In the days after the US Department of Justice (DOJ) published 3.5 million pages of documents related to the late sex offender Jeffrey Epstein, multiple users on X have asked Grok to “unblur” or remove the black boxes covering the faces of children and women in images that were meant to protect their privacy.
Tbf it’s not needed. If it can draw children and it can draw nude adults, it can draw nude children.
Just like it doesn’t need to have trained on purple geese to draw one. It just needs to know how to draw purple things and how to draw geese.
I don’t think so. Speaking as a parent.
What you don’t think?
Why does being a parent give any authority in this conversation?
I have changed diapers and can attest to the anatomical differences between child and adult, and therefore know AI cannot extrapolate that difference without accurate data clarifying these differences. AI would hallucinate something absurd or impossible without real image data trained in its model.
We have all been children, we all know the anatomical differences.
It’s not like children are alien, most differences are just “this is smaller and a slightly different shape in children”. Many of those differences can be seen on fully clothed children. And for the rest, there are non-CSAM images that happen to have nude children. As I said earlier, it is not uncommon for children to be fully nude in beaches.
We are human beings. AI is not. It never had that experience of being or caring for a child. It does not (or should not) have that data in its dataset.
that’s not true, a child and an adult are not the same. and ai can not do such things without the training data. it’s the full wine glass problem. and the only reason THAT example was fixed after it was used to show the methodology problem with AI, is because they literally trained it for that specific thing to cover it up.
I’m not saying it wasnt trained on csam or defending any AI.
But your point isn’t correct
What prompts you use and how you request changes can get same results. Clever prompts already circumvent many hard wired protections. It’s a game of whackamole and every new iteration of an AI will require different methods needed bypass those protections.
If you can ask it the right ways it will do whatever a prompt tells it to do
It doesn’t take actual images/data trained if you can just tell it how to get the results you want it to by using different language that it hasn’t been told not to accept.
The AI doesn’t know what it is doing, it’s simply running points through its system and outputting the results.
It still seems pretty random. So they’ll say they fixed it so it won’t do something, all they likely did was reduce probability, so we still get screenshots showing what it sometimes lets through.
That’s not exactly true. I don’t know about today, but I remember about a year ago reading an article about an image generation model not being able, with many attempts, to generate a wine glass full to the brim, because all the wine glasses the model was trained on were half-filled.
Did it have any full glasses of water? According to my theory, It has to have data for both “full” and “wine”
Your theory is more or less incorrect. It can’t interpolate as broadly as you think it can.
The wine thing could prove me wrong if someone could answer my question.
But I don’t think my theory is that wild. LLMs can interpolate, and that is a fact. You can ask it to make a bear with duck hands and it will do it. I’ve seen images on the internet of things similar to that generated by LLMs.
Who is to say interpolating nude children from regular children+nude adults is too wild?
Furthermore, you don’t need CSAM for photos of nude children.
Children are nude at beaches all the time, there probably are many photos on the internet where there are nude children in the background of beach photos. That would probably help the LLM.
You are confusing LLMs with diffusion models. LLMs generate text, not images. They can be used as inputs to diffusion models and are thus usually intertwined but are not responsible for generating the images themselves. I am not completely refuting your point in general. Generative models are capable of generalising to an extend, so it is possible that such a system would be able to generate such images without having seen them. But how anatomically correct that would be is an entirely different question and the way these companies vastly sweep through the internet makes it very possible that these images were part of the training
Well yes, the LLMs are not the ones that actually generate the images. They basically act as a translator between the image generator and the human text input. Well, just the tokenizer probably. But that’s beside the point. Both LLMs and image generators are generative AI. And have similar mechanisms. They both can create never-before seen content by mixing things it has “seen”.
I’m not claiming that they didn’t use CSAM to train their models. I’m just saying that’s this is not definitive proof of it.
It’s like claiming that you’re a good mathematician because you can calculate 2+2. Good mathematicians can do that, but so can bad mathematicians.