S1, E35 - Barry P. Chaiken, MD, MPH: Physician-as-Patient Perspective on AI in Healthcare

Summary

---

Practical AI in Healthcare – S1, E35

Physician-as-Patient Perspective on AI in Healthcare

Guest: Barry P. Chaiken, MD, MPH
Hosts: Steven Labkoff, MD & Leon Rozenblit, JD, PhD
Date: May 3, 2026

---

Episode Overview

This episode brings a unique perspective to the intersection of medicine and artificial intelligence: Dr. Barry Chaiken, renowned internist, informatics leader, and two-time cancer survivor, shares his journey transitioning from physician to patient. The conversation explores how being a patient altered his approach to illness, how AI can and cannot empower patients, and what it truly means to use AI tools safely and effectively in healthcare. The hosts and guest also debate challenges of digital health literacy, clinical trial innovation, and the critical need for industry-wide rapid learning and governance frameworks.

---

Major Discussion Points & Key Insights

1. Physician Becomes Patient – The Emotional Shift

[04:08–07:59]

• Barry’s Story: As a preventive measure, Barry monitored his prostate health for years after his father died of cancer. When diagnosed himself, his immediate response was not the rational, systematic approach he’d expect from a physician — but the emotional, paralyzed reaction of any patient.
• Key Quote:

  "When you're sick, you do not think rationally as you would as a clinician. You think emotionally and you don't think about things in a clear way." — Barry Chaiken [06:15]

• A friend challenged him to seek expert opinions and more advanced care, reminding him of the well-known adage: “A doctor who treats himself has a fool for a patient.”
• Reorienting with Clinical Reasoning: After being jostled into action, Barry deliberately compared clinicians, evaluated procedural experience, outcomes, and made an evidence-based decision—contrasting typical patient experience.

2. AI and the Patient Journey: Opportunities and Pitfalls

[12:01–17:03]

• AI as a Patient Resource:
  • Barry didn’t use AI during his initial cancer diagnosis because generative AI wasn’t widely available.
  • Reflects on how, today, ChatGPT and similar tools are both powerful and risky in patient hands.
• The Human-AI Interface Problem:
  • AI works very well for doctors (over 95% diagnostic accuracy) but poorly for lay patients (as low as 37%).
  • Key Gap: Clinicians know how to prompt and interpret; patients often don’t.
  • Key Quote:

    "If you do not prompt [AI] appropriately, it will come up with crazy results...The doctors know how to prompt the tool because of their background, but the average patient doesn't." — Barry Chaiken [13:31]

• Barry advocates for using AI as an explanatory tool, not a diagnostic one, especially for patients: focus on understanding results, framing thoughtful questions for clinicians, building background to support shared decision-making.
• Analogy: Using AI is like using Excel or a Da Vinci robot — the results depend on your input, purpose, and skill.

3. How to Prompt and Co-Work with AI

[17:36–20:03]

• Treat AI as a tool, not an oracle—its output reflects the quality and specificity of your input.
• Always specify what you want (e.g., “Explain these results and suggest questions to ask my doctor”) and what you don’t want (e.g., “Don’t diagnose or prescribe”).
• AI Sycophancy:
  • Barry instructs AI not to agree uncritically:

    "I do not want you to agree with me all the time. If I wanted a puppy, I would get one." [18:52]

• Self-teaching by experimenting with prompt engineering builds literacy and improves usefulness.

4. AI Literacy for Patients and Clinicians

[21:50–24:54]

• AI literacy is now essential in healthcare for both patients and providers.
• Barry’s method for patients: Use AI to understand, not to diagnose; verify all information; maintain trust with clinical teams.
• With OpenEvidence, a clinician-facing AI, he ensures every AI advice comes with references—advising physicians to read, verify, and use AI as collaborative support, not replacement.
• Important Principle:

  "Trust but verify. That really applies to AI too." — Barry Chaiken [24:51]

5. Clinician’s Superpower: Saying “I Don’t Know”

[24:54–27:09]

• Knowing and admitting the limits of your knowledge is crucial. Doctors are trained to seek advice and work collaboratively.
• Automation Bias: There’s risk in embedding AI decisions into workflows without critical review—humans may overtrust AI’s suggestion simply out of habit.
• Key Analogy: Traffic light positions—habitual behavior can override actual context or evidence.

6. AI for Everyday Healthcare Problems

[27:09–29:26]

• AI can resolve real-world issues beyond clinical care.
• Example: Used AI to craft a diplomatic, effective email to a dentist that led to a fair, non-confrontational resolution after a bad outcome for his wife.
• Emphasizes that AI’s benefits and effectiveness greatly depend on the user’s skills, expectations, and context.

7. Governing AI: The Need for Collective Rapid Learning

[31:07–37:21]

• Barry critiques current decentralized efforts (like the Coalition for Health AI): too slow for AI’s pace; too fragmented.
• Advocates for a model akin to the Aviation Safety Reporting System—a public-private rapid learning system capturing, centralizing, and sharing lessons and incidents from across the sector.
• Quote:

  "We can't afford to distribute the management of AI to all different organizations and let them figure out how best to use it...There has to be...a way...for people to share what they learn." — Barry Chaiken [31:17]

• Highlights risks of automation bias and lack of resources, especially for smaller providers.

8. Clinical Trials, Digital Patients & Interoperability

[40:16–46:23]

• AI and synthetic “digital patients” will revolutionize drug development, allowing in silico testing before human trials, optimizing patient selection, and accelerating recruitment.
• Interoperability: The industry has lacked incentive alignment; technical barriers are less daunting than economic ones.
• Barry’s Banking Analogy:

  "If we wanted interoperability, we'd have it...It's really about how do we get the incentives to be aligned..." [43:14–45:43]

9. Advice for Patients and Clinicians on Using AI

[47:48–52:32]

• For Patients:
  • Don’t use AI alone to diagnose or select treatment; it can amplify anxiety and misinformation.
  • Use AI to understand labs/results and formulate questions for clinicians.
  • Quote:

    "Don't use AI to figure out what's wrong with you and how you should be treated, because it will make you anxious. It has a good chance of giving you misinformation." — Barry Chaiken [47:48]

  • Trust and open communication with your care team are paramount.
• For Clinicians:
  • Experiment with clinical AI tools (like OpenEvidence) using hypothetical patients in your field.
  • Use AI to generate patient-friendly explanations and question lists to support shared decision-making.
  • Remain the final filter for AI output and educate others on safe and effective use.

