When Should You Trust a Machine With Your Life?

0:00 In recent studies, AI diagnosed patients correctly 77 % of the time, better than human doctors. but should you trust your health to an algorithm? Today we're joined by one of our oldest friends who happens to be an actual working doctor, Dr. Ajit Beren We'll talk about why AI is not ready to be trusted to offer medical advice, how insurance companies could use it against you, and why some conditions only human doctors can spot. And he shares a mind blowing story about a mysterious illness no algorithm could have possibly diagnosed correctly, proving why we still need human expertise in medicine.

0:45 Okay, this is the fit mess. Thanks so much for listening. We talk about all things sort of AI and wellness and mental health related. And today we're joined by an actual doctor with an actual knowledge about computers and stuff and know Ajit that doesn't make you better than me. I think you were born with like a like a syringe and a computer in your hands, right? Yeah, exactly. I just don't want anyone calling me thinking I'm their support after the podcast. Because it... You can call me, give me... Just, you know, please leave a decent review after a brief survey for me. Yeah.

1:26 Jason's here as well. That's right. That's right. All right. So, G, this is something Jason and I have been talking about a little bit here the last few weeks in the show is just how much AI is being integrated into sort of all things health and wellness. Again, you're an actual doctor, you you are on the front lines of this stuff. How much of this is a part of what you do when it comes to treating the patients that you work with? I work right now these days in a very complimentary and alternative limb, but I still have to do standardized intakes that require some knowledge. Now, usually what I see out of AI is that people are looking for a diagnosis or looking for a differential or something like that.

2:11 So it's kind of like they're done with WebMD, they're done with just a generic Google search. now they want this and this and what's funny is that Google has trained us to see the very top first. It's how they do their ads, everything. So I think that when I search for something, I hate seeing an AI generated response because it reiterates itself 10 times if you the entire thing. And it's not that knowledgeable. It doesn't give the differential for the patient. um It still needs to be worked on. It's scary enough where it could get really good if it starts asking questions back.

2:53 You know? Like, what type are you? well that case, you know, you may have this. You know? Yeah, so along those lines, mean, if people are using AI to try to understand their health care and context, and it becomes this choose your own adventure piece, would AI actually with better prompts inside of it actually asking you questions more effectively to try to explore deeper on a topic? Is that something as a medical professional you think we should be looking at and that patient should actually be paying attention to? Or is it still too nascent and too early at this point to really hope that it gets the prompt engineering stuff?

3:36 That's a great question. I think it's too early. And is that because the amount of information that's in the system itself is not actually sanitized and produce results in a way that's actually reliably of value and results in human beings kind of having the Denning Kruger effect where they think I know shit that I really don't and apply practical medicine advice to their own physical bodies and essentially start to biohack on themselves with limited controls. Yeah. For the most part, what physicians do is before we do blood tests or do any kind of scans and all that, you have experience.

4:21 That experience kind of lets you look at a patient and say, I can clearly see the reason your face is hurting is because you have a bullet in it right now. So it's not like my face hurts. So we can actually do this. pretty obvious what the diagnosis is going to be. Yeah, but an AI generated response, mean, that thing isn't even giving blood tests. Now again, if there's an interaction where it's like, hey, do you have a copy of your latest blood report? And when was it? And things like that. If it becomes more interactive, it can definitely be, definitely something that could come pretty damn close. Yeah, so...

5:05 You and I have a personal relationship where you help me with my health in a meaningful way, which is a big reason why we've invited you on the show, because we both interacted with you on a personal level and on a business level in different degrees, but also at a medical level. And because of that, some of the things that you told me to do were things that my naturopath, who I really like and really value his opinion, he did not quite go through and ask the same, ask the same type of data to be collected. that you asked for when you were going through trying to fix kind my metabolic health issues. Is it that the AI systems themselves are trained on classic Western styles of data and information collection that's much more compliant with trying to reduce the cost of overall health care to try to fit into this model that insurance companies, private payer, and these socialized medicine systems need to actually kind of narrow these costings.

