How to find out what is Human.

The Internet is flooded with generative media, and this page isn't one of it. So I want to try and collect what I find out about what differs human media from generative media and the margin becomes thinner every day.

Text

Text is something that we can universally understand and it is far easier to detect the difference in a text than in an Image. But it is also easier to fake the results of human texts so the margin for error is insanely tight. I am trying to figure out how to bridge that gap.

Mathematical Topic Gap

Research showed that if we use an NLP to vector the contents of both AI content and Human content, the AI will aim for a close hit while a Human will on average be a bit further apart from the core topic. Which means if it is to close to the topic, the post is most likely produced by AI. Problem tho is how much of a distance do we need to assume that something is human made.

So this metric cannot be used alone. But it's the most reliable. You notice that reliable in this case just means that the statistics is telling us more than just guessing and that will be common in all methods of measuring how human content is. The difference between contents are getting pretty hard to figure out.

Textural Entropy

This will be something that can also be measured in images, tho in text it is easier to find a measurement because the grouping is 2 dimensional in most cases instead of N dimensional which helps a lot in figuring out those values by hand.

It's pretty similar to the topic gap and the sauce of it is that a text generated by an AI is usually either more random than a humans text or less random than a humans text. The difference here tho is that we look at just what we can calculate.

For example the length of sentences, basically the spread between the dots. A higher spread means that the sentences are varied by size and a lower spread means that sentences are basically the same length. If you just look at the dots in this post, you see that the spread looks bit more varied and that is a sign that the sentences written reflect a thought or a random pattern. So a higher spread tells us not much, but a low spread might already feel unnatural and is unnatural for a human.

So we can calculate the length of every sentence and if the entropy is either to high or to low it tells us how a text behaves without knowing the context.

Favorite words and styles

An AI will write however the prompt told them to. A human has personality. If you read a lot of my posts, you will notice that I certainly have some favorite words that I won't stop to use and lack some grammar skills. An AI usually doesn't have those flaws and will just do as told, a professional writer would as-well but we are just looking for a probability because we can never truly tell.

Images

I will go a bit more in depth at some point but we can apply the same strategies to images. I must say tho that the human eye cannot figure out the differences in the entropic behavior or the topic centering. There is one thing we humans perceive better than most AIs.

Contextual Flaws

It happens not all the times and is getting optimized out, but an AI will more often fail to draw the context of an image. What I mean by that is easy to figure out and maybe you saw it for yourself.

An example is simple. If the road leads to nowhere or into a café where a car would crash, the AI apparently didn't learn that a drive through doesn't mean that we drive into the café where people are already sitting.

WIP Notice

The topic is currently something I am heavily thinking about and I am trying to find ways that are somewhat of a metric. I also don't know how much of my ideas I openly tell the public because I don't want AI companies optimizing things out that I want to use for tools.

All in all, I hope you find it interesting what I find out and maybe you can tell me some Ideas of yours and I will try to figure them out.