Plain English with Derek Thompson - Is AI Really About to Solve Human Disease?

Episode Date: October 3, 2025

I’ve had the privilege of talking to many brilliant people about artificial intelligence. And when you ask them to imagine the most beneficial consequences of this technology, they almost always give the same answer: medicine. The dream is dazzling. Superintelligent AI will cure stubborn diseases and disorders—cancer, schizophrenia, Alzheimer’s. It will diagnose all our illnesses, design new lifesaving drugs, accelerate clinical trials, and pair with wearables to fight chronic illness and extend our health spans. But which of these promises are realistic? Which are outlandish hype? And what, exactly, can AI do for us in medicine right now? To separate fact from fantasy, I talk with Lloyd Minor, dean of the Stanford University School of Medicine. If you have questions, observations, or ideas for future episodes, email us at [e-mail]. Host: Derek Thompson Guest: Lloyd Minor Producer: Devon Baroldi Links: https://www.newyorker.com/magazine/2025/09/29/if-ai-can-diagnose-patients-what-are-doctors-for https://www.worksinprogress.news/p/why-ai-isnt-replacing-radiologists?utm_source=substack&utm_medium=email Learn more about your ad choices. Visit podcastchoices.com/adchoices

What's up? It's Todd McShay, host of the McShay Show at The Ringer and Spotify. We're building this thing up and I couldn't be more excited to be back, talking college football and everything NFL draft with the most informed audience out there. That's you. My co-host, Steve Mention, I will be with you three times a week throughout the football season with all the latest news, analysis, and scouting intel from around the league. For even more insight, subscribe to my newsletter, the McShay report to access my mickey. drafts, big boards, tape breakdowns, and other exclusive scouting content you can't get anywhere else. It's going to be a great season. And I hope you'll be with us at the McShay show every

step of the way. Today, AI and medicine. I've talked to all sorts of people about artificial intelligence, people who think it's the most important technology in the future and people who think it's a bubble, people who are doomers and people who are optimists. And when I talked to Talk to folks who are most positive and hopeful about this technology, and I ask them, what's the best thing it could possibly give us? It's striking that they all give the same answer. It will solve disease. Superintelligent AI will cure cancers, schizophrenia, Alzheimer's.

It's almost as if, of all the mysteries in the realm of forbidden knowledge, the mysteries of human biology are the most tantalizing. And so we imagine AI as a door that, once opened, will lead us into longer, healthier lives free of debilitating disease. So I've been following the story very closely, AI and medicine. And I've watched three very different narratives unfold in the realm of AI and what it can do for us in terms of solving disease. These three narratives are the fantastic, the pessimistic, and the realistic.

The first story brings us to Harvard University. In July, the New Yorker writer Drew Kular, travel to Harvard's Countway Library of Medicine to witness a showdown between man and machine. In this contest, an AI named Cabot, faced off against a brilliant expert diagnostician named Daniel Restrepo. The game was, who can solve a deep medical mystery faster? The showdown was reminiscent of three decades ago when the chess grandmaster Gary Kasparov faced off against Deep Blue and the IBM supercomputer famously beat him.

So here was the medical mystery. I guess you can play along at home if you want. A 41-year-old man suffers 10 days of fevers, body aches, swollen ankles, a painful rash on his shins and fainting spells. He goes to the hospital. A CT scan shows lung nodules and enlarged lymph nodes in his chest.

First up, it's Restrepo, the human. He looked at the fever, arthritis, swelling of the lymph nodes, lower extremity rash. He says it all points in one direction. Lothgren syndrome, which is a rare form of inflammation. Restrepo, it turns out, was right, and the audience cheered. Then the moderator returned to the podium to announce that while Restrepo had six weeks

to prepare his presentation, the large language model had a mere six minutes. Then the AI spoke, and here I am quoting directly from Kular's article. A woman's voice, warm and casual but professional, filled the room. Good morning, everyone, it said. I'm Dr. Kabat, and we have what I think is a really instructive case that links dermatology, rheumatology, pulmonology, and even cardiology. So let's jump right in. The voice, whose style and cadence were indistinguishable from those of human doctors,

began to review the patient's medications and medical history. No exotic exposures, Kabat said, just life in urban New England with a cat that scratched him six months ago, which, you know, I keep in the back of my mind, but I'm not married to it. The audience laughed. The AI generated an array of possible diagnoses, pointing out the strengths and weaknesses of each. It noted that the patient had high levels of C-reactive protein, a biomarker of inflammation that is sometimes associated with autoimmune conditions. Putting it together, Kabat said, the single best fit is acute sarcoidosis manifesting as Lofgren syndrome.