6:01 They don't ask for some of these other tests? Do you think it's more that the models themselves have kind of pigeonholed themselves in this way because they've actually been trained to try to minimize the amount of things they ask? Or do you think it's more that they just don't know right now and they're an unreliable arbiter of information? And I'm going to specifically talk about testosterone because you actually asked me to test multiple different types of testosterone to go through to understand my testosterone, my free testosterone. my total available testosterone and then my albumin levels were part of that as well. And my naturopath looked at me and goes, yeah, we totally should test this. Like he didn't even think twice as another medical professional.

6:42 And now it's part of my regular testing routine. And as a medical professional, he should have asked, he probably should have asked that also, but he wasn't doing that because he was focusing on the pieces. And you just have these different approaches. When I think about AI, Is AI another approach or is it an accumulation of minimal data sets that trying to treat symptomatic problems if they exist today? God, that was long-winded, No, I don't think AI should treat or diagnose anything at the moment. Okay, at all. Also, I don't know your natural path, but you it almost sounded like it could have been Dr. Nick Riviera, right? It's just like, that's, that's a great idea, man. Let's, let's do that, you know, but, um but I respect natural paths, in many cases, more than I do a lot of MDS.

7:30 So, um because they're they're using the good stuff that allopathic physicians sometimes ignore, you know, So that's kind of actually where I'm kind of, you know, I think I'm a little different with things because I consider myself an allopath as well as a naturopath, you know. You guys have heard my terrible analogy or, know, like, okay, I'm gonna do it again. You know, if you got the cold, drink some oranges, get some rest. That's naturopathic. But if you have syphilis, man, I'm gonna give you some penicillin. That's allopathic, okay? So that's where... you know, it's too early for anything.

8:10 Again, if there's a good interaction, and in fact, there are great interactions that already exist, like when you go to like women's health websites or men's health websites, you know, which are predominantly like, dude, I need to get horny and I'm bald. Okay, so it's like one of those things, okay? And the same thing for women, and those are actually, most of it is done just through like a wizard. you know, just like, you know, 30 questions, upload an ID, and then in some cases, a brief talk with a physician to say, yeah, dude, I think this could help you, okay? uh Or just an email or something. I mean, it's gotten that simple. So uh those examples exist where you can actually, you know, get a good solid sort of aggregate, okay?

8:57 And then the physician will verify that. Okay, so. I don't think that AI right now is at that level yet. I think it's sending it, it could be dangerous too, very dangerous. OK. I have a much dumber question for you. So, I mean, you guys know this stuff better than I do. And that's part of why I'm here is I'm trying to understand it and incorporate it into my life in ways that are not going to, you know. send me down the wrong path. But I mean, my understanding of AI is that it's basically scraping the internet for a whole bunch of information and connecting a bunch of dots and regurgitating it back to you. So instead of you having to read 10 websites, it read 10 of them for you and wrote a really stupid paragraph and left out a lot of key information that you would have gathered had you, you know, scrolled down 10 spots.

9:40 So I mean, if my understanding of it is correct, if I'm anywhere near close to that, not much has changed like we're still Googling, you know, how come there's blood running out of my butthole or whatever's going on and then we're finding the results and putting them all together and going, my God, I have whatever terrible disease it just diagnosed me with. And then we run screaming to the doctor to verify it. If we're still using it in that way, I don't think it's necessarily dangerous unless it's the person who, you know, finds the result that's like inject bleach into your body and it's going to cure it because that's what Donald Trump said five years ago. Is that kind of the danger that you worry about is sort of that bad information? Yeah, um here's what I think AI is.

10:23 AI is kind of like, you know, why trouble yourself and read this entire page? Let AI sum it up for you real quick. Okay, so it's like, it's making us dumber and that I think is dangerous, you know? Like I love watching this one scene out of uh Parks and Recreation where Chris Pratt is trying to figure out why Leslie Knope is sick. and he's on his computer and he says, think I know what you have. You might have a network connectivity problem. Such a great scene. you know, it's dangerous. It's a good place to get good ideas. But see, for someone who's like, I used to see this all the time, like people who are trying to avoid going to the hospital, for example, when they absolutely need to go to the hospital, they will Google whatever they can to give them a better sense of assurance.

