[INTERVIEW]
Ronda Nelson: The answer was so good I almost completely missed the lie. The physiology was right, the recommendation made perfect sense, and the citations looked so legit until I checked them. And that’s when I realized the biggest risk of using AI when it comes to patient care isn’t getting a bad answer. It’s getting an answer that feels so right but is completely wrong.
Well, I am not here to tell you that AI is bad. In fact, I use it all the time when I’m writing content, when I’m organizing, or when I need a system to be developed. I use it all the time. I use AI as a thinking partner, so that I don’t have to stare at a blank piece of paper when I’m trying to write an email. It gives me ideas. It helps me think. It’s saved me so much time.
But what it’s also done is given me plenty of lies, fabricated citations. I cannot even tell you how many of those there are, oversimplified physiology, and this false confidence where I think, “Oh, I got this.” And I didn’t. And that’s the lie that’s so hard to overcome, because AI makes it sound so believable.
So the question we’re going to answer today is, what should we, you and I, hand over to AI, and what needs to stay with me? Because AI inherently isn’t making clinical practice better always, or worse. What it’s doing is amplifying our judgment, or our lack of judgment, in this case, of the person who’s using it. So when I’m using it, if I’m not asking the right questions, if I’m not providing the right context, then what I get out could be garbage, and that’s what we want to avoid, especially when it comes to healthcare.
I’ve seen doctors go in and just put in a test result, give a few little details about the patient, and say, “Hey, what does this mean?” Or just put a blood test in with no context, and it comes out saying the person’s got some crazy disease or something like that. AI loves to make things up. And it’s one thing if the fabricated answer from AI is ridiculously obvious that that’s not the case. Okay, those aren’t the hard ones to spot.
It’s the one that’s mostly right. The one when you’re asking AI to help you figure out something with a patient, you’ve got complex multiple systems happening, things that are breaking and not working, and you’re trying to figure out how do all these things work together. AI can do a great job at that, but when you give it the information and it uses terminology you’re familiar with, it follows a logical structure, it gives you a few research articles, but in the middle of all of that, AI is missing context. There’s no contraindication, meaning it’s not pushing back at you saying, “Hey, wait a minute, did you ever ask this question?” AI does not do that unless you ask it to. It might give you a study that doesn’t even support what it is you’re talking about, or make an illogical jump in the reasoning.
And if you’re not paying attention, those little gaps can actually lead you down the wrong road with a patient. And you’ll be like, “Okay, I put the test in, I asked the questions, I got it right, the AI says this, off I go.” Let me tell you, I’ve tested it, and more often than not, I’m going to get an answer that’s probably 65, 70 percent right. But that’s not good enough for me. I want 100 percent right. I want to know, on a confidence score, that this is a ten out of ten. I put in the input, and I’m getting something out that’s less than optimal.
And especially remember, AI is going to always lean, or have a bent, toward conventional or Western medicine, which is fine. There’s all this research and data, and we rely on some of that research, although sometimes I wonder how credible it is, but that’s another topic. But unless you specifically tell your AI the perspective you’re coming from, think about it like my glasses. If my perspective, my glasses, represent my functional medicine, integrative wellness, whole food, body-can-heal, don’t-like-synthetics perspective, I have to tell it that.
Because otherwise, the glasses the AI has on have just been the glasses given to it by whoever programmed it, and it’s largely going to be shaped and formed by conventional ways of thinking. Understandably so. I’m not making it wrong. I’m just saying we have to interrupt the pattern and say, “This is who I am, and this is how I want you to respond.” So if I can’t give it that context, it’s going to jump all over the place and give me information that’s more Western-based, and that’s going to lead me down a wrong path, and I don’t want that. It’s going to prioritize the things I don’t want to prioritize. I’ve seen it happen over and over again, and those are my signs. I’m like, “Okay, back it up. Start over.”