For a moment the audience was silent. Then a murmur ripple through the room. a frontier seem to have been crossed. End quote. There is absolutely no question that some large language models, especially those designed specifically to diagnose common and rare diseases, are already doing the work

of expert diagnosticians that have spent decades refining their skill. Something important is clear. clearly happening here. Now, that's the fantastic narrative. The pessimistic story isn't hard to find. For all the hooting and hollering about AI's potential to design new drugs, there is no drug available today that is, quote, AI designed.

There are hardly any AI design drugs in clinical trials. There's lots of exciting work being done at the intersection of AI and medicine, but AI inflected medicine isn't moving much faster than non-AI medicine. What's more, as Kula himself reports, OpenAI's GPT4 answered open-ended medical questions incorrectly, about two-thirds of the time. There's suggestive evidence that people are using chat GPT and getting meaningfully, consequentially wrong answers.

A poison control center in Arizona reported a drop in call volume, but a rise in severely poisoned patients. The center's director suggested that AI tools might have steered people away from medical attention. By this account, AI is not remotely ready to be or replace our doctors. And finally, beyond optimism and pessimism, there's realism.

There's no field of medicine more vulnerable or sensitive to artificial intelligence than radiology or using medical imaging to diagnose and treat disease. The vast majority of AI-enabled medical devices are for radiology. And in fact, about a decade ago, the computer scientist, Jeffrey Hinton,

declared that, quote, people should stop training radiologists now because AI would simply wipe out the profession. But, as the online journal works in progress recently reported, demand for human radiologists is higher than ever. In 2025, American Diagnostic Radiology Residence programs offered 1,208 positions across all radiology specialties, a record. Last year, radiology was the second highest paid medical specialty in the country,

with an average income of $520,000, 40% higher than when Jeffrey Hinton told everyone to stop trying to be a radiologist. So what we have here is not a simple story. What we have is, I think, a deeply complex story. Artificial intelligence hallucinates and is an expert diagnostician. It's a brilliant radiologist, but it's also not replacing radiology. So I wanted to be able to talk to somebody who fully saw this landscape clearly and could help me separate hype from reality. Lloyd Minor is today's guest.

He's the dean of the Stanford University School of Medicine. And what we do today is we go through some of the boldest claims made about what artificial intelligence will do for us in medicine. I've heard from folks that it will diagnose all our diseases, like Cabot. It will help us design new life-saving drugs. It'll accelerate clinical trials so that we can test those drugs more efficiently. And it will join with wearable devices from aura and, Apple to fight chronic illness and extend our lifespans.

These are the promises being made of artificial intelligence in medicine. But are they true? I'm Derek Thompson. This is plain English. Dean Lloyd Minor, welcome to the show. Thank you, Derek. It's great to be here. I'm really grateful you're here.

There are so many wild claims flying around about what artificial intelligence will be able to do for us in medicine. It will diagnose all our diseases. It will help us design new life-saving drugs. It'll accelerate clinical trials that we can test every drug that we invent faster and cheaper, and it will join with wearables to fight chronic illness and extend our health spans. These are incredibly dramatic claims for technology that is still in its infancy. And what I want to do with you in our brief time together is interrogate each of these claims very directly and narrowly to understand what can this technology actually do for us right now, right? Like, the future is a foreign country. Like, if AI cures every cancer in 2013, I will report in 2013 that AI has cured every cancer,

but it hasn't done that, so I don't want to get ahead of our skis. Does that sound okay as a general landscape of the show for you? It sounds great, Derek. All right. So four claims that I just outlined and that I want to scrutinize in greater detail. Claim number one, AI will help us diagnose disease more efficiently in a way that will save lives and money. And to give a little bit of context from my understanding of this space, in the last few years, we've gotten studies from MIT, Harvard, Stanford, and other schools that have repeatedly found that chat GPT doesn't just equal doctors at diagnosing several illnesses. It beats doctors, including sometimes doctors who work with AI to diagnose diseases. Just this year, Microsoft released a paper that found that its technology, quote,

solved, end quote, more than 80% of case studies, which was about four times more than the human doctors that it surveyed. Now, that's a Microsoft study. They're of course going to say that their technology is four times better than humans, but that's the context. Is this effect real? Do you think AI's already a better diagnostician than most human doctors today? And what would be the significance of that? Well, Derek, I think, first of all, at the broad level, and you outlined four very exciting topics that we're going to cover in our conversation today. But I'm reminded at the beginning of our conversation, I'm reminded of the statement attributed to many that we tend to overestimate what we can accomplish in the short term and underestimate what we can accomplish in the long term. And I think perhaps we're seeing some of that play out today as we deploy AI in all aspects of life, but in particular what we're talking about today, AI in the application to health, health care, and discreet.