---

Notable Quotes & Moments

• On Prompting AI:

  "It's just like that Excel spreadsheet is only a tool and it doesn't know anything. It doesn't have scruples and morals and objectives and goals and ethics or anything. It's just really a tool. And if you don't control that tool, it will give you junk, it will give you misinformation, it will hallucinate." — Barry Chaiken [17:36]

• Healthcare’s Unique Trust Requirement:

  "There's no industry like healthcare. There's no place where you implicitly...trust every single person that you interact with in that place, from the doctor to the janitor...That trust is a sacred trust..." — Barry Chaiken [49:30]

• On Governance and Learning:

  "Aviation doesn't let each airline figure it out and let them compete on safety records. That reminds me when people talk about just let each health system figure it out. We have a shared problem in a pre-competitive common space." — Leon Rozenblit [35:35]

---

Key Timestamps for Main Segments

• 01:50: Introduction to Barry and framing “doctor as patient”
• 04:08–07:59: Barry’s cancer diagnosis, reaction as patient, friend’s intervention
• 12:01–14:29: The promise and challenge of AI for lay patients vs. clinicians
• 17:36–20:03: How to approach, prompt, and learn from AI tools as a patient
• 21:50–24:54: AI literacy and OpenEvidence experiences
• 24:54–27:09: The problem of automation bias for clinicians
• 27:09–29:26: Using AI for practical, non-clinical health problems
• 31:07–37:21: The necessity of centralized, collective AI governance and rapid learning
• 40:16–42:33: Digital patients and AI-powered clinical trials
• 43:14–46:23: The economics and incentives problem with interoperability
• 47:48–52:32: Actionable advice for patients and clinicians on AI

---

Takeaways for Listeners

• AI is powerful but double-edged: It can clarify or confuse, depending on knowledge and approach.
• Prompting is everything: Explicit instructions and boundaries matter—like any tool, outcome reflects operator skill.
• Trust and literacy are key: Both patients and clinicians must build AI literacy, prioritize reliable information, and foster trust in the clinical relationship.
• Systemic solutions are essential: Healthcare needs collective learning and robust incentives for safety, interoperability, and rapid, equitable AI implementation.

---

Transcript

[A]

Welcome to Practical AI in Healthcare, the podcast that cuts through the noise to spotlight real world solutions delivering real world value. From patient care to clinical research, from life sciences to patient engagement, we focus on what's truly moving the needle in healthcare. No hype, no theory, just practical insights where AI is making a true impact. Welcome aboard and let's get to it. As many of our listeners know, Leon and I work very closely with the DCI Network Division of Clinical Informatics at Bethesdril Deakins Medical center in Boston. This June, the network is hosting Patient powered Digital Health 2026. The conference will bring together patients, innovators, industry leaders, healthcare providers and policymakers to shape the next generation of real world patient centered solutions. The meeting will run from June 22nd to the 24th in Boston at Harvard Medical School. We've arranged for our listeners to get a discount on registration to the meeting. If you register between now and May 15th and use promo code PracticalAI June no spaces, you'll receive 30% off your registration fee. You can learn more and [e-mail] patients2026. In addition, we're always looking for sponsors. If you or your company are interested in becoming a sponsor, please reach out to admincinetwork.org see you in Boston. Hello and welcome to this week's edition of Practical AI in Healthcare. My name is Dr. Stephen Lapkoff and I'm here as I am every week with my co host, Leon. How's it going?

[B]

Hey Steve. Doing great. I'm so excited for our guest today.

[A]

So they say that doctors make the worst patients. They think they know everything about their disease. Sometimes they get really into micromanaging their own disease. And there's a saying that says a doctor who treats himself as a fool for a patient, and that's a long standing adage they teach you in medical school and it's to try to dissuade you from doing exactly that. And yet doctors are patients and doctors become ill. And today's conversation is going to be with a doctor who got sick. He's someone I happen to know for a long time. His name is Barry Chaikin and Barry's going to share with us his journey through a diagnosis, a difficult diagnosis, and how he brought AI into helping him manage that throughout the course of his disease. Barry's an internist and he has been in this field for a number of years, at least a couple of decades actually. And he's held a number of roles in places like himss. He was the board chair of Himss. And he's also very well versed in healthcare AI. So with that as a background. Barry, welcome to the podcast.

[C]

Thanks for having me. I really appreciate it.

[A]

So, Barry, why don't you take us back to how you got where you are today. We often ask our guests for their superhero story. How do they get their cape? How'd they get their tights, How'd they. How'd they get to where they are today? Just give us some background and then we'll dig into your journey.

[C]

Well, there are a couple of things that I always like to suggest that make you have a great life. We know about health, we know about a good family, raising as children, a good education, things like that, intelligence. But one of the things that I think is so important is curiosity. Because if you're curious, there's no day that's ever boring. There's always something you're going to be curious about, interested in, that's going to excite you. You're going to learn. And that's really been the journey of my personal and my professional career. I've always been curious. Went to medical school, did preventive medicine residency, a little of internal medicine, joined corporate America in the 90s, did some research a bit before that. I've had amazing experiences throughout my professional journey because of the fact that I was curious, and I'm really grateful for it.

[A]

So why don't you take us back to your diagnosis? What happened? And when did you go from being Dr. Barry to being patient Barry?

