AI Doctors Make Mistakes Eventually

And now they're putting arrows on where there are visual artifacts. So like Yeah. Look at just like that hand getting warped up. What? What's going on there? It's really bad, eh? Ah, we're finally here. All you have to do is just wait a little bit and then eventually Yeah. makes a mistake. Specifically, I'm talking about a YouTube channel that's had a lot of people scratching their heads over here.

And I'm also going to talk about a lot of things that many people seem to have not noticed yet. Now, I'm not the only one who's obsessed with this growing AI doctor problem, but I thought it was worth it to share my opinion as somebody who builds software in AI for healthcare. So, the first thing that I want you all to witness is that one, the channel does not tag any of the videos as AI and please leave a comment with your testimony as that might be necessary or important later.

Two, it does not say that AI is being used on the channel description at all. Three, in fact, the channel recommends reporting AI channels. It also does so in the community posts, and I think that's some hilarious irony. All right, so let's begin with a quick obvious question. Does a real doctor with that name exist?

Well, there is a GMC registry matching his name. He's appeared on a video during CO and you can find his name under this hospital, Bradford. So, the answer is obviously yes. My concern is will this video make the real doctor lose his license? I've been contacted by multiple NHS doctors who wish to remain anonymous and in discussing this with them, they think that it's possible but not necessarily.

It depends on multiple factors such as is the real doctor involved with that channel. Now, I personally refuse to believe that that is the case because how could an emergency doctor of all people take such a huge risk in not declaring AI? It just doesn't make sense. But then again, I also believe that people need to do what they think they need to do in order to survive.

So, it's possible, but that is just my opinion as a human. So, I doubt his involvement and I think after this video, a lot of people are going to be dying it as well. Okay, so let's play this YouTube video that was made by this channel called Brainy Medicine. Okay, so I changed the banner from How Not to Die to How Not to AI.

And now they're putting arrows on where there are visual artifacts. So like fingers disappearing. Um, ears disappearing. Yeah, a lot of finger disappearing. So, like, yeah, you see the hands disappearing over there. Like, I don't need this to be able to think that this channel is AI, but I think this does or should help a lot of people.

Yeah. Look at just like that hand getting warped up. What What's going on there? That's really bad, eh? Yeah. A lot of ears disappearing and reappearing at random times. Look at that watch just like shifting in position. Notice that it tends to happen near the edges of things. So, for example, uh when the watch is behind the hand, then the watch kind of moves abnormally.

Look at that finger appearing, disappearing again. Some more abnormal uh watch movement. So, there's a lot of these visual artifacts or aberrations in this video. Uh but let's just go with this last audio. Alpha Tokerrell. Never heard of that. Tell me how to pronounce that in the comments. Alpha Taperrell.

Never heard of that. Tell me how to pronounce that. So what the channel caught is that alphael is vitamin E. Most doctors would know about that one. You should probably ask a doctor. Just quiz them for fun, I guess, on what alphael is. They should know it's vitamin E. This was caught by the medicine YouTube channel.

So hopefully now you agree at least that there's something suspicious there. But the question is, would something like this hold up in court? Now, if I was called as a subject matter expert, I would say yes that this is likely an indication of an AI artifact, but that is just my opinion. You've got to weigh all the factors.

Uh, which again, my opinion is that it does cross that threshold for me. But maybe you know software engineers and data scientists that don't agree with me. Just feel free to discuss this with them. I have no issue. You shouldn't trust me. You should just look at the data and make a decision for yourself.

Okay. So, let's say that you think that this may be suspicious. Now, your next question is probably, is this actually on the channel? You want to see that this was posted on the channel. So, let's go ahead and watch the videos together before they get deleted. Okay. So, let's go to the channel and we'll go in the order that they were laid out in that brainy medicine video.

Uh, first is on this video right here. I'm going to skip through the to the time stamp. So, it's right before he says or the avatar says remember. That's when you see the first artifact. Okay, let's go to the next one. Uh, it's the same video, but it's at 7:01. So, Brainy Medicine just decided to focus on one video, but I'm I'm sure it's all over the place.