11:17 Like, my God, this sneeze is just nothing, okay? I don't have brain cancer, you know? um That's where it can be dangerous. We want people to go get checked out. We want people to do that. But it's also good to have an understanding though as well. Hey, what are the treatment options now since I've been diagnosed? And in that regard, it's really helpful so I can say, well, what are my options? Who is against chemotherapy? Who's against this? So those things, obviously just like anything, just like the internet from the inception, is good and bad. Helpful, not helpful. What's your take on, we talked about this a few episodes ago, there was a study done about, basically they set up these trials where patients were walking into doctor's office X, interacting with a robot, answering all the questions, and then going and seeing a real doctor, answering all their questions.

12:09 And supposedly, according to the results, the robots got it right 77 % of the time, the actual human doctors got it right 66 % of the time. Man, that isn't a Theranos uh type of, uh no, man, is that really a thing of the robot? What did the robot diagnose the person with? Was it something? Okay. don't have it up. should have actually pulled it up in front of me. Maybe I could get AI to bring that up for me. em would love to read that. All I can say is... if like the things that people go to the doctor for are so common. Like, like, when when you were in a situation seeing patients sort of all day long, is it that like, is it a lot of people that all get the flu at the same time, and they all come knocking on the door going, I don't know what's going what's wrong with me.

12:59 I feel terrible. Well, guess what everyone has the flu right now. Like it is that that does something like that factor into this. Uh, yeah, it does seasonal stuff, obviously where your location is. So it's like, you know, winter time, you know that flu, you know, cases are going to rise. So yeah, there's seasonal, environmental, and regional variations that help everyday primary care physicians come to a better answer of what this could In addition to blood tests and things like that, if one needs to get that done. So when we were talking, just before we hit record, we were talking about sort of how much you interact with AI and you mentioned that there are times when, you know, you'll use it and search things up, but then have to like verify the facts or you know, the information that's been given to you.

13:43 Like, what is that? What has your experience been with that? How much of it are you seeing is just like completely just wrong? Do you do you find most of the time you're like, damn, that's spot on? Like, what's your experience been? it's definitely spot on, okay? Sometimes, but a lot of times it's a very broad, uh it's a broad answer and you need to kind of dig into the specifics of what it could be. um A great example is like an upper respiratory infection. That could be a... you know, a viral infection, could be bacterial infection, it could be an allergy, you know? And so that is where AI can't help you too much.

14:28 You you would need to have that trained eye. But for the most part, you know, it should be able to give you a few varying factors of what it does. In fact, I think I saw something one time that was something like a cough, like it could be a cough, and then I think it cascaded down to uh just something terrible. uh But it gave all the key, again, I forgot what it actually was. I think it was a type of cancer, but it gave varying answers. But I think all of it should be coupled with, if you're looking for a diagnosis, just use this as a stepping stone to go to the hospital, to go to your physician and get checked out. You know what helps me think clearly enough to tell the difference between good and questionable health advice?

15:11 Magic Mind. And their promise is one you can trust. They only use the world's best ingredients. test every component rigorously and inspect every bottle by hand. And unlike most companies that make you mail back a half empty bottle for a refund, MagicMind offers a no questions asked 100 % money back guarantee. No need to return the bottles, though why would you want to? This stuff actually works. Get your mental performance shot at magicmind.com forward slash fitmas20 and use the code fitmas20 for 20 % off because your brain deserves better than WebMD and robot doctors. So I am doing another search at the same time that we're talking to try to look at the way that AI spits back data sets.

15:58 And part of what I am seeing on a consistent basis is that AI goes through and looks at, I can prompt engineer the hell out of my data sets. And I can give it all kinds of parameterization and I can lock all kinds of component pieces in and I'm trying to compare it. with some of the data sets that I have and some of the information that have. like, you brought up a really good point, you know, here's my lab data. ChatGPT, looking at my lab data, what can you tell me? Format not compatible. That's what it tells me because I'm uploading a PDF of my blood work and it's like, ah, I can't really tell because this might be this and this might be that.

16:43 Form data gets confused, field data gets mixed up, and suddenly what looks like a normal level for one thing is slightly in the wrong spot. So this is a data indexing problem. The exact same problem that database engineers have. trying to run these things through. Now it's not that I don't think the data itself is of value and available and it can make these pieces work, but I think like you said, Ajit, a human eye looking at this thing would have seen my data and gone, there's the problem. And I could look at my data because I've done this long enough and I'm not trained, but I guess I'm tuned to look at stuff at this point. And I can go, there's the problem. But the AI can't.