[C]

Okay, so when I was in med school, actually, my father was diagnosed with prostate cancer. And soon, a couple years after that, he passed away. Really horrible. They didn't have treatments back there in the 80s, so he really, really suffered. So basically in the late 30s or so, I started doing regular physical exams and a PSA to check my prostate to see if I was developing a cancer. Was always normal, normal, normal for many years. But I still checked it every single year. One year it went from 1.4 to 2.8. Under 4 is normal, but when you double, it's always a concern. Went for a biopsy, found out that, like my dad, I had prostate cancer in six of 12 cores for the biopsy and decided to do surgery to see if we could cure it or at least put it on the back burner. So I had the surgery, did incredibly well. I was really, really lucky. It helped that throughout my career, I've always been athletic, running marathons and cycling for cancer in honor of my dad. And I think that helped a lot with my Recovery. So that was my experience. But I have to tell you, being a patient is a really different experience than when you're a doctor looking at and treating your patients.

[A]

So when we were doing our previous session with you, you'd mentioned that in the course of this situation, you had a friend who basically said to you, you know, what did this friend do? And say to you that would. That gave you a change in your direction, change in your course.

[C]

So if any patient came to me with a cancer diagnosis, I would immediately ask them, where are you being treated? Did you go for a second opinion? Do you think your cancer is appropriate for a clinical trial? Did you visit a medical center or cancer center designated by the US Government? That's what I would do. But why didn't I do that?

[A]

Well, did you?

[C]

Here I am in Boston. Why didn't I do that? I mean, what is the difference between me being a physician and being a patient? And you know what the big difference is? You're the one who's sick. When you're sick, you do not think rationally as you would as a clinician. You think emotionally and you don't think about things in a clear way. I had a friend just say to me, what are you doing? Just like that, what are you doing? Why aren't you seeking a second opinion? Why are you going to be treated in a community hospital where the physician is treating you is not an expert in prostate cancer or in GU cancers? So don't you think you should go ahead and look at a different place? And oh, by the way, at that point, for over 30 years, I had been raising money for the Dana Farber Cancer Institute through the Penmas Challenge, which is a two day, 200 mile ride from the middle of Massachusetts to Provincetown. And like, why aren't you going to the Dana Farber to get your treatment? Why aren't you getting evaluated there? And I'm like going to myself, literally. I said, you idiot, what are you thinking? Why aren't you thinking like a doctor? Why are you thinking like a patient?

[A]

And there goes the adage, the doctor who treats himself right. Well, it's good that you had a friend. It's good that your friend sort of jostled you into another perspective. You did have the clinical training, and to your point, it's hard when you're a physician to hear these diagnoses when it's you. And it really can knock you a little bit for a loop. But it also raises questions about, you know, how did you, how do you actually do this? Yourself. And then how did AI get into the picture for you? And how did you bring that to bear in the course of your own disease?

[C]

Well, actually, it didn't back then because I was diagnosed before ChatGPT was made generally available.

[A]

Okay.

[C]

But it's worth sharing this point. So I told you about how I was being a patient initially. Well, once I got jostled into being a doctor, what did I do? I had to choose the surgeon who would actually do the prostatectomy on me. And I went to Mass General, and I was referred to, through a colleague of mine to two different doctors. One of them did an open resection, and one of them did it laparoscopically, which is similar to the robot for those people who aren't familiar with that type of surgery. And in that, I asked them this question, how many have you done and what have been the results? So the first physician says to me, oh, you're thinking of going to the other doctor? I won't mention their names. And he said, well, we're colleagues. So I keep track of my outcomes for all of my patients for the following years, as long as I can stay in touch with them for at least five years. And my colleague does the same. And then we get together on a regular basis and we compare our notes and. And here's the results of what we found. So that's being a doctor, right, and making a decision. And then what I decided to do, I decided to do the laparoscopic because I said to myself, hey, I'm relatively thin, I'm very healthy. It makes sense for it to do laparoscopically versus somebody who might be a little bit overweight and therefore a little more complicated to visualize where the surgery would be done. And the choices I made, I'm really pleased with them. I was very lucky with the end results.

[A]

And to this day, you're doing. You're doing well, thank God, and you're able to lead a very much a normal life. You're back in action, you're lecturing, you're speaking around this topic, and you're involved again in healthcare AI at this point. So you know what, where did you go from there?

[C]

So that was the one cancer that was, quote, unquote treated. My PSA is undetectable and there's no signs of disease, thank goodness. I have colleagues and friends who have not been as lucky as I have been. So I'm grateful for every day that my activities of daily living and my sequelae from the surgery are really minimal, if non existent. Several Years after that I was diagnosed with a rare lymphoma called Waldron's Jumps, which is an elevated IgM in your bloodstream. Now again diagnosed by how my PCP saying hey, we have to do regular testing, right? And found out that an elevated protein, just a random typical blood test that they might do for baseline for someone of my age, elevated protein, worked it up, found out that it was an igm and then guess where I went for treatment this time it was directly to the Dana Farber. And I have a funny story about that one too. So being treated at the Dana Farber, it turned out that there was a Bing center for Waldron Strum's lymphoma, which is amazing that they would actually have one of the world renowned centers being literally a 40 minute walk from my house. So they would always ask me how long did it take you to get here, how many planes you took, how many trains you took, how far did you drive? And I said, I walked over, it's only 40 minutes. So I'm incredibly lucky there. And I've been treated really well and doing very well with that too. So I didn't really get. I know you were going to ask me about AI and disease, so I'll go ahead and let you do that.

[B]

Well, Leah, let me pick it up from you, Barry. I mean one thing I found fascinating about your story is that you had significant clinical training to fall back on and you still almost went down the wrong path by just being a patient. Right. And being freaked out like all of us are when something like this happens. And that raises a question of what tools could help an average patient. Right? You had your own experience with AI. You talk a little bit about how it helped you as and how you learned to use it. And also how would AI tools help an average patient who's not a clinician in the same situation?