discovery science, life sciences. Now, on the first topic of how is AI impacting today, AI impacting the diagnosis of diseases and the selection of the correct treatment for diseases and in general, how do we get, identify the right patient for the right therapy at the right time? I'm reminded of, you know, I'm probably best known in my field, which is the way the inner ear balance system works and in the field of otology and neurotology, because I described an inner ear. ear disorder in 1998, developed a surgical procedure to fix it. For many, many years after those first papers were published in which I, my colleagues described the syndrome and then had a bunch of follow-up studies, still, most of the patients coming to see me diagnosed themselves from doing a Google search

at the time. And, you know, there was a lot of concern when the Internet first came, you know, widely available, and people are talking about Dr. Google and how Dr. Google is going to get us in trouble. And for sure, there are examples of that. But the fact is, I think if you look on balance, and we're now, before we talk about the AI era, just the value of search, search has really democratized access to medical information, and I think has helped many, many people and many, many, many physicians. Fast forward now to AI, and everything that I think search has enabled us to do and improving access to information and then applying that information to diagnostic situations, I think

we're seeing that the effects of AI and taking that to a whole new level. You know, is it going to eliminate the role of a physician-diagnostician in the near future? No, I don't think so. But just in the short years, you know, what Chat GPD was introduced in November 2022, just in the relatively short period of time, less than three years since its introduction, we've already seen evolution of the models. We've seen fewer hallucinations, and we've seen more accurate, focused information coming back for them. So today, I think many physicians are using large language models to assist in diagnoses. What we've done here at Stanford and many of our colleagues at other institutions have done the same

is that we've in-licensed models so that we can actually put elements of the medical record into a large language model, but in a secure, closed environment so that there's no breach of patient privacy. But in order for physicians to get information back on a series of rare, findings in a patient, and how do you fit all those together? And in particular, when you're selecting a therapy and you're selecting a medication, and maybe you don't know what the probability of an adverse reaction to that medication is in that patient, because that individual patient has a constellation of medical findings that isn't necessarily readily apparent in the knowledge

base of the physician or in the medical record. I think it's thrilling to think that artificial intelligence has access to this digital cosmos of symptoms. and disease and etiology and can be a genius at answering questions that would require a typical doctor to find a needle in a haystack. I mean, at a very personal level, my dad died of a vascular cancer

for which the entire medical literature had something like 12 cases. And so if there's an artificial intelligence that can see those 12 cases with a clarity that a human might not be able to, my dad might be alive. So I come at this from a place of hope. That said, I have two worries, and I want to make sure I articulate these worries clearly.

One worry is over diagnosis. So I think about artificial intelligence interacting with the existing financial incentives in healthcare. And if AI is incorporated into a system that rewards treatment, then its ability to find the tiniest little blips in radiology or pathology reports seems to me to be very likely to lead to too much treatment. Some of it necessary. Some of it may be unnecessary, some of it may be harmful. And one could imagine how that would lead to, say, health care costs going up in ways that don't necessarily extend people's lives. Are you worried at all about doctors who don't yet entirely understand how to use this technology, using it to find the tiniest little cysts that mean absolutely nothing to people in a way that raises health care

costs and adds over-treatment? Yes, I think it's a valid concern. And I believe, we've seen that risk and the manifestations of that risk with the introduction of every new medical technology. For example, now a standard way to evaluate the function of the heart is through an echocardiogram. It's a non-invasive study involves no exposure to ionizing radiation. The equipment and the technology has become very sophisticated. You can get accurate measurements of the size of each of the chambers of the heart, the pumping, of the heart, the thickness of the wall of the muscle, the diameter of the valves in the heart, huge amount of information.

