doesn’t it follow that AI-generated CSAM can only be generated if the AI has been trained on CSAM?
This article even explicitely says as much.
My question is: why aren’t OpenAI, Google, Microsoft, Anthropic… sued for possession of CSAM? It’s clearly in their training datasets.
Yeah but the article suggests that pedos train their local AI on existing CSAM, which would indicate that it’s somehow needed to generate AI-generated CSAM. Otherwise why would they bother? They’d just feed them images of children in innocent settings and images of ordinary porn to get their local AI to generate CSAM.
It’s certainly technically possible. I suspect these AI models just aren’t good at it. So the pedophiles need to train them on actual images.
I can imagine for example AI doesn’t know what puberty is since it has in fact not seen a lot of naked children. It would try to infer from all the internet porn it’s seen, and draw any female with big breasts, disregarding age. And that’s not how children actually look.
I haven’t tried, since it’s illegal where I live. But that’s my suspicion why pedophiles bother with training models.
(Edit: If that’s the case, it would mean the tech companies are more or less innocent. At least at this.
And note a lot of the CSAM talk is FUD (spreading fear, uncertainty and doubt) I usually see this in the context of someone pushing for total surveillance of the people. It’s far less pronounced in my experience than some people make it to be. I’ve been around on the internet, and I haven’t seen any real pictures, yet. I’m glad that I didn’t, but that makes me believe you have to actively look for that kind of stuff, or be targeted somehow.
And I think a bit mure nuance would help. This article also lumps together fictional drawings and real pictures. I think that’s counterproductive, since one is a heinous crime and has real victims. And like, drawing nude anime children or de-aging celebrities isn’t acceptable either (depends on legislation), but I think we need to differentiate here. I think real pictures are entirely on a different level and should have far more severe consequences. If we mix everything together, we kind of take away from that.)
This is not strictly true in general. Generative AI is able to produce output that is not in the training data, by learning a broad range of concepts and applying them in novel ways. I can generate an image of a rollerskating astronaut even if there are no rollerskating astronauts in the training data.
It is true that some training sets include CSAM, at least in the past. Back in 2023, researches found a few thousand such images in the LAION-5B dataset (roughly one per million images). 404 Media has an excellent article with details: https://www.404media.co/laion-datasets-removed-stanford-csam-child-abuse/
On learning of this, LAION took down their database until it could properly cleaned. Source: https://laion.ai/notes/laion-maintenance/
Those images were collected from the public web. LAION took steps to avoid linking to illicit content (details in the link above), but clearly it’s an imperfect system. God only knows what closed companies (OpenAI, Google, etc.) are doing. With open data sets, at least any interested parties can review, verify, and report this stuff. With closed data sets, who knows?
Training an existing model on a specific set of new data is known as “fine tuning”.
A base model has broad world knowledge and the ability to generate outputs of things it hasn’t specifically seen, but a tuned model will provide “better” (fucking yuck to even write it) results.
The closer your training data is to your desired result, the better.
That’s not exactly how it works.
It can “understand” different concepts and mix them, without having to see the combination before hand.
As for the training thing, that would probably be more LORA. They’re like add-ons you can put on your AI to draw certain things better like a character, a pose, etc. not needed for the base model.