Sometimes I just get mad, and I’ll delete the whole chat and say, “I’m done. That’s it. I’m starting over. Forget it. We’re going to pretend that you’re dead to me right now. Forget it.” And I just start over, because it gets so screwed up. And that’s whether you’re using Gemini, Perplexity, Chatty, or Claude. I actually have a thing, I think Chatty, which is what I call ChatGPT, does a better job at this stuff. Perplexity also does a good job. But you have to give it, think of the glasses again, a framework or a lens to look through. It’s going to look through its 20/20 lens of Western medicine, and I don’t want that. I want it to look through my lens, which is different than everybody else’s lens.
So here’s how I use it. I’m just going to give you how I do it, because I’ve seen practitioners come to me and say, “Oh, AI gave me this answer,” and I’m looking at it going, “Yeah, but what about this, and what about this, and what about this?” And the practitioner’s like, “Oh, yeah.” Okay.
So when it comes to connections or relationships between symptoms, I will ask questions. I never ask for a diagnosis. Ever. Ever. And if you take one thing away from this show, that is your takeaway. Never ask for a diagnosis. Always ask why. What is the correlation between these symptoms? What is the mechanism by which these may be related? What are other explanations that I haven’t considered, that should be considered? Are you guys tracking with me here? You get this?
I’m not going in blindly, just letting AI do my job. My job is to think. My job is to think critically and objectively about the facts that are in front of me. And I love having AI, but I’m going to ask it, “What are the connections between these things? What are the mechanisms? What are other explanations that I should be considering that I’m not considering?” Because sometimes you don’t know what you don’t know. That’s why they call it a blind spot, because it’s blind. You can’t see it.
I’ll sometimes ask, “What am I overlooking? What pieces of information here are important that I need to know?” Let AI start to pull that stuff out for you. And then sometimes I’ll actually ask about a competitive line of thinking. I’ll say something like, “Okay, if my hypothesis is X, what would argue against that? My hypothesis is anemia, but what would be an argument against that?” I let it defend itself, and I’ll say, “Well, this is what I think, what do you think?” That’s a powerful way to leverage how you use AI in your practice as far as it has to do with healthcare. There are lots of ways to use it in marketing, in administration, in streamlining office processes, in SOPs, and all that. I’ll talk about that another time. Today we’re just talking about it in healthcare.
Then I’ll sometimes ask, “What is another system, breakdown, or condition that could also bring about this pattern? What could help me distinguish between the two?” I just want AI to broaden my thinking, because I learn that way, and I’ve learned so much. Like, I just recorded the methylene blue episode. Before I recorded that, I did not know methylene blue actually has a medical use case, for methemoglobinemia. I did not know that. I’m like, “Oh, well what is this?” And I go search it up, and I’m like, “Ah, how about that?” I didn’t even know that was a thing. So it will teach you. You’ll learn things, only when you’re curious and inquisitive.
So then I’ll have it kind of pressure test my hypothesis. I like to do that. I’ll often ask, “What are the assumptions I’m making? Where is my line of thinking weak? Is there a simpler way to explain this? Am I making this too complicated?” And then sometimes I’ll ask about safety. What medications should I be wary of or concerned with, with these types of symptoms? Or the patient is on these medications, what are the top side effects for these medications combined? I’ll let it give me that information, because it saves me time. I don’t have to go look it all up.
And then I always, always, this is a little bit of a side note, we’ll get to it in a minute, but I’ve been lately saying, asking it, to give me in any output, a confidence score of ten out of ten. Do not hallucinate. Do not make things up. Do not lie to me. Do not placate me. Do not tell me I’m beautiful. You know how it’s like, “Oh, that was lovely, you are the best, you’re so right.” I hate that. Do not tell me those things. Challenge everything I say. Be the devil’s advocate. Tell me the differentials. Tell me the reverse, the alternatives. I want to know all of that. Do not placate me. It makes me mad. It really makes me mad in life too, if someone placates me. My opinion is worth considering. Do not disregard me and say, “Oh, you’re so smart.” Don’t do that. So you want to make sure that when you’re asking the questions, you’re also, again, going back to our glasses, putting that framework in place so AI knows how to give you information back.