And guess what? There are findings probably more commonly than not. The valves size is a little bit different, or the Arctic root is maybe a little bit larger than we typically see in a person of that age. And then what do you do with that information? So that's, and we go through waves where perhaps until we can, get enough medical knowledge to know that that slide enlargement of the aortic route or whatever is not that significant. And maybe the echocardiogram should be repeated in a couple years or five

years, but nothing should be done at the time. But until we got that knowledge and understanding, there were probably people that did undergo further workup and maybe even some that underwent procedures. So I think any new technology has that concern, that risk, and it ultimately is dependent upon its responsible use. However, ecocardiography has vastly improved our treatment of common cardiac diseases, enable us to know when we need to intervene, enable us to know how to adjust medications, how to judge a patient's risk factor for a variety of different conditions later on. So I don't think we would do away with it, but there's been a curve, an evolutionary curve, of understanding how it can be appropriately used

and how findings that may deviate from the, quote, normal need to be interpreted in the broader context of the patient's health and of the status of the heart. I really like that answer, and one way that I'm synthesizing it for myself, is that I think some of the most boosterish predictions about AI say, oh, not only is it going to find all the diseases, it's going to save us so much money.

And there's a tension in there. If AI helps us find more illness, it will probably lead to more treatment. And in the capitalist healthcare system that we have, more treatment is going to cost more money rather than less. So it's not as if we're going to enter this, I think, Valhalla, where health care costs plummet and we find all the diseases and everything is hunky-dory forever,

it's possible that the actual real successive AI at detection will raise overall health care costs in some ways rather than reduce them. The second thing I wanted to push on is I'm interested in the idea that some doctors probably won't use large language models for diagnosing enough, but some doctors might use it too much. So, for example, there was a paper that was published in, I believe it was Lancet, gastroenterology, and hepatology, which found that after three months of using AI to detect cancers in colonoscopies, doctors reportedly got worse at finding growths on their own.

And I think that's so interesting because so much the literature on AI and medicine is about, these Luddite doctors rejecting the technology when it could make them better. But here we have evidence of some doctors accepting AI, but it makes them, it, it, it, it diskills their human intelligence. Is this a valid fear, or maybe another way to put the question, is are you more afraid that doctors won't use this technology or that they'll use it so much in a way that they're actually deskilling themselves? Yes, let me return to your first point, just to make a conclusion. a remark, I do think that's a risk in terms of AI actually driving health care costs up.

And we've seen that in a lot of cases with the initial introduction of new technology. However, the real hope and opportunity is that AI will ultimately lower health care costs because we will diagnose diseases earlier and therefore treat them more effectively. So, you know, in the United States, we have a great sick care system. We spend a huge proportion of our medical expenditure is on care of complex diseases. If we could reduce some of the complexity because we've diagnosed and intervened earlier, or maybe in some cases figured out specific preventative strategies based on our risk factors that AI enables us to identify, then ultimately we can lower health care costs.

So that would be my concluding thought on our first topic. Now, on the second topic, absolutely, I worry a lot about that. And as a medical educator, how should we be educating the physicians of the future? How should we, those of us who are practicing physicians, be thinking about the responsible use of AI as we have more and more knowledge, not only information, but knowledge at our fingertips. You know, medical education, the best analogy I know of has been it's somewhat akin to learning foreign language. What do you have to do to learn and communicate in a foreign language? You have to know the vocabulary.

you have to know the grammar. Then you've got to be able to synthesize and put together the vocabulary and the grammar in order to communicate with someone else and interpret what you're hearing back from them or reading in a case of reading the text in a foreign language. In medical education, we have to know the vocabulary of medicine. We have to know the grammar, which is how the body works, the biochemistry, the way the organ systems interact with each other. And then we have to put that together when we're seeing a patient to synthesize a diagnosis, come up with a treatment plan and then, you know, carry out that treatment plan, obviously

in combination with the patient and their family. Now, where along that spectrum will AI be a disruptor? Well, in terms of the vocabulary, knowing sort of the medical information, as you pointed out, AI is really good at that. It's only going to get better. There's no turning back. And we'd already started to move away from, thank goodness, some medical memorization. I mean, when I was in medical school, not only did we memorize, you know, the names of drugs, their mechanisms of actions, we had to memorize their dosages. And that was terrible because human brain isn't designed to keep arcane numbers accurately. And the consequence of making an error is significant.