Well, the third thing. So, right before he says, the third thing, which is an edge on the video already going reasonably well, the third thing just the finger is disappearing. Okay. And then at let's see where it's 240. It's Let's skip to this one. Okay. fourth minutes and we're at the results page. So, right before he says over that 12, the thumb disappears a little bit.

So, there's a lot of disappearing appendages. Now, let's go to that alpha toeral part, just so you know. Um, that one is is pretty bad. So, I'm going to go to that video right here. should be around 4 months ago. February 24th. There you go. Okay. 6 minute 5th seconds. Tell me how comments. We've got flavorings which are nutmeg extract.

We've got antioxidants which are tell me how to he's just listing some things are in I think in a canva. Alpha tocaperel. Never heard of that. Tell me how to pronounce that in the comments. So yeah, alpha took a fereral vitamin E. The real doctor probably knows what vitamin E is anyways. Yeah. Oh well.

Okay. So, they're definitely on the channel. Those videos are definitely on the channel. Now, the best part is I think I actually caused this. So, a couple of months ago, I made this video saying or arguing that the channel's variance is too low or the lack of randomness was too low. And that's what made me suspicious.

And so, if you browse through the channel, you'll notice that it has increased its variance since um maybe even the perceived quality of the videos as well because that requires more variance. The beautiful part is that when you ask AI to increase variance or randomness, it actually lowers factual accuracy. This is a core principle of AI as I understand it.

And if you don't believe me, uh you could read the research for yourself right here. Uh so for this for example this paper was published by a little company uh that you might have heard of is called Nvidia but yeah let's read it together quickly. Okay so we've got the paper here it's called factuality enhanced language models for open-ended text generation.

So the first author is Naon Lee and the last author who probably supervised the work is Brian Kenzer and I'm highlighting this line here. So popular sampling algorithms in open-ended text generation can harm the factuality or accuracy due to the uniform randomness introduced at every sampling step. So what does that actually mean?

So in in MLMs, it's just basically a next token or next character generation machine. And you have this parameter called temperature or top P for example. And if you turn that up, if you increase the temperature, then you get words that are more random, more variant. It sounds more variant. It sounds better, but you increase the likelihood of introducing errors just because you're drawing from words or characters that are less likely to occur.

Vice versa, if you turn the temperature down, then you get the highest probability words, which is most likely going to be the most factual or most accurate. And so when I challenged the channel to increase its variance back in March, it increases variance, but it began to introduce more errors. This principle that as AI increases variance, it introduces more error is so powerful and I'm betting on it so hard that it's the foundation for what I'm building as a tool to help screen for YouTube channels for AI usage.

Now, right now it's just text based. For example, it reads all of the channels videos, finds out what it keeps repeating over and over. Um, so repeats are low variance and most importantly, it gives you the video links and timestamps so you can check them for yourself. Loss of calf muscle is one of the strongest predictors of frailty.

I don't want this to fall into the wrong hands. So if you want to be a user, you have to apply at the link in the description. So how effective at exposing AI is applying pressure to increase variance and therefore error rate. So when a channel increases their variance and error rate, they have two options.

Let more errors fall through to the videos or two increase their quality assurance process and therefore costs. So for example, let me show you what I mean. It's really simple to understand. Let's assume that you're trying to generate an app picture of yourself. So, I'm just going to do that right here.

If you do it once, you might get something that looks off. Maybe the jaw is different or the nose and you'll think, "Oh, AI is really bad." And you just stop trying. But if you keep trying, you'll eventually get something that looks pretty good. And you can accept that. That is just quality assurance.

So AI factories do quality assurance and applying pressure to increase their variance forces them to do more of that. It increases their costs and eventually the facade becomes too expensive to maintain and so you end up with hilariously revealing details. So, for example, this comment by M. Kyle 2, uh, who noticed that the channel says um, canola oil when everybody in Britain refers to it as rape seed oil.

