Why AI fitness videos always look slightly wrong

AI keeps cloning fitness videos on social media. The three most welldocumented cases are all female and I tried to find a single male I could not. This video isn't going to be a list of AI tells or videos. It's more going to be about the engineering. So you'll be able to understand the AI fingerprint and why most of these videos are women videos.

My name is Nas. I built AI software in healthcare. Let's go through those three cases to understand the AI fingerprint and the system underneath it. So the first video is by Demi Bagby. In the original, when she rolls over, the camera moves with her. In the AI version, the camera doesn't move. So that's already off.

And then the hair flip is next. In the original, the hair has a weight of momentum and it looks natural. In the AI version, it kind of just looks pasted in or photoshopped. And then at this point when Demi has her mouth shaped in a funny way, the AI does not do the same. And here's why. AI generates by averaging when it's mostly seen.

The training data is slow, calm movement, hair that doesn't flip over that quickly, and neutral faces. So when real life gives us something rare like Demiy's facial expression changes and the camera motion, AI really struggles to replicate that. As engineers and data scientists, we call that low variance.

And that's just how much data differs from the average. So low variance is the fingerprint. Low variance means probably AI because it's in that average point. When you have a moment that's really awesome or rare or weird, AI version of that moment looks polished and dead. The second creator, the variance is not in the face.

So the next creator is Natalia Barldi. In the original Natalyia Barldi, her upper back muscles fire hard and precisely and the contraction has to happen very synchronously with the movement for the movement to work. In the AI version, the movement happens, but the muscles don't contract. It kind of looks like a dead muscle.

That kind of rare and precise muscular contraction is rare in fitness data, especially rare in general data. Most fitness videos do not show synchronized contractions cleanly. So, an AI generates a similar movement, it doesn't have enough examples to copy from. The muscles don't show up. And it's not that AI is blind to muscles or anything.

It's just that so far nobody has trained it on that data. Now, the third creator here is the hardest case. The variance is not in the face or muscle. It's in her environment. The third creator's name is Ninja Yari. If you look at the original, her video, she's got a shadow, and the way she's throwing also has a shadow as it flies about the video.

Now, in the AI version, the person still has a shadow, but the weight does not have a shadow. It's just an object hanging in the air with nothing under it. or the shadow is blurred out. So, at that point, it's not really a workout video anymore. It's more a magic trick of how the shadows disappear. So, the variance in that video came from two sources.

One is a shadow, and the other is the fact that it's a fastmoving object. Both of those combined are especially rare. So far, we've seen three creators, but three different sources of variance. The face, the muscles, and the environment. But again, the same fingerprint each time. the variance. So, here's the rule.

AI generates the same average thing every time. So, if you have a rare face, a fastmoving muscle, or shadow that's jumping around, that's probably a sign that it's not AI. The fingerprint is low variance. You will not see those in an AI generated video. The fingerprint is low variance, the absence of high variance features like those.

So, if you ever find yourself looking at a fitness video and wondering, look for the highest variance moment and see if it's been flattened. If it's been flattened, then it's probably touched or generated by AI. But here's the most disturbing part here. I've covered three videos. In another video, I've done six, and they've all been women or females.

Unfortunately, that's not a coincidence. You might think the AI prefers women or that it's a small sample and a coincidence, but that's not the case. The malicious operators actually engineered it this way. Female content gets way more views, female fitness content, than male fitness content. And so, it was a question of maximizing those views for the engineer.

They weren't just guessing, right? There is data behind this. I use this Vid IQ thing. And if you look at women's workout, there's 19,000 searches a month for that. Men's workout, 9,700. About twice as many. Female fitness, again, about 19,000. Male fitness, 5,000. Almost four times. So, look at this flow diagram.

People watching things on YouTube or social media flow at two or four times the rate into female fitness content than they do in male fitness content. Once they're in, the algorithm shows them more of the same content and the pool only gets bigger. The operators didn't pick women per se. They picked the pool that grows the fastest and AI is just the cheapest way to fill it.

So, you've been watching an engineered system. On one side, it's leveraging psychology and therefore using female-based content and the other side AI is landing out the abnormal or out of variance features. So, what was engineered is not the fact that it's women, but the fact that they want more views.

Again, the test is look for the rarest moment in the video, and if it's flattened, it's probably AI. If you're a female fitness creator, you're probably a target for being cloned. I recommend adding shadows or stuff to make your video highly variant and therefore harder to copy. AI factories are designed to flood your feed with email lists.

It's a bit more protected. So, if you want to sign up for my early access email list, I'll put a link in the description. Thank for watching.