Well, now, of course, we don't do that because dosages are generally calculated based on the patient's height, weight, and other characteristics. And we maybe have some general knowledge, but that's taken care of with a lot of checks and balances in the background. One can envision AI, you know, taking that a few steps further and actually identifying what the correct medication is. And then the real issue from the standpoint of the role of the physician is how do you evaluate that information coming back from AI to know, yeah, this seems right? Or, no, I don't think so. And I'm going to push it a little further. I'm going to question it. And the risk is, as you said, that we become, and you cited the example of colonoscopies and AI identifying suspicious lesions.

and then gastroenterologists not becoming as accurate as they were before without AI. And that's a concern too, although there are other cases where technologies have sort of gone beyond what traditional skills may have been. For example, before brain imaging, neurologists, the neurological exam, neurologists could do exquisite. accurate, fine-tuned exams that could take hours, and they can generally pinpoint the lesion from very, very subtle clinical findings. In the age of MRI,

the need for that has really essentially, has gone away. It's an elegant, elegant form of medical, physical exam and reasoning, but it just doesn't have much of a place when we can get such accurate imaging, or at least not the same place that it did,

you know, 40, 50 years ago, before we had MRIs. So I think there's some evolutionary process here where, yes, there are some things that the AI is going to take care of and take care of once it's trained better than humans will. But until we get to that point, we definitely need to be careful about moving too quickly away from the knowledge and skill set that is enabled us to get to the point where we are. The second claim that I want to interrogate is AI will help us design new wonder drugs. I feel like every few weeks I read a headline that claims that AI has designed some drug to

combat a superbug or be the next super penicillin or maybe even take on certain cancers. I would love you to tell me, as concretely as you can, how are scientists using artificial intelligence to design drugs that they wouldn't be able to design without AI. And maybe another way to get at this question if it unlocks something else is, if AI is a tool in the drug discovery process, what is that tool good at that's new? Right. Well, one thing it's good at today is, you know, Deep Mind, which is now part of Google, a number of years ago published a large language model or made available a large language model that can predict and calculate the structure of any protein by knowing its amino acid sequence.

And it turns out to be amazingly accurate in doing that. Okay, if you're designing a protein biological drug, protein-based drug, and you know the structure of the receptor, You know the structure of your protein, but you want to manipulate the structure of your protein just a little bit. Then you can, with the simulation, move amino acids around and look at how the confirmation, the structure of the protein changes. And make a prediction then, oh, this is probably the molecule I want. This is the protein I want. That has to be experimentally validated.

But before you'd be doing a lot of guesswork, you could do some calculations based upon the, you know, the charge. of the charges of the amino acid, how they interact with one another. But the model is so much better than any calculations that were made before alpha-fold was introduced. Now we're seeing a similar approach being used for nucleic acid structure and function. So today has the application of AI to processes of drug discovery massively change the drug discovery process? No, it hasn't. A lot of the approaches that are being used today by the many companies that are deploying AI

is to get as much biological information as you can. So if you're working on the liver, you know, get slide after slide after slide of liver disease and pathology in a variety of different situations, all very, very well annotated, and then, you know, train your models based upon that biological data

that you have absolute confidence in, and then start to do manipulations and make inferences from a database that is as pure as you can get for the need that you have in your drug discovery process. That's a logical approach. It's also, you know, in this approach, the AI process is actually not the most difficult part of the process, the most difficult part is getting all the biological data in to train the model as well as you can. And that takes time. And so, again, it's back to the statement.

We tend to overestimate what we can accomplish in a short period of time, underestimate the long period of time. I think on a scale of five to ten years, that AI is going to dramatically change drug discovery processes. But it's proving to be a longer time scale than I think some people had thought when it was first-centered. when we first started seeing companies focused on this, say, five years ago, four or five years ago. So I remember- Let me jump in there because once again, I don't want to just repeat this over and over again. But like I come at this, like wanting this technology to work. Like who doesn't want AI to figure out how to design certain proteins or certain, you know,

certain immunity blockers that ends up being absolutely fantastic for helping to cure certain diseases? But I do think it has been, what, four years now? since that alpha-fold result was published, I believe, in the Nature Journal. And there were some people saying, look, this is going to revolutionize drug discovery immediately. You know, it's been four years. And to my knowledge, there is no drug that exists that patients can take that is so-called

AI discovered, right? We were mostly in the pre-clinical phase of working out these drugs. I don't want to necessarily say it's been disappointing. But maybe unpack a little bit more about what we've learned about why it's taken a little bit longer than we thought to use this technology to design drugs that help people. Right. And Derek, you know, when the human genome, the first map of the human genome was published out 25 years ago, it was like, ah, we've just cracked all of disease. And look, we have made tremendous progress because of that initial work. And of course, now genome sequencing is available at a relatively low cost. However, it's proven to be much more difficult to actually intervene and treat or prevent disease just from genome sequence information. I think there's an analogous situation of what we're seeing in AI-enabled drug discovery.