[C]

You know, I'm really glad that you asked that question for a couple of reasons. One is I'm going to make believe where it's today and I was diagnosed. Many people have heard a lot about Google health. Right? Right. ChatGPT Health. There's some. Claude and others have these AI tools for healthcare. If you don't use AI frequently, you don't really understand what its wonderful greatness it can do, but also the bad things that it can do. And I'm quite nervous about these consumer facing applications in healthcare because the application is driven by what you share with it and if you do not prompt it appropriately, it will come up with crazy results. There was a recent study in a medical journal, I believe it was in Nature, who specifically showed this, where when the AI tool for healthcare was used by doctors, it was successful over 95% of the time in getting the correct diagnosis and recommendations. But when it was used by individual patients, the variation was enormous, where only 37% or so were able to get to the right diagnosis. And for something that was life threatening, it went from everything, from recommendations, went from everything. Visit the ER immediately. So you're okay, just go to sleep and take an aspirin, essentially. And the problem is, is that the doctors know how to prompt the tool because of their background, but the average patient doesn't have that type of a background to prompt it appropriately for it to give you the correct answers.

[B]

Yeah, Barry, that's a really important insight, right? That study, in your interpretation of it, suggests that a lot of the times where we think the problem is an EA model, the deeper problem is in the interface between the human and their knowledge state and the AI. What you know, I think that's a really good insight, but I don't know how to fix it. So how do people who know how to prompt effectively do really, really well? How do we go through a process of teaching an average patient to prompt more effectively or just interface with the system? I don't want us to be stuck in the prompting metaphor because that's just one way of interacting with these systems.

[C]

So I've used open evidence, which is open to physicians and other clinical people of a national practitioner ID number for this, specifically asking these questions, not about the disease process itself. But I'm not an oncologist. I know nothing about breast cancer or supraventricular tachycardias for cardiology, at least enough to be able to instruct a person appropriately. But I've used the tool to be able to say, I have this patient. Here are their parameters, about them, their history, the lab results, things like that. And I said, produce for me a document that describes all of the results in layman's terms and what questions that I should ask my clinician when I go to see them. So clearly, unless you're a practitioner, you don't have access to that type of tool. But if you're going to use any of these healthcare tools, you do the same thing. Ask it. Here are all of my results. I don't want you to give me a diagnosis and I don't want you to give me a treatment. And it's really important you tell the AI you do not want that. What you want is the purpose of my query with you, my interaction, my chat with you is I want you to help me understand what the disease I have and what the results mean per test and what questions I need to ask my clinician when I go to see them, my oncologist, my radiologist, whatever you go, who you ever going to see, ask them those questions. And what happens then is two things. One is you obtain a background, a basis for understanding your disease, but you also then continue supporting the trust you have with your clinician, who will then explain to you what's going on in a way that is not only you will understand, but is unique to you and who you are. And that's incredibly valuable. AI can't do that. Your clinician can. And what then happens is you become an active member of of your care team, which is invaluable to your health and getting better, but incredibly valuable to your clinician, who then knows your part. They know what you're thinking, they know that you're getting educated, they know what you need to do to get well. And it really helps them as well as helping you.

[B]

So Barry, I think what drove you to AI is just your natural curiosity and love of figuring stuff out. You discover ChatGPT early, like a lot of us did who thought this is too cool not to use. Right. So, but that the result was a self teaching process that you are now sharing the results of. Right. And I loved your reflection of like, what would I do differently if I got the diagnosis today? Can you talk about that self teaching process and how do you think an average patient can emulate that kind of immersion learning?

[C]

Okay, so the first thing you have to understand is if you're going to use AI, is that you are the knowledge person. You're the one who controls the tool. You don't go and open an Excel spreadsheet and say to the Excel spreadsheet, oh, give me a financial analysis of all of my savings and investments and stuff and provide for me a financial plan that I can use for the next 20 years. It doesn't work that way. Even if you use a basic Excel spreadsheet, all it does is create numbers and has formulas and can do math for you. But it doesn't create the worksheet that you use to be able to accomplish a task. You have to look at AI similarly, where people make a mistake is they say, oh, the AI is going to be able to answer my questions and do this kind of work for me. And it doesn't, it's just like that Excel Stream spreadsheet is only a tool and it doesn't know anything. It doesn't have scruples and morals and objectives and goals and ethics or anything. It's just really a tool. And if you don't control that tool, it will give you junk, it will give you misinformation, it will hallucinate. So what you need to do is you need to put in the time to tell it what information you have, what your goals and objectives are, what what it should and should not do to be able to make it effective. And one of the most important things that I do to AI is an instruction that I embedded in its memory. I said, I do not want you to agree with me all the time. If I wanted a puppy, I would get one. I don't need a puppy to lick my face, okay? So puppies are much more fun to be with than AI telling me how wonderful I am. And that's really important. Tell it not to always agree with you, to challenge you, but. But also give it all the information you can and also tell it what not to do and not where to go.

[B]

I'm tempted to emulate my ChatGPT experience, Barry, what a brilliant insight that was, right? But yeah, the anti sycophancy is really important. So one thing that jumped out at me, Barry, was that use both ChatGPT and Claude to write Navigating the code. Right? Your most recent book and use fed it. You were feeding it source material, you were generating chapter drafts and then like editing down 12,000 word chapter drafts down to type 2000. What did that process teach you about how these tools actually work versus how they think people think?