And we also have that in US, but we mostly refer to the thing as canola oil. People who are in denial about the channel will probably make excuses such as, well, it's probably being made for an American audience, blah, blah, blah. And I also think that's hilarious. Unfortunately though, it's not all hilarious.

So, the channel has invested a ton in looking realistic. For example, you can see at the beginning, it looks very dolllike. And more recently, it looks so believable that it probably would have fooled me if I didn't look at the overall pattern. But can we say the same for the information that it's spreading?

Okay. Well, back in December 2025 when the channel first started, it hallucinated this Mayo Clinic cortisol study. Let's just boom. A study from the Mayo Clinic tracked thousands of professionals over several years. They measured cortisol patterns using salivory samples throughout the day. What they found was really interesting.

People with flattened cortisol curves, meaning their cortisol didn't vary much from morning to evening, were five times more. So that one isn't so bad in itself. I have other videos for that. But generally or more importantly, is the information that the channel spreads bad? Well, a lot of the videos are just basically different ways to tell you to diet and walk.

And I think most doctors will tell you to walk unless you have some kind of contraindication. Again, check with your doctor. They are the only one legally bound to care for you. Anyway, has the channel invested more in trying to reduce hallucinations? I'm not sure. I think I may have to do a deeper investigation later.

Um, but the vitamin E example that we saw earlier with the alpha tocaper, that does not bode well in my opinion. The real problems in my opinion are much bigger than the channel. For one, let's go with a hallucinated cortisol study example. It's now a part of the internet. So if you Google about it, you get that video as a valid source.

[sighs and gasps] This is really annoying to me because I grew up on the internet and now it's being littered. The second reason is to do with the truth. Whether the real doctor is involved or not. Those videos look AI generated to me and they're not being tagged as AI. Now, you might you might argue that tagging it as AI would reduce the likelihood that people walk.

And you could make the argument that in some way it's like Santa, a white lie, uh, a necessary lie to help people do better. I think it's a personal preference. Maybe you sided that. I personally prefer the truth and I am betting on people that don't like being lied to. Now the third problem in all this is a lot bigger than those two in my opinion and it has to do with AI stealing the work of humans.

Now you probably know that AI just is inspired or generates things based on what humans have in the past. I believe that's what's being done in this channel, my opinion, and it's not declared and it's plowing over real human creators that are putting in so much authentic, genuine effort to create value, new value for people like you and me. In my opinion, a lot of the things about that channel seem to have been inspired.

So, for example, the how not to die. That's an actual book. Dr. Mike Scrubs. Those are Dr. Mike Scrubs. Like the branding is just a copy from that tweaked a little bit. I don't want to live in a world where the internet is littered with AI generated information. The truth is hidden and real human creativity is just getting absolutely clapped by AI.

So luckily there is a way to prevent that from happening. Even on YouTube, if you go to a channel, there is this life tab. I'm just going to show you on mine because I know I actually stream. [gasps] And this live tab allows human doctors to go live like on on air or whatever stream and answer your questions in real time.

This is currently pretty hard for AI to generate. It can be staged and and things might change, but right now it's the best thing we have. And that's what I propose. In my way, it's the easiest way for a doctor to at least prove that they are involved with the channel, whether or not they are using AI on their videos.

So, if you want to do your part and you're a viewer, look for doctors that go live and engage with them there in the comments and they should answer you live. And if you're an ambitious doctor and you want to create value in a new way, then go live. Go do live Q&As's, but you know, make sure you cover yourself.

If we go in that direction, it should save the internet. It should get us closer to the truth and it should help humans and human doctor creators. Now, if you're a non-technical viewer and this didn't convince you, then you should discuss this with technical people uh that you trust. Now, if you like this video and you want to know how easy it is to AI generate a doctor, I actually have a video about that just to help frame things or orient you.

And I'm also sure that whatever I'm doing right now is pretty good, too.