It for sure can be used to design molecules. It can be used to curate biological information. But by the way, and you mentioned the fact that as of today, there's no AI-generated. drug therapeutic that's gone all the way through the process, FDA approval and is on the shelf in pharmacies across America across the world. And there are lots of reasons for that. There are oftentimes off-target effects. You can have what seems like a perfect drug. And by the way, this happened long before AI. And then there's an off-target effect. If it's a small molecule that you never expected, you didn't really have a reason to expect it,

but it showed up and it delays or in some cases cancels an entire program. So it's a tough business, and I don't mean that as an excuse, but also I'm optimistic that in the end, the drug discovery process, we're going to have more effective therapeutics at a broader array of targets than we've had with the conventional approaches in the past. I think though that's we're really looking at a before we see that sort of transformation we're probably looking more at the three, five and ten year timeline

than we are and there were people that were hopeful it would be much faster than that but it's proven it's proving to take a lot longer Convierte your passion in a new with Shopify and batter records of ventas with the form of pay with a better conversion of the world Has you know, the best one? I wonderable system of

Shopify facilita the free of your website on your website, and in the world. That is music

for your ears. No, you'll be more your business will a super-exit

with Shopify. Empeas your period of a month in Shopify. coms bar records. I wonder if

what we're missing with AI and drugs is a new platform technology. So with diagnosis,

we have it. The large language models can memorize millions of case studies, it can formulate a theory of symptom and disease, and it can articulate its answers to patients and doctors in words. So the interface of large language models like chat chabit is perfect for diagnosing patients who want to know why they're sick. But if the interface, if you will, for drug discovery isn't complete, that seems key here. Like, let's say you design a drug

that blocks a key protein. Okay, well, what knock-on effects? Is that drug going to have throughout the body? We have no idea. We have no idea whatsoever. We have no platform technology to give us that answer. Is there a missing invention

that we need with artificial intelligence and medicine that we're just missing right now? Well, let's take one specific example off-targeting. effects, which are oftentimes the reason that small molecule studies fail. And we can see that with biologicals, with protein-based therapeutics also. If we had knowledge of all the receptors of all the cells and the interactions of all the receptors of all the cells and the detailed metabolomics

of every cell, then we might be able to do a, we should be able to do. We should be able to do a much better job of predicting off-target effects. We have some knowledge today, particularly with regard to mechanisms of liver metabolism, as to when diseases are likely to be toxic, when therapeutics are likely to be toxic to the liver. But there are many other cases where we simply don't have the detailed knowledge or the assimilated knowledge of all the metabolomics of cells to be able to make predictions of these off-target effects. So I think that's a data acquisition and curation issue, but that would certainly be of great utility in the AI deployment for drug discovery. As we get better at simulating the metabolic functions of cells, and then we can computationally calculate how interfere it with one.

particular pathway is going to affect the overall physiology of the cell. Already that's being done, but that will undoubtedly be done more with greater and greater sophistication as our knowledge of the pathways and our simulation of those pathways increases. So, I mean, those are a few examples of where I think, you know, the, where I think additional studies will help AI be even more impactful. drug discovery. But it sounds to me just for my own edification that like you're saying, the breakthrough here is that we've digitized an understanding of protein construction and we've digitized the design of certain molecules, but we haven't digitized a map of the effect those molecules

would have throughout the body and throughout tissues in the body. And so it's sort of like if drug discovery is like a maze, it's like we found we're like the opening of the maze is, but we haven't actually created like a digital record of what the actual maze looks like inside the body. And so we invent these molecules, and it's great, and it seems like it's going to work exactly on the target. But we have no idea, once we launch that pinball, what it's going to do inside of the body.

That kind of makes sense to me. I think that's a great analogy, great metaphor. And it demonstrates the complexity of all these processes. Right. All right, claim number three is that AI will accelerate clinical trials, which will allow us to test drugs faster and cheaper. This might sound a little bit nerdy to some listeners. If you're in the medical community, you know how unbelievably essential this is.