Now, sometimes I find that if I have a really complicated situation where there are a lot of blood tests, the temptation is to just upload all those blood tests and say, “Here you go, make sense of this, please. That’s what I need. Make sense of all of it.” That is definitely a temptation, but I don’t want you to do that.
What I want you to do first is think about building a timeline. A timeline, and you should be doing this with the patient anyway. Where were you born, where did you grow up, were you vaginal or C-section, breast or bottle, did your mom have fillings, were you on antibiotics as a child, were you ever ill as a child, were you hospitalized as a child, what was your diet like as a child, where did you go to school, did you homeschool, were you in a public school, were you in the military, did you move around, did you live in a place where there was a lot of chemicals, were you in the Midwest where there’s a lot of spraying in agriculture. You want to be asking those environmental questions, lifestyle questions, medical health history questions. So I like to build a timeline, and that helps me also build the symptom pattern in that timeline.
So here’s one of my favorite tricks with AI. Before you ever, ever, ever, ever put a test in, and that’s always our default, “Oh, here, I got a complex patient, here’s a little bit of information, and here’s the test,” you’re never going to get it right. I’m here to tell you, it’s not going to diagnose it right. It won’t. Just won’t. If you had a test with high lipids, high cholesterol, high LDL, low HDL, an ApoB positive, let’s say you put that in there. How many different reasons could there be that the cholesterol is high? There could be a whole bunch of different reasons, and it will give you some of them, but it will never give you all of them, and there could be secondary and tertiary reasons why that’s high, or the cholesterol isn’t even the problem, which is often the case. There’s other things going on.
So what I like to do is build out my timeline, build out the symptom timeline over the top of that, and then I use the little microphone on the AI, and I do a narration. I’ll say, “Okay, this patient is X number of years old, female, born, breastfed, not bottle fed, vaginal birth, antibiotics at two, three, four, five, and six for chronic recurring ear infections, three doses per year minimum, had strep throat, got into school, was home for three months with pneumonia.” I will narrate the whole case, everything that I know up to this point, medications, what supplements they’ve been on, other providers they’ve seen, things they’ve tried, because usually by the time they get to me, they’ve tried multiple things. I just tell the whole story, because now that story gives context.
Do you see the difference between just uploading a test with a few key points? Now what I get out is much more comprehensive. And I always say, “Please ask me questions, any questions you need to clarify.” This is one of my favorite parts, because Chatty, which is what I use, will come back and ask me the questions. “Well, what about this?” And I think, “Oh, dang, that’s actually a really good question, I do not know,” and I write that down. I’m going to ask them, I’m going to find out what the answer is to that. Or I’ll think, “You know, I have that, I just forgot to say it.” So I love doing the case narration.
Now, once I do that, I start asking, “Okay, given this information, what other questions should I be asking the patient? Or what are the associations, what are the mechanisms between these symptoms that could be causing this?” I’m not asking for a diagnosis, I’m just saying, “How could these connect together?” Resist, resist, resist the temptation to give Chatty your opinion first. You make AI give it back to you first.
And then I’ll say, “Now, would you please give me the research? What is your research that backs up your recommendations, your decisions, your clinical whatever?” And then I start asking for the publication. This whole process takes me, it’s pretty fast really, the narration probably takes the longest amount of dedicated time. But it doesn’t take very long until pretty soon my thinking is aligned in the right direction, because I’ve had some pushback from Chatty. Chatty’s asked me some questions about it, “Well, what about this?” And I say, “Where am I missing it? What else should I be thinking about?”