[C]

Well, first off, if I had to do that today, it would be so much easier because they're so much better than they were when I first wrote the book. But I want to address one thing about using AI for writing. There's a huge controversy about this. And you might have read about some people losing their jobs and because they used it to write a book or they used it to write an article and such. Here's the thing. On my website, I specifically say in my terms and conditions that I am responsible for everything that I write that appears on that site. And whether I use AI or not doesn't really matter. If I write something that's stupid, if I have a wrong link, if I go ahead and print something that's an hallucination from AI, that's not a reflection on AI, that's a reflection on me. So I take full responsibility. But in this instance, what I did is I constrained it, meaning I said, don't go out to the web, don't use any sources on the web unless I share something with you, only use what I share with you. And I used at the time, ChatGPT wrote better than Claude, wrote better than ChatGPT. So it was really a mishmash. I asked ChatGPT write it based upon my instructions. I did the same with Claude. I then put them together and asked Claude to rewrite it again. At the end of the day, I ended up with 12,000 words, and then I honed it down to 2,400. And I don't care what anyone says. That's my thinking, that's my writing, that's my doing. And I'm responsible for everything that's written in that book. And I checked every single reference that is in that book to make sure that was correct. Again, AI is a tool, just the way Google search was the year 2005. If you were writing a paper for your professor at college, you know, it made it easier than going to the library, but that was about it. You're still responsible for it.

[A]

So, Barry, you're raising a point that we have been talking about on the podcast now in several sessions, and it has to do with AI literacy. Leon and I are members of the Division of Clinical Informatics, DCI Network. And at our last conference in September of 25, the topic of AI literacy got a lot of attention, and you hit on it in a very interesting way, whereas you took your own personal literacy journey. Now, most people take a similar journey, but they don't always arrive where you did, which is getting very literate. And certainly with your healthcare background and training, that gives you a leg up. But you raised a couple of points that I think are really important, which are about how to interpret all of that information, how you advanced yourself. Maybe you could speak a little bit about that and how open evidence either helped or hindered that process.

[C]

Okay, so first off, let's separate them. Outside of open evidence, if you're just using Google and such for the general public, just be careful that it's going to hallucinate and you don't put the parameters around it, you're going to get in a lot of trouble. You always have to be skeptical of what it actually shares with you and what it says. So make sure you do that. Open evidence is a little bit different. I don't know the founders and how they put together their models and how they feed them anything, but I have to guess, considering the people who are going to use it and their ability to critique it fairly well, they've worked really hard to make it much more robust in terms of it not hallucinating. And what's interesting about it is every single time it tells you to do something, it gives you a link to the reference that tells you to do something, which I find to be incredibly helpful. Obviously, when I use it, and I suggest any clinician who does use it, they go ahead and read through it and make sure that resonates with what their knowledge base is. And if they question it, they should go ahead and do some of the research on it. I find that to be really useful. But I used it for two friends, actually, patients, one with cancer, another one with a cardiac issue. And one of them is actually walking in today with their oncologist and is going to have these questions. And when she was diagnosed with her breast cancer, she was incredibly scared, right, and nervous and such like that, and was concerned of what to do. And she was on the edge of treating one way or the other. The fact that she just had the questions was incredibly valued to her. It calmed her down. And then now she was going to work to develop her relationship with an oncologist. So open evidence is really, really good. And I know that a lot of physicians are using it. As a matter of fact, I was introduced to it by my pcp. But again, going back to the general public, just be really careful using these health tools in AI because they have problems with it and don't. As a famous president, Ronald Reagan, once said, trust but verify. That really applies to AI too.

[A]

You know, you made the point that clinicians have an advantage of knowing when it is what they don't, knowing what they don't know. Or at least the good ones know what they don't know. Doctors are oftentimes okay with saying to themselves and to others, I don't, I don't know what, what this means. What are the implications of that with AI and how is that going to evolve in, from your perspective, in how clinicians are going to be using it and as they go forward.

[C]

So I think clinician superpower is particularly doctors is using those three words, I don't know. I work for corporate America. If you wanted to get fired, the best way to do it is to go and say, I don't know. I know I'm taking that to extreme, but it's a different mindset. Trained as a clinician, you are actually proud to say I don't know, because that means you're going to go to somebody else who does. And you work collaboratively. And that's always the best way to work. We as clinicians, particularly in 2026, do work as a team. Doctors, nurses, therapists, NPs, et cetera. So that's really important. Using AI. I think the physician needs to be careful about trusting it as being an authoritative source per se, and trusting it even more than they might trust a colleague. And that's one of the risks we have with it tied to it, is that if we implement these AI tools within a workflow that automates the use of the AI, we risk at suffering from something called automation bias, where we go ahead and approve what the AI suggests without taking a moment to pause and think that maybe it's wrong or maybe it doesn't make sense, or maybe I should check it. And if for those of you who don't understand automation biases, just think of this. If you go to a if you drove into a town where the red light was on the bottom and the green light was on the top, at a stop sign, at a stoplight, many of you would drive through the red light because you automatically assume the one on the bottom is green and the one on top is red and your brain can't be bothered to try to figure out the color, just knows the position and that tells it what to do.

[A]

So you told us again in the pre visit that we do with our guests about I guess your wife had some dental issues and as a non dentist advising dentist, how did that go? Used open evidence.

[C]

In that case, I didn't use open evidence, but what I did was I used AI to help me with how to address the issue. Here's a really key point for all of you out there who are listening to this podcast. Did you know that if you see your dentist and they do an expensive procedure and you're not happy with the results that you can ask for your money back? I didn't know that. I didn't know that either. And my colleague and longtime friend, we went to high school together. We reconnect. A couple of years ago, he was one who said to me, oh Barry, you know, if you don't like what happened and my wife had a horrible outcome, I won't of course mention the dentist. You just ask him for the money back. Which we did. But here's the thing. I identified working with my colleague, what kind of message I wanted to give to that dentist to actually get my money back and what would be most helpful. And I went online and it was able to help me write the email to the dentist that in one instance would basically say, give us our money back and let's all avoid the problem. And that was a way to address the issues of the dentist. I'm not here to go sue people or any of that. That's not my makeup. It was just a bad outcome. There were some problems along the way and he really needed to give us the money back because he made some mistakes and mistakes in follow up, which of course is the other problem in dentistry. Dentists never get sued for malpractice for their procedure. They get sued for not following up after the procedure. This was the case in this instance. So the AI was incredibly helpful to be able to do that for me and made it so that I was able to have a reasonable settlement with this dentist that was not accusatory, not escalating, didn't need to require attorneys and anything. It was just a fair settlement for everybody involved and saved everybody a lot of grief. And it was. There's something that I would never could figure out how I could have written it in a way that would have conveyed that message.

