A new report warns that the proliferation of child sexual abuse images on the internet could become much worse if something is not done to put controls on artificial intelligence tools that generate deepfake photos.
Images, yes, but mixing concepts is a mixed bag. Just because the model can draw, say, human faces and dog faces doesn’t mean it has the understanding necessary to blend those concepts. Without employing specialised models (and yes of course the furries have been busy) the best you’ll get is facepaint. The pope at a beach bar doesn’t even come close to exercising that kind of capability: The pope is still the pope and the beach bar is still the beach bar, and a person is still sitting there slurping a caipirinha.
I mean if you train a model on porn with adult actors and on regular photos with children, it shouldn’t be hard to generate the combination.
You probably wouldn’t even need any fancy training data but if you really wanted you could pick adult actors that look young or in other ways similar to the children to help the process.
Knowing what a nude adult looks like doesn’t mean that the model knows what a nude child looks like. I’m quite sure it’s easy to generate disturbing images like that, but actual paedophiles I think won’t be satisfied with child faces on small adult bodies.
Ordinary deepfakes actually have a very similar problem: Sure you can take a picture of a celebrity and tell the AI to undress them – but it won’t be their actual body. The AI is going to be able to approximate their overall build but it’s going to be a generic adult body, not the celebrity’s body. Or, differently put, AI models aren’t any better at undressing people with their eyes than teenagers.
I see where you’re coming from but that’s a technical issue that will probably be solved in time.
It’s also really not a black and white; sure maybe you can see it isn’t perfect but you’d still prefer it to content where you know no one was actually harmed.
Despite what reputation people like that have (due to the simple fact of how reporting works), most are harmless like me and you and don’t actually want to see innocent people suffer and would never act on their desires. So having a safe and harmless outlet might help.
I see where you’re coming from but that’s a technical issue that will probably be solved in time.
You cannot create information from nothing.
So having a safe and harmless outlet might help.
Psychologists/Psychiatrists are still on the fence on that one, I wouldn’t be surprised if it depends on the person. And yes the external harm produced by AI images is definitely lower than that produced from actual CSAM, doubly so newly produced CSAM, but that doesn’t mean that therapy, even in its current early stages, couldn’t do even better.
Differently put: We may be again falling into the trap of trying to find technological solutions to societal problems (well, this is /c/technology…). Which isn’t to say that we shouldn’t care at all about models trained on CSAM, but that’s addressing symptoms, not causes. Ultimately addressing root causes is more important: The vast majority of paedophiles are not exclusive paedophiles, often they’re not even really attracted to kids at all beyond having developed a fetish, they’re rapists focussing on the most vulnerable, often due to having been victims of sexual abuse themselves.
Arguably that’s exactly what generative AIs do. Which is not what you meant, but yeah. I was going more for like “given current progress and advancements in how we curate datasets and whatnot, there is no reason to believe that we won’t have 100% undistinguishable AI-generated pictures eventually”.
We already know that you don’t need to have stuff in the training dataset to have it show up meaningfully in the output.
Psychologists/Psychiatrists are still on the fence on that one, I wouldn’t be surprised if it depends on the person. And yes the external harm produced by AI images is definitely lower than that produced from actual CSAM, doubly so newly produced CSAM, but that doesn’t mean that therapy, even in its current early stages, couldn’t do even better.
100% agree there. What I would like to see is more research, but that’s currently kinda impossible with CSAM being as criminalized as it is. Which is kinda sad.
Therapy seems to work on most help-seeking people (and there are studies proving that), so this should be a last ditch effort.
The rest of your post I don’t agree with. It isn’t really (definitely not exclusively) a societal problem - some people’s brains are simply wired in a way that’s just bad and there isn’t much you can do with it, and either these people suffer by living with it, or they cause harm to others because of it. Both is bad.
The vast majority of paedophiles are not exclusive paedophiles, often they’re not even really attracted to kids at all beyond having developed a fetish, they’re rapists focussing on the most vulnerable, often due to having been victims of sexual abuse themselves.
