Justice Thomas too forecasted that “technology may evolve to the point where it becomes impossible to enforce actual child pornography laws because the Government cannot prove that certain pornographic images are of real children.” In his view, “if technological advances thwart prosecution of ‘unlawful speech,’ the Government may well have a compelling interest in barring or otherwise regulating some narrow category of ‘lawful speech’ in order to enforce effectively laws against pornography made through the abuse of real children.”
This is a big contingency in your argument, and it’s not an assumption we can make.
The stochastic copy and paste machine needs similar input to what it is expected to output. That’s why there is an immense effort to collect as much data as possible from everything right now. If the relevant data isn’t in the training set then it can’t produce a related output.
You can train models on the output of other models, but at some point the original csam was produced and then processed for patterns by the machine. If you just copy those tagged patterns your still basing your output on very real csam.
Historically the US has banned all forms of csam including hand drawn or generated. This is a shift in that paradigm which now opens the flood gates for investigators as it dilutes real cases with facsimiles generated from real cases.
I’m not a free speech absolutist. There are categories of speech that should be protected and categories of speech that should not; and there are actions and products that absolutely should not belong in the “speech” category to begin with. Spending money is not speech, taking and owning pictures is not speech, and generating pictures of children being abused is not speech. There’s no reason any of that should be bundled up in our overly broad web of “speech” just cause we want to err on the side of personal liberty over common good and protection of this vulnerable in society.
Sort of, but not necessarily to the degree this comment suggests.
A good image generation model can create a photorealistic image of a child riding a three-headed dragon flying over the surface of Mars. It has definitely never seen a real image of that, or one of the major components of it. The copying is at a lower level than the whole scene, and the presence of an element in the output does not imply exactly that element was in the training data.
There have been cases of models being withdrawn after discovery that they were trained on actual CSAM, and allegations that Grok was trained on CSAM in a recent lawsuit. It seems reasonable to me to treat a model trained on CSAM as CSAM itself in most cases.
The model has seen images of dragons (likely even multiheaded ones), children, riding, and Mars. Those existing images are what allows it to pick the appropriate patterns and draw an output.
There are significant physiological differences between a child’s body and an adult’s. If you don’t have a collection of naked children to pick from it will be difficult to produce naked children to the extent indistinguishable from actual pictures of abuse. You might can get partway there building off a chain of descriptors based on an adult model, but I don’t think you’re getting all the way to photorealistic child abuse material without the context trained on child abuse material.
I’m absolutely not willing to test my hypothesis, however, for what I hope are fairly obvious reasons.
There are adult porn models who have popularity based on how youthful their appearance is. Small, even almost flat chests, for example. Which is to say that AI knows how to draw those bodies proportionately. Take those models, add in pictures of children in swimsuits, and AI has all the information to infer what a nude child looks like.
And that’s before opening up the extremely gray area of perfectly legal photos which nonetheless depict nude children.
The reality is if you try to scrape all available images from the internet you will unavoidably ingest CSAM, but even if you could perfectly filter that out, AI would be able to easily create such depictions.
This hints at another issue: most nudity is not pornographic. A model could, for example be trained on medical and scientific imagery, or just innocent family photos; it’s my impression photos containing nude young children were more acceptable in the pre-digital era. It’s certainly distasteful that AI companies are training on such material without explicit permission from the subjects, but that’s true of almost all AI training material.
I share your unwillingness to test how good image generation models are at making a realistic facsimile of CSAM. Legal or not, that’s yucky and I don’t want to look at it.
To play devil’s advocate: those training pictures already exist, and AI generating fake images means another child wasn’t harmed.
I didn’t make any argument and thought I was clear that I wasn’t going to make an argument, one way or another. I also stated that it should probably be illigal though I’m torn on it.
I’m also not a free speech absolutist. There are things that need curtailed. I’m just hesitant to add more to that list as it can easily go too far, depending on who is in power. We see that world wide right now with Free Palestine.
Free speech isn’t just about something spoken, it’s about expression as well. Taking and owning pictures is part of free speech. Limiting porn to 18+ is a great limitation and should not be changed.
These are all arguments. Non committing doesn’t mean you aren’t making arguments. I was focusing specially on your 3rd argument as it was the most unfounded in my opinion.