at our place told me about a patient who had a AI enabled CT scan and it read, uh, it read the breast shield as the skin. Ah. And therefore said that the patient had bilateral lung nodules, which turned out to be her nipples. Yeah, I see. <laugh>. And, you know, that's the sort of thing that the AI will sometimes do if it doesn't understand the context and having a human look in is helpful. But it's something we gotta think a lot about in our training world. Yeah. Because, because if we just automatically give the answer to our trainees, they will get dumber. Yeah. Over time, it's not their fault. Yeah. We've gotta figure out how do we prevent that from happening. Yeah.

Euan Ashley: I think we can, we can feel it ourselves. I mean, those of us said, you know, you, you and I have both for the last couple of years, I think pretty readily turned to these models, uh, to explore them and understand them. But I've found myself, caught myself recently in the last few weeks thinking, oh, maybe I should spend a

Robert Wachter: Bit longer do this myself, thinking

Euan Ashley: About this myself. Yeah. Uh, for whatever it is. Sometimes it's, oh, what's the name of that song? Or, right. Who was the person in that movie? Or, you know, what are the fourth and fifth things are my differential, uh, it's, you're,

Robert Wachter: You're a better man than I'm Right. <laugh>

Euan Ashley: <laugh>. But it's

Robert Wachter: Just the fourth, the fifth things are the differential. I would say. Yes. Let me try to think. I don't wanna de-skill. Yeah. What's the name of that song? My threshold to go to GPT is zero. It's like, I'm not gonna waste neurons on this. Yeah. Right.

Euan Ashley: No. Well, I applaud you for that <laugh>, but I, some little, I, I don't know. I mean, we think about our, our older patients and thinking about cognitive skills and thinking about trying to continue to en you know, engage their brain Yeah. And everything and, and even the, the novelty of learning new things, whether it's a new language or something as you go on. I do do worry a little bit about that. There's the classic study with the brain scans and I think the London cab drivers when GPS came in Yeah. And that was another, another example.

Robert Wachter: Yeah. I mean, none of us knows how to reap and that's Yeah. Sort of reasonably useful de-skilling. It's like I don't have to Yeah. And you could argue we should all know because what if, you know, Google's down. Right. It's like, I, that's not worth, yeah.

Euan Ashley: We need the North star <laugh>. We need to Exactly. A compass. Uh, this is how we, this is how we get ready. Yeah. We'll get, we'll get by. Yeah. Yeah. But, uh, well, Navi, navigating medicine is, is the thing that we're, that we're all, uh, trying to do. Um, overall, you're, you're, you're optimistic, I think. How, how do you see the next It's, it's, I mean, it's dizzying how fast this moves. I mean, giving a talk, you know, a month later on ai Yeah. Uh, I have to change a whole bunch of the slides. Um, but you, you have this really great kind of historical view in the sense that you've sort of analyzed different generations of medicine. So I'm really interested in, in, in particular, and we're here, like, this is the future of medicine. Yeah. So we know where we are today. And it's pretty, it's somewhere that would've been hard for us to guess, I think two or three years ago. Mm-hmm <affirmative>. Certainly five years ago. No, no chance. Great. But even looking ahead two or three years, I won't say five or 10. Yeah. You know, where, where do you think we are and what are the things that do change? And what are the things that don't change? You have the great story about Jeff Hinton saying that radiologists gonna be outta jobs and

Robert Wachter: Yeah. In 2016 he said, do we stop training new ones? 'cause clear by 2021, we'll never need any 2021 and we can't hire radiologists fast enough. Yeah. Same

Euan Ashley: Here. Yeah. You know, we have a, we have a shortage, so we're clearly not great at, at predicting the future. At least Jeff, although he was great at many things, remains great at many things at the moment. He's, he's a pretty, uh, a bit of a doomsayer about the future actually. But, but in, but in terms of medicine, where do you, where do you see us in, in just two

Robert Wachter: Or three years? Yeah. I, I spent a lot of time thinking about this and interviewed over a hundred people to try to get as smart as I could. Yeah. And landed in a very optimistic place. Now, part of that is because the technology is unlike anything we had before. And part of it is because the system is so screwed up. Yeah. That I think we need it. And I, I think as I saw what this technology could do, in many ways it was, uh, it was perfect fit for many of the problems. In fact, many of them, as you said, were created by technology. Right. Um, I'm also sort of a student of politics. I was a poli-sci major in college, is sort of how I think about the world. Yeah. And I think part of my optimism is I think the, the most important incumbents in the health system or the doctors, maybe to some extent the nurses.

Robert Wachter: Yeah. And if either of those two guilds were worried about losing their jobs, they would push back very hard on this. It would slow it down a lot. I don't think they are, and I don't think they should be. Right. I think for the foreseeable future, the unmet needs are so profound that the technology will actually enable them to do their job better. And actually, you know, they'll be happier. Yeah. And we see this in radiology where the, the radiologists say, bring it on. I want this stuff. Yeah. And I, you know, they're smart. They know maybe 15 years from now, I don't have a job, but for now I have to get through my day. So I'm optimistic because I think we're being smarter about how to implement it. We're starting with relatively low hanging fruit where we're being successful and that before we start on really ambitious stuff, uh, because the incumbents are mostly welcoming it.

Robert Wachter: Yeah. Because it's solving a lot of the problems in the health system. I think one of the big questions is, will it save money? Mm-hmm <affirmative>. Um, we're really good at healthcare at figuring out a way of not doing that. Yeah. <laugh>. Exactly. We good. Like, and spending more money. And spending more money. And there of course is the argument that these tools right now are enabling us to bill better. Yeah. And keep our, our prices high and all that. But you can imagine a world where the tools, and this may depend on a change in the incentive system, but the tools driving us to deliver more cost effective care, the tools lowering the administrative cost of medicine and all the bureaucratic stuff. And that probably does mean some layoffs in the billing department. Yeah. Um, but you know, as I net it all out, I think that, um, that it's net positive over the, over what I see as the foreseeable future. 15 years from now, I still think it's probably net positive. But now you're in a world that's so far over the horizon that anybody speculating that I can predict. Think about, you know, what you might have said about social media as it was rolling out. Yeah. And how nobody was predicting what it would do to, to education or to teenage mental health or to politics. This is gonna have all sorts of effects that I think are hard to know, but it's took over a five year time horizon. I think it's pretty positive. Yeah.

Euan Ashley: Well, thank you for leading us into the future. Thank you for writing such a compelling book. Oh, thank you. And thank you so much for visiting Stanford.

Robert Wachter: Always a pleasure. Thank you. Thank you.