[B]

Yeah, I love that insight, Barry. It sort of. I keep being struck by the differences in outcome that different people have using AI tools where some of us are like, this is amazing. We're getting the best results and other people just think it's garbage. Right. And there's this enormous difference in what we're bringing to the table and the expectations and the background. And I think one colorful metaphor used that I'm going to steal going forward is that the da Vinci surgical robot, like you could give one to me, not a surgeon, not an md and say, go ahead, Leon, perform some surgery. I am not going to do a good job. Right. So maybe you or Steve would get. Would only need a month of training to do a good job, whereas I'd need several years just not to kill people. And I think in some ways AI is a little bit like that. What we're learning is that there's some tacit knowledge that some of us are bringing to the table that we have to convey to folks who aren't. But let's, you know, a lot of us are trying to figure out, including Steve, who's been leading that effort on DCI side effort, trying to figure out what digital literacy means in the AI context and how to improve it. And there are also others that are working on governance models that are trying to make AI safe, equitable, effective and trustworthy. And I'm involved myself along with Steve and DCI and also in the DIA AI consortium. But let's talk about what's not working. Like, I usually take the bad cop role at some point. Right. Like lots of things that are moving forward and great. You know, you've been critical of some of the efforts. I mean, I know you've had some fairly sharp words to say about what the Coalition for Health AI is doing. And I know none of us are criticizing them as an organization. And I think we all agree that what they're doing is really important. But talk a little bit about what the problem you see with their current approach.

[C]

Okay, so AI is different from other technologies that we've had. Let's even go back to, you know, search and Internet and all that. It's just evolving so incredibly quickly and it's making things so easy to do, so to speak. But the problem is that the doing part, it does so much that we can't afford to distribute the management of AI to all different organizations and let them figure out how best to use it. We have to, in a lot of ways, bring the best minds together to be able to figure out how to use it, to understand what the problems may be and to share that information collectively. I use the argument about the aviation safety reporting system, which unfortunately we have not modeled in health care, which is a terrible shame. But people should realize the reason why very few people die in plane crashes today is because of that aviation safety reporting system. And if you don't know about it, just go back and read a short Wikipedia article on how it started with crashes in D.C. that should have easily been avoided, but there was no way to report the problem in the first place. They didn't even have strict definitions of what go around meant or clear to land meant at those days. Today that doesn't happen. So it's incredibly rare that we have those accidents. But let's look about AI. The problem is that the current system is trying to distribute their management of the AI or the research around how to best use of AI. The problem with that is that you have all these organizations who are working to discover the same exact fact. And there's no easy way to be able to share that types of information with each other. And it's just the technologies is so powerful and so advanced, it's just not an effective way to do it. There has to be using the aviation safety reporting system as a model, a way to be able for people to share what they learn about how to use the AI, best ways to deploy it, both workflow and otherwise, so that it can quickly. Everyone can quickly respond to what they learn from it. Creating model cards and stuff, it's all great, but it's just going to be so slow and so far behind how fast this technology moves. And. And I think people are going to be hurt by it. Let me give you just a simple example. When we had meaningful use and we had electronic medical records, we. At the beginning, there were some organizations that had the pharmacy department put together the pick list for the doctors. They put together a pick list the way they thought about pharmaceuticals, not the way doctors treat patients with those pharmaceuticals. What happened then? The doctor would look at a list and they'd have to scroll to the bottom to get to the one that's most likely appropriate for this particular patients, or most of the patients in this particular ward say it's cardiology. So it would be cardiology medications. Okay, so what happened over time? Doctors got rust, they got distracted, things would get problems, and they would accidentally pick the wrong one. That's a workflow problem. It's not a doctor problem. It's an implementation problem. And I can see the same thing happening in AI where AI becomes part of the model. There aren't specific stops in that workflow. And the doctor gets tired late at night, what have you, and goes ahead and picks something that isn't exactly good or exactly correct. You know why? Because they're a human being. And human beings get distracted, and human beings get tired, and human beings, because of those things, make errors. That's what we did in aviation. We tried to remove the human factor from causing a problem. We have to think of AI the same way. And I don't think the current models that are being used or current approaches being used are really able to do that. One last thing. I'm not advocating government to do it. I'm not saying that I know all the right answers and how to do this. What I'm saying is I think what we're doing today is incorrect. And I think that if we built together a public private partnership with all stakeholders, vendors, patients, clinicians, organizations, institutions, pharmaceutical companies, et cetera, I think we'd be better off to create a system where everybody can benefit from it and still have a stake in it, to secure safe, quality health care with great access, at the same time allowing our private sector make the various revenue they need to do to report to their private companies, I think all of that is. Is possible.

[B]

Yeah, Barry, and what you're saying is very close to our heart, and it's really well aligned to the way we've been thinking about it at the DCI Network and the couple of papers that we've produced on private public partnerships, I really want to emphasize for the audience that the aviation safety reporting system is an amazingly important model to look at. And I really recommend people looking into it as an enormous success story of creating a rapid learning system. The alternative would have been imagine let each airline figure it out and let them compete on safety records. That reminds me when people talk about just let each health system figure it out. We have a shared problem in a pre competitive common space. And if we can create a shared learning experience, we can move the entire industry forward faster and increase patient trust and patient safety at the same time. The other thing you mentioned, which is humans being human and planning for human error is part of the process, absolutely essential. And one particular way that that comes out is in automation bias. As we know, when you set up a process where the Automation is right, 99% of the time that 1% gets ignored. Right? Which is why the TSA employees, when they're on the job, actually have to have red, you know, red team drills where somebody comes in and pretends to sneak a gun through because otherwise they just stop paying attention. Because it's never a gun, right? Like, nor, you know, you see, you see a thousand bags and you know, it, it all looks like a gun, but it's never actually a gun. So the question is, what do we, what is the equivalent of that? How do we change the workflow around clinical AI to handle that automation bias that makes people start ignoring processes where they only have to pay attention 1% of the time?