Drug development has gotten exponentially more expensive in the last few decades. In the reporting for abundance, I spoke to one researcher at the Broad Institute who told me that on a per-patient basis, the cost of making an FDA-approved drug, has increased by 10x in the last 20 years. What makes you most optimistic that AI could accelerate clinical trials that we could test drugs faster and cheaper? Several things. First, large language models today are enabling us to identify patients who qualify for a clinical trial. For example, we have a clinical trial. For example, we have a at Stanford have developed our own large language model based upon others that are available

commercially, but focused on looking through medical records in order to call out patients that may be eligible for a new clinical trial that we're just bringing online. Previously, how has this been done? How have we reached out or identified patients who may be eligible for any, whether it's a phase one, two, or three clinical trial. That is, whether it's a, it's a brand new drug, or it's a drug that's in the last stages before FDA approval. The way we've done that is, generally a physician who takes care of patients with diseases for which that drug is applicable goes through their data and says, oh, I have this patient I saw, you know, a few months ago, and maybe there's there's a record of some of the mutations in a particular

tumor, if we're talking about a cancer. But it's been very, very ad hoc. And that's why it's been difficult to get patients enrolled. The other thing is just more sophistication of monitoring the progress of clinical trials so that we can adapt, you know, so-called adaptive clinical trial design, which is rather than saying, we are going to enroll 1,600 patients in this phase 3 trial, and then we will break the code and decide whether or not the therapeutic has been effective. Now, pharmaceutical companies and others conducting these large trials are able to set up criteria to monitor in real time and make adjustments in the trial, again, based upon statistically valid methodology, adjustments in the trial in real time enabling the trial to be much more efficient and effective

at the end at knowing if the drug is effective and if so, at what doses is effective for what categories of patients. So with claim one, AI diagnosing diseases, I think we can say, yes, this is happening. With claim two, AI designing new drugs. Yes, it's certainly trying to design new drugs, but we don't yet have a drug that's AI designed that someone can, you know, go to their doctor and be prescribed. To what extent with claim three of AI accelerating clinical trials,

is this happening? Is it like kind of happening and that we know that AI is being brought to bear on clinical trials, but we don't yet know if it's making the trials faster or cheaper? Like, what is the level of reality of claim three compares to? compared to claims one and two? I think the level of reality is it's being deployed today. We don't yet know whether or not it will have incremental impact or transformational impact. I think it'll take a couple of years to know that.

But if we're predicting and believe me, I don't have a crystal ball for any of this. And I've learned that many, many times in my career. but if we're predicting that the full impact of AI on drug discovery is going to be on a three, five, ten year time scale, I think the full impact of AI on clinical trials is probably a two to five year time scale. That is to know, is this going to be a total game changer in terms of extending the efficacy and the prevalence of clinical trials or is this going to be an incremental advance? I think we'll have a pretty clear idea in the next two to five years. All right, that brings us to claim four.

This is the claim that AI combined with wearables will be a key weapon in the fight against chronic illness. So speaking personally, I've got an aura ring. I'm wearing it right now. It tells me heart rate variability, sleep time, body temperature, all of that. Apple Watch has done studies showing that it can detect AFIB and potentially, therefore, be a warning signal for heart issues. So these devices are telling us how long we sleep, right, body temperature, heart rate variability. it's not hard for me to just look up my HRV, which was, I think, 55 yesterday, and determine if that's, like, good or bad for someone my age. So I wonder here, what can AI plausibly give us that these devices aren't already telling us?

Like, what more does AI theoretically bring to the picture of wearable, individual data, and helping us stay healthier for longer? It's a great question. It's perhaps the area where it's most complicated because, as you point out, we're already able to get a lot of information about the function of our body through devices that are available off the shelf today. I think where AI will offer the most benefit is in knowing how much of that, let's put this way, we're getting a lot of data from those devices. Now, how do we derive information from that data?

Okay, so we're showing a larger heart rate variability with exercise. Is that significant? Or maybe just reflect that the person is fatigued. You know, their overall physiological cardiac function is not as efficient as it was because of fatigue or other factors. So I think that's where AI will be helpful. It is making sense out of all the data that we get. I think it's great we're getting all the data. You know, one of my late colleagues, Sam Gambier,

late chair of radiology here at Stanford used to remark, every time we fly on a commercial airliner, those jet engines are being monitored hundreds of times a minute with data being beam back down to Earth. And routine maintenance is being done based upon real-time information coming from the engines. it's enabled them, you know, it's enabled airlines to reduce the, you know, the incidence of catastrophic engine problems and made their maintenance much more efficient.

We don't do anything like that for the human body. And maybe we shouldn't for everyone, but for people who have multiple medical problems, people with congestive heart failure who are constantly teetering on the brink of being either fluid overloaded or fluid depleted. well, it would be nice to have a lot more information and be able to intervene proactively and keep people healthy at home rather than constant trips to the emergency department. But still, and probably there we have the best developed algorithms. But with general health, knowing the significance of a myriad of different findings

and changes that we're picking up from these devices is hard to do today. And AI should help us determine that significance. So about three weeks ago, I got my first blood test in, I won't even say how long, too long. I get the blood test back and the website where I'm reading the test results, it's showing me my number for certain indicators like LDL, bad cholesterol, and a bunch of other words and acronyms that I don't understand that tell me of my body. And it says the number and then it gives me this range. this healthy range in which my number either is falling inside the healthy range or falling outside the healthy range.