17:23 And this has to do, I think, with the fact the models are still learning. They're still trying to understand these things in context. Now, that being said, if I take a specific data field in there and I say, here are my albumin levels. And here's what it says normal and what abnormal and all these different things are. And I say, me information on it. It's really good at that. But I can also see that it's querying the internet and going out there and compiling these things across multiple different sources. And I can go through and I can look at that. And on top of that, I can go through and I can click on the link to the data source itself. And it actually says, this link created for AI generated content.

18:05 I think what we're seeing is that the companies that do the smart things already are already trying to take this into consideration and lock these things into place to try to make them better over time. And it doesn't supplant the physician piece, but I think it's a clue and an Easter egg that we keep following it. It's going to eventually reveal that. We don't know what we're doing at scale and in mass because these guys keep saying, I'm making this stuff to answer the questions for generative AI. At what point do physicians have to create their own bank of information that's not generated by AI, for AI to spit back responses?

18:46 Because as far as I can tell, the internet seems to keep getting swept up in all these kinds of weird motions and we keep deleting information and things and a lot of the index information that was available for like NIH is suddenly gone because we have a new administration that's come in and anytime the word race or equity or anything like that shows up they delete it which are all terms that are used for good reasons and then AI is further convoluting that data by crunching those pieces going across. How are physicians going to be able to adjust and work in this space with so much inconsistency in the Well, don't think physicians, so long as there's human physicians, will have to deal with that too much.

19:30 I think they're going to be dealing more with the really cool technology that's coming out in the work of uh blood chemistries, urine analysis, scans of the body. uh That stuff is getting so amazing. That's the stuff that's going to give you your definitive answers. And I think that's what they're going to be leaning towards. um I think part of AI learning is that it's going to understand certain population groups too. And that might be based on your IP address, Tell me more about that. How does my IP address play into that? Like, let's say you, me, and Jason have an IP address and it's located, you know, in Yemen. They're gonna be like, yeah dude, those guys are probably gonna die from terrorism, okay?

20:11 You know, something like that. Whereas like, you know, just think of an IP address as being something regional, short of it being firewalled and like, know, VPN'd and all that, okay? But, you know, you could, like, just like how you use Yelp, right? Like, you know, you're over your phone, If you have your settings right, it already knows your location. So whether it be location, GPS, or IP address, as long as there's no type of mirroring or safety mechanism in place, you could pretty much get an idea of what it's like. I'm sure it could see what kind of metabolic cases are in certain areas by certain hospitals, how many dialysis centers are in one area and stuff.

20:53 So there's a lot of good predictability. You know, like I think there was something, I'm sure you guys have read something like this, areas that have a Whole Foods market tend to be a little bit healthier than places that just have a Piggly Wiggly or something like that, you know? Right. Yeah, you bring up a good point. mean, the idea of understanding geolocation based upon these health indexes, but not just that. Like if I'm using a browser and I'm using like chat GPT to go through and do this information, I go through and actually tell it use history or not use history. So if I tell it to not use history, it's going to try to do everything from scratch, assuming I'm somebody brand new. If I tell it to use history, it'll use contextual clues to try to figure out who I am, including things like browser ID, log in information.

21:37 location of other devices that are running this application set. Like it tries to track all those pieces and pull them down and it uses inference to try to make sense of it in context. The same way humans do in their own local geographies. That being said, if I'm a physician, how soon do you think it's going to be before physicians start taking this type of metadata that's produced by apps and produced by other interactive tools? and incorporated that into their practice. Like, I mean, you and I work together on a uh fitness product that takes information from smartwatches, and it can include lat-long information, geolocation, movement, all those pieces.

22:19 But doctors aren't really using that data today because they think it's kind of novel in the way that it's approaching those pieces, or maybe they don't trust the source of it. But it's clear that we're machining things up a lot more. And the only way that you can actually handle the volume of data that's coming from the telemetry perspective is to have these tools in place to be able to make sense of it. Do you think doctors are gonna be forced to start taking these types of data sets into almost into practice and into consideration when doing diagnosis and understanding their patients or are we still just in that novel phase where, you know, we're just got to wait and see. I think we're in the novel phase, it's uh the first part.

