• Allero@lemmy.today
    link
    fedilink
    English
    arrow-up
    9
    ·
    2 days ago

    That’s exactly how they work. According to many articles I’ve seen in the past, one of the most common models used for this purpose is Stable Diffusion. For all we know, this model was never fed with any CSAM materials, but it seems to be good enough for people to get off - which is exactly what matters.

    • DoPeopleLookHere@sh.itjust.works
      link
      fedilink
      English
      arrow-up
      1
      arrow-down
      3
      ·
      2 days ago

      How can it be trained to produce something without human input.

      To verify it’s models are indeed correct, some human has to sit and view it.

      Will that be you?

      • TheRealKuni@midwest.social
        link
        fedilink
        English
        arrow-up
        5
        ·
        2 days ago

        How can it be trained to produce something without human input.

        It wasn’t trained to produce every specific image it produces. That would make it pointless. It “learns” concepts and then applies them.

        No one trained AI on material of Donald Trump sucking on feet, but it can still generate it.

        • DoPeopleLookHere@sh.itjust.works
          link
          fedilink
          English
          arrow-up
          1
          arrow-down
          2
          ·
          2 days ago

          It was able to produce that because enough images of both feet and Donald Trump exist.

          How would it know what young genitals look like?

          • JuxtaposedJaguar@lemmy.ml
            link
            fedilink
            English
            arrow-up
            3
            ·
            1 day ago

            If you train a model on 1,000,000 images of dogs and 1,000,000 images of cats, your output isn’t going to be a 50/50 split of purely dogs and purely cats, it’s going to be (on average) somewhere between a cat and a dog. At no point did you have to feed in pictures of dog-cat hybrids to end up with that model.

          • JuxtaposedJaguar@lemmy.ml
            link
            fedilink
            English
            arrow-up
            3
            ·
            1 day ago

            You could probably make some semi-realistic drawings and feed those in, and then re-train the model with those same images over and over until the model is biased to use the child-like properties of the drawings but the realism of the adult pictures. You could also feed the most CP-looking images from a partially trained model as the training data of another model, which over time would make the outputs approach the desired result.

              • JuxtaposedJaguar@lemmy.ml
                link
                fedilink
                English
                arrow-up
                4
                ·
                1 day ago

                It doesn’t matter if it’s accurate or not as long as pedos can get off to it, so just keep going until they can. According to our definition of what a pedophile is, though, it would likely be accurate.

      • Allero@lemmy.today
        link
        fedilink
        English
        arrow-up
        4
        ·
        2 days ago

        Much as all in modern AI - it’s able to train without much human intervention.

        My point is, even if results are not perfectly accurate and resembling a child’s body, they work. They are widely used, in fact, so widely that Europol made a giant issue out of it. People get off to whatever it manages to produce, and that’s what matters.

        I do not care about how accurate it is, because it’s not me who consumes this content. I care about how efficient it is at curbing worse desires in pedophiles, because I care about safety of children.