Do you have any statistics proving this? It’s exactly the bias that already makes non-acting pedophiles unlikely to seek help. Obviously these kinds of people are the ones you hear most about, but I wouldn’t be so sure that they’re the majority (even if they’re most of the problem).
My point is that if you take it as people who need help and actually manage to provide it, you should be able to get the number of abuse down overall except for the people who truly can’t be helped. And it really doesn’t matter much how you provide that help, even if it’s morally questionable like using artificially generated CSAM.
Images, yes, but mixing concepts is a mixed bag. Just because the model can draw, say, human faces and dog faces doesn’t mean it has the understanding necessary to blend those concepts. Without employing specialised models (and yes of course the furries have been busy) the best you’ll get is facepaint. The pope at a beach bar doesn’t even come close to exercising that kind of capability: The pope is still the pope and the beach bar is still the beach bar, and a person is still sitting there slurping a caipirinha.
I mean if you train a model on porn with adult actors and on regular photos with children, it shouldn’t be hard to generate the combination.
You probably wouldn’t even need any fancy training data but if you really wanted you could pick adult actors that look young or in other ways similar to the children to help the process.
Knowing what a nude adult looks like doesn’t mean that the model knows what a nude child looks like. I’m quite sure it’s easy to generate disturbing images like that, but actual paedophiles I think won’t be satisfied with child faces on small adult bodies.
Ordinary deepfakes actually have a very similar problem: Sure you can take a picture of a celebrity and tell the AI to undress them – but it won’t be their actual body. The AI is going to be able to approximate their overall build but it’s going to be a generic adult body, not the celebrity’s body. Or, differently put, AI models aren’t any better at undressing people with their eyes than teenagers.
I see where you’re coming from but that’s a technical issue that will probably be solved in time.
It’s also really not a black and white; sure maybe you can see it isn’t perfect but you’d still prefer it to content where you know no one was actually harmed.
Despite what reputation people like that have (due to the simple fact of how reporting works), most are harmless like me and you and don’t actually want to see innocent people suffer and would never act on their desires. So having a safe and harmless outlet might help.
You cannot create information from nothing.
Psychologists/Psychiatrists are still on the fence on that one, I wouldn’t be surprised if it depends on the person. And yes the external harm produced by AI images is definitely lower than that produced from actual CSAM, doubly so newly produced CSAM, but that doesn’t mean that therapy, even in its current early stages, couldn’t do even better.
Differently put: We may be again falling into the trap of trying to find technological solutions to societal problems (well, this is /c/technology…). Which isn’t to say that we shouldn’t care at all about models trained on CSAM, but that’s addressing symptoms, not causes. Ultimately addressing root causes is more important: The vast majority of paedophiles are not exclusive paedophiles, often they’re not even really attracted to kids at all beyond having developed a fetish, they’re rapists focussing on the most vulnerable, often due to having been victims of sexual abuse themselves.
Arguably that’s exactly what generative AIs do. Which is not what you meant, but yeah. I was going more for like “given current progress and advancements in how we curate datasets and whatnot, there is no reason to believe that we won’t have 100% undistinguishable AI-generated pictures eventually”.
We already know that you don’t need to have stuff in the training dataset to have it show up meaningfully in the output.
100% agree there. What I would like to see is more research, but that’s currently kinda impossible with CSAM being as criminalized as it is. Which is kinda sad.
Therapy seems to work on most help-seeking people (and there are studies proving that), so this should be a last ditch effort.
The rest of your post I don’t agree with. It isn’t really (definitely not exclusively) a societal problem - some people’s brains are simply wired in a way that’s just bad and there isn’t much you can do with it, and either these people suffer by living with it, or they cause harm to others because of it. Both is bad.
Do you have any statistics proving this? It’s exactly the bias that already makes non-acting pedophiles unlikely to seek help. Obviously these kinds of people are the ones you hear most about, but I wouldn’t be so sure that they’re the majority (even if they’re most of the problem).
My point is that if you take it as people who need help and actually manage to provide it, you should be able to get the number of abuse down overall except for the people who truly can’t be helped. And it really doesn’t matter much how you provide that help, even if it’s morally questionable like using artificially generated CSAM.
deleted by creator