Why AI fitness videos always look slightly wrong

We're going to go over the top six fitness videos that were stolen by people using AI. Now, I'm sure this is just the beginning. So, in this video, I'm going to show you the single science-based reason that allows you to detect AI slop and differentiate that from content that cannot be copied by AI. We're going to do that using six content creators, each of which have a video that has a unique takeaway.

By the way, if you're new to the channel, I've done health research with the world's best. And for the last six years, I have been an engineer where I make software that saves lives every single day. Health software. I work with AI and health all the time. The first video that we're going to look at is by Demi Bagby.

And the takeaway from this video is that it focuses on a unique human feature. And there is something that is not just like uniquely human. So at the second second, notice when she rolls over, the original video has a lot of camera movement. The AI video doesn't. So AI is still really bad at transitioning the point of view in human videos.

As a result, you can see that the hair flip in the third from the AI video is very unnatural. The hair does not look like it's even really in the video. Yeah, it looks like it was photoshopped or something. So at the eighth second you can see that the mouth is inconsistent when it has to move. And the reason for this is AI is not great at copying data that's out of the normal distribution.

So what I mean by that statistically normal distribution has neutral faces in the center. So Demi Bagab's face in that second is far from what is considered neutral. It's very expressive. With that said, it is still possible to copy that if the person that used the AI had anything close to a functioning brain.

So, in general, you want to look for something even more nuanced than the facial expression. And that's exactly what we can get out of the next video. So, here you can notice that the AI fails to replicate the muscular contraction that is necessary for the movement that this next creator is making. So in the original video, the upper back muscles contract very precisely to control the motion in the way that it needs to go.

The AI has no muscular contraction. And this is again because that contraction is out of the normal distribution. You don't often see the back muscles contracting that strongly. So the takeaway from the second video is look for muscular movement. Muscular movement is really hard for AI to replicate. at least one that is accurate, especially one that is fast and as violent as that muscular movement is.

All right, so in this next one, hopefully everybody knows this one by now. If you don't, you should definitely see it. The creator already identified it. And you can see here that the finger disappears for a bit. And the reason why it actually did that is because the finger went into the shadows. So here AI is struggling with darkness.

And that's because darkness is out of the normal distribution of videos. Most videos are well lit uh in the center. Whereas if you go dark or super bright, those are highly variant. Basically, there aren't many videos filmed in the dark. So AI currently struggles to have enough information to be able to replicate that.

And that take away actually becomes necessary and key to detect the AI in this next video. So if I didn't have shadows or darkness on the mind, I probably would have not noticed it. So the first time I looked at this video, I thought that the AI user nailed it. But then I was looking for darkness and I realized that the original video had the shadow of the bar that she was holding.

So I immediately I looked for the shadow in the AI clones video and it wasn't a straight line of the bar like in the original. So the takeaway for this video reinforces darkness shadows. Look for those and if it seems abnormal it's probably AI and if you add a lot of darkness or shadows to your videos then it's going to be harder for AI to copy it or steal it.

So you can see that for the next content creator. Same thing when she's throwing away that weight, the shadow, you don't see the shadow. So, the AI1 doesn't have a shadow. And another reason why it's so hard to copy the shadow in this one is because the weight is being thrown really quickly. So, that is an element of variance.

It's not normal for things to be moving that quickly in videos. But this video also has an important unique takeaway and that's actually very close to the important unique takeaway from this creator's videos. So here we look for a place of quick and large variance and we find that the original person at the 7second flipped a pancake and the pancake went up in the air and you can see their eyes following the pancake and it all makes sense.

A lot of going on in that that has to be consistent. And now if you look at the AI when it tries to do the same thing, it completely fails it. I don't believe that anybody would have fallen for that. Hopefully not, but you know, you could be surprised. For this last content creators videos, the takeaway can be extended also to long form videos.

So the original video didn't have a large variance on the person unless you look really closely. So, because she's running so violently, so aggressively, so quickly, when her feet hit the floor, you can see the impact go up her face. You don't see that in AI, and you certainly don't see that in the center of a normal distribution curve.

It's really outside on the highly variant part of the curve. But that one was really hard to notice. There is another thing. So most of the variance in this video is not in the person, it's in the overhead shadow at the third second and the fact that there are other people in the video moving. So those two are helpful enough.

The AI has even lower variance. There are no overhead shadows or other people in the video. But to summarize all these, humans can do more facial expressions. Humans have natural muscle movement. AI struggles with dark spots, shadows, etc. AI struggles with motion. And the fifth one is that AI can't add too many characters without messing them up.

And that's why you don't see other people in the video with the runners. Okay, so you might be thinking that AI struggles with physics in general, or you might have thought that before the video. Yes, it still does, but in a lot of these videos, the user used what is called or known as character replacement.

So basically you can just have the real person do the movement and frame for frame or picture for a picture swap the real person for a fake AI generated person or another person. That is much harder to do still if the video itself has a lot of variance. I hope you know what that is by now. So my advice for creators is to maximize variance.

So add those elements that make it hard to replicate for now. You know, eventually AI is going to get more and more data and maybe different models will allow it to replicate things, but I think we are a ways away from that. All right. So, the other thing that you might have noticed is that all of those videos were by female creators.

I didn't do that intentionally. I couldn't find a single video where a male's video was stolen by AI. And I think the reason for that is people who are using AI maliciously, they are trying to target people who look for female-based content. So yeah, that sucks. But you know, I just gave you the tools to be able to fight back.

I hope you could use them. It's only going to get harder.