“CSAM generated by AI is still CSAM,” DOJ says after rare arrest

jeffw@lemmy.worldmod to News@lemmy.world – 291 points –
“CSAM generated by AI is still CSAM,” DOJ says after rare arrest
arstechnica.com
216

You are viewing a single comment

No, that is quite literally not how it works. Insisting that that's how it works doesn't make it true.

This isn't secret data or a hidden process. You can just read how it works. It does not work how you are insisting it works. It was literally designed from the ground up to work in the way that you are insisting it doesn't.

Because I do not want to type gross things, let's discuss sauron, from the Lord of the rings.
He is never depicted without his armor. He is referred to only with male pronouns.

Lord of the rings, sauron

Entering a prompt into my local stable diffusion instance of Lord of the rings, sauron, naked Negative prompt: Armor returns an image of a naked figure with a predominantly male body and musculature, no pubic hair, vulva, and bony protrusions from the head. (The crown seems to have been converted into organic material).
Since it has no training on the specific notion of a nude sauron, it extrapolates from what it knows about the shape of sauron and typical words used with sauron, to what it knows about nudity.

This is clearly not the image I was looking for.
Giving it more specific prompting: Lord of the rings, sauron, wearing comical heart print underwear Negative prompt: Armor. This returns an image of a man with a similar build as before, but now lacking the bone spikes. This man has a prominent feature on his chest, similar to the glowing shape of the armored sauron but it's a heart shape, and he's wearing brief style underwear with only one heart, and a multitude of rings or rings shaped things on them.

This is the process by which one gets the system to produce something that it has never seen before in the fashion that one wants.
If you were to just ask it as though it should know what that immediate concept is, it will make vague random guesses that fit with what it knows.
Further concepts are required to guide it in the direction you're looking.
It does not know what sauron looks like under his armor, but it does know how to leave armor out of an image, what a naked man looks like, and what comical print underwear looks like, and that sauron is spikey and frequently associated with "dark, ominous, fire, volcano".

It does not know what child pornography would looks like, but it does know how all of the words you would use to describe it in detail change an image, and how to combine them together so they fit with each other.
That's why the individual in question is mentioned as having extremely detailed prompts. That's required to get it to do something it's unfamiliar with.
If you want a naked woman you can just say naked woman and it can go from there, because it knows that concept. Well enough that you often have to put in some work to get it to not include a naked woman, since it can associate them with anything.

Again, I've never tested the exact subject in question, and I'm not about to. So yes, I guess I am making some speculations here.

But I remember my own stages of puberty, and there are distinct differences between a young boy and a grown man. All I'm getting at is that if AI can generate a convincingly realistic image of an underage pre-puberty nude, then it must have been partly trained with underage images.

You are painfully dense or trolling.

One last try before I give up: can you describe what a pre-puberty individual looks like without referring to them as such? Congratulations, you know how to ask an AI that has never seen an inappropriate image of a child for one.

So yes, I guess I am making some speculations here.

That is abundantly clear, as you've been repeatedly corrected about how those speculations are entirely incorrect.

More painfully dense than anything.

Don't blame the AI systems, they don't know anything more than what humans have taught it. Blame those that fed in inappropriate material, whether it's the images or the words and prompts. Blame those feeding AI the garbage, whether it's the developers or the users, or the sites the AI scrapes data from.

AI is neat and all, and often funny even, when used properly. But at the end of the day, whatever it generates was somehow created by human input.

Find and prosecute the humans that are using it for inappropriate purposes.