It’s impossible to completely stamp out thought crime.

Also, trying to do so through law and enforcement sets a dangerous precedent.

I suspect it would be better to approach it as a public health issue.

In some countries all forms of description of underage sexual activities are illegal. So the sentance "She was having sex" is perfectly legal, but add an age marker and it is illegal. "She was having sex on the day before her 18th birthday".

It is hard to legislate around as there will always ve ways to avoid it and get around it. But all this just sounds like the normal hype => fear => hype => fear, etc cycle that all new tech goes through.

This is especially damning on the internet, because genuinely intolerable pursuits directly benefit from lesser problems being treated as equally bad. Filesharing networks work better with more users. Chasing merely distasteful people toward paranoid systems softens the reputation of those systems and makes the worst minority of traffic easier to hide.

The problem is that it's not just cartoon characters, but also realistic looking people. That makes it, especially in the next years when the techniques improve, impossible to know what is fake and what is not and thus the fake ones should also be banned. And these models are trained on images of actual abused children, which of course is the main problem with this.

Did nobody in this comment section read the video at all?

The only case mentioned by this video is a case where highschool students distributed (counterfeit) sexually explicit images of their classmates which had been generated by an AI model.

I don't know if it meets the definition of CSAM because the events depicted in the images are fictional, but the subjects are real.

These children do exist, some have doubtlessly been traumatized by this. This crime has victims.

I think a lot of people are arguing that the models which are used to generate these types of content are trained on literal CSAM. So it's like CSAM with extra steps.

In most (all?) countries no such distinction is made, the material is illegal all the same.

Prove it's fake when some of it of your daughter is making it's way around school.

You've missed the point. Fake or not it does damage to people. And eventually it won't be possible to determine if it's real or not.

Eh, if you train a ai with CSAM to make more CSAM that a different story. But in general yes.

What data is it trained on? This isn't meant to be a "gotcha" question, I'm wondering about it.

He is a deepfake of luke smith.

There's an option to speed up generation but it will make it less deterministic, like in it's 98% the same image but a little different. Also it's very hard to reproduce the same hard and software generation. That's the first issue.

The second is: I had examples of images with generation data, that I could reproduce to look 99% like the original and then just updating a single word or part of the training data (different Lora version for example) , switched the person away or their appearance changed a completely. (Imagine a picture of a street and a car is suddenly not there, or it's blue instead of red). It will make reproducibility not a reliable option. Backgrounds of images are even less reliable than the focus object.