22:59 I definitely see something like that happening as not necessarily a way to benefit the patient, but rather to keep these asset houses that have been purchasing systems uh essentially cost down. So I think it's going to be kind of like a very automated chat kind of like thing. You know, when you call customer service for like an airline or something. you the first thing that they say is, you know, hey, you know, you can find us on the web. And I'm like, yeah, dude, I know I'm calling you for a different reason right now. Okay. And so with that, I think I think that physicians will use it, they will use it, but they may actually also be forced to use it in the style that the overarching corporation would like them to use it.

23:52 And that could redefine litigation, that could redefine malpractice. In other words, I can see a world where something today you could be sued for, and 10 years from now they're like, no, we've rewritten that, that's not a suable thing. I can't really give an example, just to add off the top of my head, but well, no, no, I can't. mean, there are certain things that you can see in insurance that are paid for, partially paid for, and just not covered whatsoever. And I think that would probably be the lean that where insurance companies and hospitals may start working together and using AI as an aggregate. Because the thing is that if they use that as a tool to minimize costs and give, I don't know, okay, well, That is where the consumer can go Google themselves and get the same pretty much the same answer as well, which is also right.

24:47 So it's like what my physician is now bound to basically to do for me, which you know what they're telling him or her to do is the same information I'm able to Google. So it must be right. I you're you're both touching on something that the more Jason and I talk about this stuff, the more terrified I get about the power that it wields. And just the connection of this information to the insurance companies who are denying stuff that they should absolutely not be denying, you know, every single day to have some sort of sharing of this information that I think will just continue to disqualify more and more of these costs because the insurance companies their sole purpose is to stop finding ways to pay these things.

25:29 And I think this is only going to make it easier for them. And so in a system where you know, it isn't socialized medicine, it is very much a uh massive for profit institution. ah It's terrifying. how far this could go and really harming people's lives when it comes to this particularly in more ways than it actually will benefit them, I think. I would like to switch topics slightly. Okay. AI for mental health. Using AI as a counselor or a friendship or a relationship as a physician.

26:10 What do you think of that? Dude, you all saw 2001, man. Okay, that's what I think of that. Okay. I'm sorry, Dave. I'm afraid you won't get your antibiotics, you know. Hahaha You know, um that's a Again, I can see a collaborative approach with AI, where it's not just you type in one thing, it's like, okay, here's a great example. Whenever anyone Googles the best way to kill themself, the first thing that pops up is there is help for you, there's suicide prevention lines, it's a first, they're not answering your question.

26:52 And so I think that's gonna be, that's kind of like what I can see happening. there will be fail-safes in place. Or at least the ethical, you know, internet should, for stuff like that. In the same manner of suicide prevention. It's remarkable how much this is coming up. We just talked about this a few weeks ago and how, you know, when we started talking about this in medicine, I was like, I'm, I would be curious about it as a mental health thing. And so I've actually used it as sort of a stopgap therapist, a few times just like, feeling a little stuck, you know, like ruminating on something, and I'll just kind of dump it out there and it asked incredibly profound questions and offered advice that like, a typical human therapist might not right, like I was, I was like really hanging on to a stupid softball thing related to my kids game.

27:38 And like a normal therapist wouldn't necessarily go, Oh, well, here are some drills that you can work on with your kid that might help her be better and help you feel like you've contributed, right? Like, my therapist would not go, Oh, I know these five drills, you should totally try these, these would help they would go, hey, Google some stuff to get better, right. uh So like, for me, I've we've talked about using it as a bridge between your actual appointment with your actual therapist, you know, when when you just need somewhere to dump those thoughts and have someone sort of thoughtfully reflect back your experience. And and for me, it's been helpful. And I was just reading again this morning more articles from people that are like, my God, I tried this as a therapist, and I was terrified at the results because it was so good. I think this is a thing that is becoming more and more popular because more and more people need help and don't know where else to go.

28:24 And as we create these these relationships, with this artificial intelligence, it's becoming kind of a natural go-to that like, I've got a question. I wonder if this can answer that. Yeah. In regard. speaks to what you're saying is like that collaborative approach, like there there can be, you know, in combination with a therapist, especially if you take those threads, right, if you save them, bring them into therapy and go, here's, here's kind of where my head was at. And here's the questions that I got back and what I did with them. What do you think of that? And the therapist can go, you know, that was really terrible advice, you should have done this, or that's not bad, right? Yeah, yeah. You know, I think these models, what's funny is that if, if, if let's say something totally automated happens in the world of mental health, okay, where the consumer is better.