Now, when I get these research links, I go check them out. Sometimes they’re wrong. So do not get caught sending the patient, “Oh, here’s a link,” and it goes to some random dog food therapy site. I’ve had some of the most random links show up, I don’t know why. But make sure you check those links, get the citations, make sure the paper exists, make sure it matches what Chatty or the AI is talking about, and that the conclusion or recommendation AI has given you is not exaggerated based on what the conclusion in the article or paper actually says.
And once you’ve distilled all of this out, you can go back and explain it to the patient in a better way. I will often ask AI, especially if it’s complicated, “Can you explain this to me using a word picture, or a story, so that I can relate this to the patient?” Something the patient will relate to. If it’s an 80-year-old patient, the story AI gives back to me is something for, you know, an older generation, something they’ll relate to, not some trendy pop culture reference that’s relevant now. So it’s really good at matching the story or the analogy with the age of the patient. And then I usually ask it to give me a summary. Tell me the summary of what your thinking is here.
And again, I resist the temptation to give test results, or to ask for a diagnosis, until I get to this point. It’s so hard sometimes, because A, I might be in a hurry, and B, I just want an answer, because I’ve got things to do, I’ve got to move on. But using your time to really expand your thinking, letting AI help you and teach you, that is definitely safe territory for AI.
But where I draw a firm line is it cannot generate an affirmative diagnosis. I will never, ever, ever, ever, ever go with what it gives me. Now, it might be right, but that right decision is going to come from my head and my analysis of seeing the patient, asking the questions, listening to inflection and voice tone, whether they cry, is there emotion when they talk about it, what’s their breath like, what do they look like, is their skin broken out. I mean, I can narrate that, but how do you say broken out, is it cystic, is it just rashy, does it look like eczema. I’m the one that, at the end of the day, has to look at the individual and say, “I think we ought to start supporting this system.” I’m going to support your digestive system, or I think we need to support your nutrient absorption. Now, you don’t diagnose per se unless it’s within your scope, but I’ve got all of this extra information from AI, and now I can go back and look at the patient and say, “All right, this is what I think.”
AI can’t determine the clinical priority. That’s my job. So don’t ever ask it, “Tell me what’s phase one, phase two, and phase three.” No, no, no, no, no. You do not give that away. You don’t give the diagnosis away, you don’t give the clinical priority away. And definitely, this is a big one, do not ever let AI create your protocol.
I have had AI give me wrong supplement names. Make them up. Flat out makes up these supplement names. I think, “Oh my gosh, what in the world?” But it just makes them up. And so now what? I’m like, “There is not a Standard Process product named whatever.” No. And it’ll make up ingredients too. I’ll say, “Well, what is this ingredient?” Makes it up. I’ve tested this too, and this is why I’m bringing it to you, because I’ve done both sides of this.
It’ll also almost always give too many supplements, almost always too many interventions, meaning, “Do this, avoid this, take these supplements, get your bed, go to sleep, get up for first morning daylight, make sure you’re exercising every…” It gives you too long a list of lifestyle recommendations, and that’s never going to be good for anyone. And there’s a lot of theoretical stuff that gets thrown in too, junk that patients don’t understand and I don’t have the tolerance for. AI will never consider cost, or that maybe there’s a different supplement you want to use that isn’t Standard Process or MediHerb because it’s a little more expensive. Maybe you want to go with Gaia, or Doctor’s Research, or maybe you have a different brand you like. Maybe you’re going to use CellCore for something and not another, or Supreme Nutrition. Whatever you recommend, that’s under your purview. That is definitely not the job of AI.
And then, of course, it probably goes without saying, but never, ever, ever give the patient’s information. Be very careful about when you get to the point where you’re going to upload those blood labs, but remember that’s way later. Never start with it. You’re going to do that way later, after you’ve had your back and forth conversation. “What am I missing? What other angles could I be looking at? What are the questions I should be asking that I’m not asking? What are the mechanisms, how do the symptoms connect together, why could this be happening?” You’re not asking what’s wrong with the patient.