[C]

You know what, we have to figure out how that workflow actually has to work. Whether that's some AI check on that AI or a check on the actual result, the user interface for the clinician that forces them to do something. The example would be the COVID over the switch that turns off the fuel supply to the plane's engines, those types of things. And you know, there are a lot of people smarter than me for sure. We could figure that out and we can test different things. Here's the other thing that's wonderful. Let's test different ways to do it and figure out the one that actually works the best. What I don't want to happen, I don't want the tech companies to decide how to do that. That's my fear. Because their motive for the tech companies is profits. And there's nothing wrong with that at all. That's their job. They have their responsibility and if anybody thinks they're bad guys, you're not thinking correctly. We live in a capitalist environment that has given us great riches. We just need to take care of those externalities, allow the tech companies be interested in optimizing the use of their technology. That's great, but let's not give them the keys to the kingdom. Let's make sure that we have other interests involved, safety interests, patient interests, clinical interests. To be able to do that, we can figure out the workflow relevant to all of this is. And a big fear that I have is if you're one of the large AMCs in the country and you have a lot of money and you can do this research and associated with an informatics department and such, you might be able to control what happens and make some changes. I'm worried about the smaller institutions that are around the country who are approached by a vendor who says, hey, I got this great AI tool. We're going to implement it. It's going to do all of this. And they don't have the resources to be able to check it, monitor it, do surveillance on it, see if it's working, and they'll just trust the vendor. Again, the vendor's not the bad person here. Just, we need a system that protects those smaller institutions that need to be able to receive technology that they know it works and have systems to be able to monitor it. One last thing. If you are working for a small organization and you do decide to go that path, some of the questions you can ask is, how often do you update your model? How often do you test your model? What type surveillance tools do you use to make sure that your model is correct? How do you use my data? How are you going to report back to me how my model is performing? And this is not on a yearly basis. This is at minimum on a quarterly, if not a monthly basis. Then you become an active participant in that technology and help ensure that it doesn't hurt you financially or hurt any of your patients clinically.

[A]

So, Barry, we're going to pivot because, as you know, I've spent a large part of my career in the life sciences. I knew your sister when we both worked at Pfizer. She's in the life sciences. And, you know, as we're again prepping, we talked about clinical trials, digital clinical trials, and how AI is changing that space. And we've had a number of folks on the podcast talking about it. What's your take on that space?

[C]

I think it's really exciting. Here's why we're Pretty close, maybe. We already have this, this digital person, the digital patient. You can have all their laboratory values, clinical stuff about them, genetic genomics, et cetera. And you go ahead and you, you. And we know from deep work on DeepMind and how molecules interact with each other in a biological environment, we can then ahead and do clinical trial in silicon, in the computer to figure out what's the promise of this particular medication itself. Okay. And then what we then can do is we can do two things. One is we can accept or reject whether we're going to try to do a clinical trial with people based upon those results that can increase our probability. We're not here to get an exact answer. We're here to increase probability that this trial will be successful. Then we can also identify the types of patients we want to have in that clinical trial to be able to really determine whether the medication's effective and safe. And I think all of that can be accelerated through AI also. And this is an extension of that. With our electronic health records, we can be able to determine which patients may be candidates for clinical trial. And we want to be able to do that while the patient is a novel patient versus being treated in other ways. That's really important in doing a clinical trial. So imagine if we had a way using AI to say, hey, this patient came into my office as a PCP in some rural American place, and we were able to identify, hey, you're a candidate for doing this clinical trial, and we can enroll you and use telehealth to be able to do that. I mean, one Steve, you know, one of the big problems in clinical trials is enrolling people in it and how long it takes to do that, we can identify them better, we can increase the probability the drug's going to work, or medications can be effective and safe. It accelerates all of that, and it's really good.

[A]

So, you know, the problem that we've had with health, it has been a perennial problem. Since I started in informatics, I've been in this field now, geez, I figured it out the other day. Is over well over 30 years. And interoperability has been dogging us pretty much the entire time. It's always just around the corner. Meaningful use was trying to get to interoperability. The USCDI is trying to get to interoperability. What is interoperability going to mean in the face of AI? Is AI going to make that a better problem or a solvable problem, or will the problem melt away, or will it just continue to confound us an Easy question.

[C]

No, that's okay. Here's my bias. If we wanted interoperability, we'd have it.

[A]

That's a good point.

[C]

Actually, that's the problem we didn't want. I mean, I can go through all the reasons. Probably other people on this podcast have given you all the reasons why there's no interoperability. If we really want it, we can have it. Here's, here's the thing.

[A]

That's your hymns hat you're putting on there for a second that you're wearing your hymns hat.

[C]

Yeah, no, I'm not going to wear the hymns hat here. I'm going to wear my banking hat.

[A]

Okay.