29:15 I mean, you can't argue that it may be terrible behavioral, you know, therapy. But if the person says, you know, well, you know, I don't want to kill myself anymore. Or you know what, I think elementary school kids should live. You know, like, you they, they, you know, I think that's a positive thing, you know, but there's no way of me validating that. That's a scary thing, you know, you because already too many improper uh prescriptions for uh antidepressants and, you know, anti-psychotics are prescribed too many sleep medications are prescribing. It's. You know, I. I think that could be a really scary thing, you know, but again.

29:58 if it's collaborative, if there is a human that can just check off like, that makes sense. I see validity. Medicine is just one of those things where you can automate almost anything, right? And uh medicine is all right too, but there's nothing that's gonna take away that actual clinical and physical diagnosis. Mm-hmm. a couple of years ago here at UCSF and a girl had come back from somewhere in Asia. And she was experiencing cramping, urethral discharge, something, uh she got checked, had an STD panel, had a cancer panel, but nothing was turning up at all.

30:48 Mm. But there was one physician there who happened to be Chinese and was from the area that this girl had come back from. And was like, I'm gonna get you tested for something else. I'm curious. And sure enough, uh it uh was an STD, but it was a thing. It was almost like a parasite, or like a bug that only in that one region of the world gets transmitted. So why the hell would we be testing for that here? And why, if you type stuff in, You know, you would get a completely different, you know, an incorrect response. You know, the AI would probably say, oh, maybe you've got a dermoid cyst going on in your pelvic region or some shit like that. And and when in fact, like, no, dude, it's some crazy South Asian bug that crawled up someone's dick.

31:32 OK, so like that's that's really what it was. You could Google this. I swear it's it's it's the world is full of insane things. So can you imagine if I said that actually you might have an insect that crawled up your dick? OK. I would fire my real doctor and hire that robot doctor immediately because that robot doctor is great. Yeah. I don't disagree. there anything else Jason you want to talk about here before we wrap this one up? Sure. need to do a couple of episodes and a few rounds of this because I think that there's more to learn and more to discover. I really, really am interested in hearing Ajit's perspective uh with some of the things that we've found over time with our interactions with AI and its medical space.

32:21 And yeah, this has been great. My perspectives only stand out due to my analogies. Okay? The analogies are gold. They're gold, absolutely. ah Ajit, if somebody wants to connect with you, get your medical advice, whatever you do, I always hesitate to ask you where people can contact you, because you do 4,000 things. So if you want someone to be able to contact you or follow you or any of the work you do, the floor is yours. Well, I don't have a professional Instagram. My wife does, but I've got my own Instagram. I guess for right now, uh I assume we'll do another episode, which we'll have more time now since I know how the routine is.

33:02 I mean, my Instagram is macdaddyachit, but I don't think that would be, I don't think I should get DMs about that thing saying, hey dude, I think I got colorectal cancer. What do you think? Yeah, I think something crawled up my dick, dude. Or like, think something crawled up my husband's dick, because he's a total asshole right now, you know? Well, then you got the diagnosis, okay? So, yeah, yeah. I'll get something set up, for sure. Something professional, ugh. Ajit, thank you so much for your time. of course, will do this again because we have 4,000 more questions, both dumb and smart. So we'll get to all of them next time. love it. Thanks for having me. I love seeing you guys All right, that's all the time we have for today.

33:45 I want to give a huge thanks to our friend, Dr. Ajit Baran Dhillon, for sharing his insights on AI and healthcare, his perspective, and those analogies. Always absolutely gold. about how AI is changing medicine and mental health support. Before we go, I want to remind you to check out Magic Mind if you're looking for a mental performance booster that actually works. Magic Mind has been a game changer for me. It helps sharpen my focus, boost my energy and reduce stress without the crash that I expect from coffee or other energy drinks. Head over to magicmind.com forward slash fitmas20. Use the promo code fitmas20 to get 20 % off. Again, magicmind.com promo code fitmas20. And finally, if you enjoyed today's conversation, please hit subscribe, leave us a review, and share this episode with someone who might benefit from it.

34:31 Thanks again. We'll see you next week at thefitmess.com.