So when I get to the point where I’ve narrated the timeline, done all of that, now I might be willing to give the blood labs, but you’ve got to truncate, get rid of, black out, whatever you’re going to do. Edit the PDF if you need to. That’s how I do it. I go in and edit the PDF and take out all the fields that have identifying information, from an accession number to a date. I might leave the date of birth, but never the patient name, and not even the date of the test. I don’t leave any of it in there. I don’t want any chance.
And make sure you don’t save the file name with the patient’s name. Initials are fine, but don’t save it as something like Mary.Smith.labs, or with a date attached. You could truncate and take out all the information on the test itself, but then forget to change the file name. So just make sure you’re not giving AI any patient information. That would be a really big no-no.
And then another thing AI doesn’t know is when to refer out. Sometimes it will say prematurely, “You need to refer this out, if cholesterol is 300, that needs medical management.” Okay, fine, and maybe it does, but you do that within your scope. It doesn’t always know, because you’re looking at the patient, you know the history.
So I’m not opposed to using AI at all when it comes to patient care and getting information, but I think you want to think first about using it to generate questions, not answers. So if I had one takeaway for you today, it would be that: use AI to generate questions, not answers. It should be a point of investigation. You and AI are investigating common causes, mechanisms, plausible explanations. That’s where it has to start. Ask it to challenge you. Ask it to verify anything that could affect this, like medications, dosing, contraindications, or safety.
And then make sure that at the end of the day you ask it to give you an explanation in story form, something that helps you understand and remember, or allows the patient to have a better grasp of what’s going on. Our brains are wired for stories. That’s why movies are so addicting and wonderful, because it’s a story. And anytime we can use a story to explain something complicated to a patient, something with a lot of moving parts, just like a movie, with different plot lines happening all the time, it really is helpful for the patient.
But ultimately, at the end of the day, the final decision lands on me. I’m the one who has to make the decision, but I’m taking the information that AI helped me synthesize, that broadened my perspective, opened my eyes, helped me look at it a little differently, so now I can go back to the patient and say, “Here’s what I think is going on,” because I have a better understanding of it. And now I can use maybe a word picture, or maybe I don’t even need that, I can just explain it, and that way everybody’s on the same page. But I have not delegated or relegated my responsibility to take care of the patient within the scope of my practice, or violated any HIPAA, or taken the lazy route and asked AI to interpret a blood panel for me.
I can’t help it. Look, you said yes to this profession. I said yes to this profession. But when we said yes, that does not mean we gave up, or should relinquish, our ability to think for ourselves, because that’s what the patient needs. They need us to think for them, to be the answer giver. And if there’s a tool I can use that’ll help save me time, I’m doing it, and that’s going to be AI. Otherwise, I’d be trying to research all of this myself, and you and I both know we never have enough time to do it like that.
So AI is really very helpful. I just want you to use it the right way. That’s all.
[CLOSING]
Ronda Nelson: Even though it can give you more answers and more possibilities, just be very careful. Be very careful. If this is the type of information you love, if you want to learn more about using AI in your practice for building your practice, that’s what we talk about in Clinical Business Academy. You can find me there. Schedule a free practice strategy call. Go to rondanelson.com/cba, Clinical Business Academy. Schedule that call with me, and we’ll talk about it.
Once you have AI down for your patients, now it’s time to leverage that in your business. Just like it’s making lighter work for you clinically, we can use AI to make lighter work for you in your business, creating systems, follow-ups, email sequences, and automations, so that, just like we talked about last week, when you do get a referral, you’ve got a flow, you’ve got a process, you’ve got a way to capture and nurture those people and get them on your schedule. That’s all the things we do in Clinical Business Academy. I’d love to chat with you. No pressure. Just jump on a call with me, and let’s have a conversation. If you’re ready to stop trying to figure this out on your own, I got you, friend. Go to rondanelson.com/cba, and I will see you next week.
Be prepared, because I am on a soapbox next week. It’s going to be epic. You are not going to want to miss it. See you then.
[END]