[C]

Okay. So I'm old enough to remember that if you had an ATM card, the only way you can take money out of the ATM was going to the bank where your ATM card was. So I know most of the people on this podcast don't know those days, but you can look them up because the banks saw that the ATM was a way to encourage savers, investors, what have you, clients to come to their bank versus another bank. So in a lot of ways, that ATM machine was a cost center for them. They would lose money on the ATM machine. But they said, ah, it's like a marketing cost. It'll bring people into the bank. Then the banks realized, aha, that's really stupid. What we should do is we should build a network of interoperability between all our banks and, and then charge everybody if they went to somebody else's bank. So on one hand, they turned their ATM from a cost center to a profit center because they now didn't care. If you were a TD bank and you went to bank of America, they didn't really care because they got paid for it. And oh, by the way, so did bank of America. So everybody won. So can we create an environment where the incentives are for everybody involved to promote interoperability, where they all can benefit from it and we don't have that case. Look, patients benefit from interoperability because they don't have to enter the forums twice, maybe, or they know the doctor they're seeing in a different place gets their records. Their inconvenience is relatively minor, I would like to say. I mean, in the sense that they don't have the voice. But I bet you if they were paid for using their data, boy, they would probably have a much louder voice in terms of how it would be used and push more for the interoperability. That's one. I'm an amateur economist and one of the most Important things government does is it takes care of externalities to allow a competitive marketplace to work. And here's a place where the government has not intervened to force the interoperability. ONC has done its best to try to promote it, but the tech companies have incentives not to be interoperable. Oftentimes, the healthcare providers have incentives not to be interoperable, too. So let's figure out how we can fix this. And I think we can fix it. The AI can fix it. Technically, I'm not worried about normalization or any of that. Smart people like the two of you can figure that out and other colleagues that we have. It's really about how do we get the incentives to be aligned to promote interoperability.

[B]

Barry, I really would love to see a world where we get to experiment with that. Right. A world where hospitals get paid for providing interoperable data that could be used by other hospitals would be a very different world. I do want to put a footnote on the banking example as well as the other example that always comes up. Right. The Sabre Airline Exchange. Those are great examples of industries develop interoperability models. But notice that both of those are developed in environments where there were prior intersubjective definitions of what is what you're being interoperable about. There was a contract for an airline seat or there was a amount of money, and we've already gone through a lot of institutional negotiation to say we all agree on what that means, because otherwise finance wouldn't work and airline travel wouldn't work. And by, in a way, what's, you know, my position is a little bit different, which is I agree with you, incentives matter. But I also think we're dealing with a different problem in a healthcare space where we haven't yet agreed what it means, what it is we're being interoperable about. Right. But that's a longer, definite fight we can have. And I'd love to actually bring your economics perspective. And I'd also enjoy talking about externalities and Coase's theorem. But let me. Let me start bringing this to Atlantic because I think I want our audience to have a couple of field notes from this. So if you could sit down with a patient who just got a serious diagnosis, somebody who's scared, who doesn't know where to start, what would you tell them about using AI?

[C]

Don't use AI to figure out what's wrong with you and how you should be treated, because it will make you anxious. It has a good chance of giving you misinformation. It will cloud your thinking and increase your worry and will not help you take care of yourself. If you have. Here's your diagnosis. I know you're very concerned about it. Here's what I'm going to do. I'm going to give you a sheet that's going to explain to you what is going on, what your disease is about, and what your laboratory values mean. Okay? If you'd like, after you reviewed it, you will have questions of me. Just let me know what they what. Let me know you would like to ask questions, and I will send you a list of questions that you might ask me about these different, different things. Or you can just post them and I'll prepare that other sheet for you. All of this can be done by AI without the hallucinations. And boy, when you know what's going on and you have confidence that you're in good care and people listening to you and see you, which is a common phrase of meaning listening to you, it makes you feel a lot better. Here's the thing, this is really important to me is there's no industry like healthcare. There's no place where you implicitly, without review and background checks and all those other things, you walk into a hospital, you walk into a doctor's office or a clinic, and you trust every single person that you interact with in that place, from the doctor to the janitor, is focused on you and cares about you and wants to make you well. That trust is a sacred trust between you and that organization. And 99% of all the doctors and nurses and others that I've met understand that sacred trust. And you, those organizations have a responsibility to maintain them and honor them the same way you have a responsibility to honor them when they break your trust. And I have examples of that, which, not for me, but with a loved one, I won't share here. But when they break the trust, then you know, you don't go back. But you have to work to build that and maintain that trust.

[A]

And.

[C]

And going off and looking for crazy stuff that you might find in the Internet and on AI and such doesn't really do anything to build that trust. All it does is create barriers. You don't want your doctor trying to answer questions that don't make sense with your disease. You want your doctor and nurses and others to answer questions that are pertinent to who you are and what disease you're facing that are meaningful to you.

[B]

So, Barry, let's just take 30 seconds and take that question from a clinician's point of view, if You're a clinician listening and you haven't yet dug into AI. What's the first step in becoming an AI user?

[C]

It depends on what your interest is. Let's talk about using it in a professional space. Okay. And I'd probably go right to open evidence and utilize that. And I do it two ways. One is I utilize it based upon your clinical space and ask questions relevant to your clinical space. So give it a hypothetical patient that you have and see how it responds to get confidence in whether it's thinking the way you are, it has the knowledge that you have, or is there something you can learn that you didn't know before? That's always important. The second thing, and this is just as important, use the tool as this patient information sheet, particularly if you're dealing with patients who have a chronic disease that's scary like a cancer or otherwise newly diagnosed patients use it for that purpose and give that to that patient. And that can be so comforting to them about you understand who they are and you see them and you're giving them background that they then become part of that care team, which is important that that gives them a certain level of trust. And what it does is it makes them trust you even more. And more than likely when it does ask you questions, it'll be relevant ones and it also then you're now educating the patient about being the disease. They're more likely to follow your care plan. And I'm going to tell you, if they don't already think highly of you, they'll think even more highly of you.

[B]

So Barry, I wish we had more time, but we've used it up well, well. And it was wonderful for you to share your perspectives. Folks in the audience can find [e-mail], his books are Future Healthcare 2050, Navigating the Code, and he's got a terrific LinkedIn newsletter that I hope you all will follow. Barry, it was thank you so much. It was wonderful to have you with us and we look forward to chatting with you again. And I want to thank Steve and our audience and hope you can all join us again on the next exciting episode of Practical AI and Healthcare. You.

[A]

Thank you for joining us this week on Practical AI in Healthcare. If you're ready to go beyond buzzwords and hype and explore how AI is truly transforming healthcare, stay tuned for more conversations that get us to what works. Until next time, stay pract.