When I got this panel back via email, it just looked like a mess. I didn't know what any of it actually meant. So what I did was, I took a screenshot of, like, maybe eight screenshots of every single part of this panel. I uploaded those screenshots to chat chabit, and I just entered my information. I said, I'm a 39-year-old man. I just got these blood tests back. tell me what this means, and tell me in particular if there's anything here that should strike me

as very worrisome or on the opposite, like very encouraging. And I just got this brilliant readout of my blood panel. And someone might significantly ask, Derek, how do you know it's brilliant? You're not a doctor. Well, my actual doctor called me 24 hours later, and I swear to God, she gave me the exact same information with the exact same context that Chat ChbT had given me in 15 seconds. What does that experience tell us about the future of artificial intelligence as a new or more sophisticated layer between patients and medicine? Oh, I think they indicate very positive things about the future of our use of AI medicine. When we started our conversation today, I mentioned earlier my career describing an inner ear disorder. It took a while for

other specialists in the field to get to know it. But patients, had access to the information instantly through the internet and started diagnosing themselves. So this is, and of course, the advantage of AI is it takes to the whole next level, because you can give it a complex series of data points and ask it to assimilate those into some sort of a reasoned analysis of your risk factors or is this significant or isn't it. And so, yes, I think it's going to do all of those things. help bring knowledge to our fingertips. People who have no training in medicine

can now understand things that previously it would have been difficult to have deciphered or it would have taken a lot more time and effort to understand or put together the significance. And now it's available like for you with a simple search. Last question. I'm quite hopeful about AI's effects in medicine in the long run, as you said.

I think the short run is going to be difficult. I think medicine is just incredibly hard and understanding new tools and how they can actually work in medicine is just going to require a little bit of a learning process. Where I'm worried about artificial intelligence and this connects, I think, to health is that I think we live in time where people are disconnected from each other, that loneliness is rising and alone time is rising and screen time is rising. And in many cases, artificial intelligence is not necessarily going to help this. There's going to be people who don't go out with friends because they're staying home,

you know, looking at TikTok, which is, you know, an A algorithm. They're not going to see a therapist because they say, I've got Dr. GPD that I can talk to. Maybe they don't see it at PCP because they're like, why do I need a PCP? I've got my aura ring. And we're learning more, and I've done recent episodes about the degree to which social connection isn't just this woo-woo thing that mom said was good for you, like you should make friends. social connections seems quite protective neurologically. People who are super-agers and retain fantastic memory at 80 seem to, one thing that correlates

with super-agers seems to be their high levels of social connection. There's all sorts of research that seems to indicate that connection seems to be cardiovascularly, metabolically, good for us. Being around other people is good for our bodies. And I wonder what your thoughts are about this world of artificial intelligence that we're moving into that might keep us sort of more trapped with digital devices and digital selves, which in many cases is what AI sometimes seems to us to be, and the health risks of that, and how we can balance both the best of AI while not giving into the worst of it by allowing it

to accentuate our aloneness crisis. Derek, I think it's a real concern. Of course, that is the isolation and disconnectedness. But it begins. of course long before AI. I mean, that's a complaint commonly stated with regard to social media in general that people can spend all their time interacting with TikTok or watching videos online, whatever. Whereas before, in order to have that sort of connectedness, you had to be sitting in front of someone or a group of people.

So it's a challenge then to know, well, I don't think we never turn the clocks back on technology, but how do we responsibly use the technology or maybe think of ways that AI actually helps people to maintain more connectedness rather than less connectedness because AI identifies, again, without violating privacy, AI identifies connections with others in, in, ways that might not have previously been known, or there are discussions around AI that bring people together in order to have those meaningful discussions that wouldn't have occurred had there not been an advance in AI that provided a reason for the discussion. But for sure, the loneliness problem is a huge issue, and it is growing, not declining. And like you were saying,

there's great data that a variety of different health indices are improved. by social connectedness. Dean Lloyd Minor, thank you very much. Thank you. It's been wonderful being with you today, Derek. Thank you for listening.

Plain English is produced by Devin Beraldi, and we are back to our twice-a-week schedule. We'll talk to you soon.