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.
Since this labour is likely to be farmed out to innocent people in developing countries
You don’t quite seem to understand how easy it is to train these AI models, and because of that, you’re missing a critical point - with open-source technologies like Stable Diffusion, which has models that can be refined and run on a consumer-grade graphics card, the people using models to generate images and the people creating and refining those models are the same people. People who want to generate brand new pokemon sprites can train a model on all the pokemon sprites until it looks good. A few absolute galaxy-brain nerds who want to generate MIDI spectrograms from a text description and convert the output into audio… can apparently do that. And of course, people who want to generate lots of hentai or photorealistic porn can create and fine-tune a model, or multiple models, all by themselves (I won’t link any of these, but hundreds are readily available, and thousands exist in total)
In other words, people who already consume CSAM are the people working on models for generating CP, and a subset of those have definitely been trying to make it work with only legal images so that the model itself can be distributed and used without breaking any laws, maybe even hiding in plain sight pretending it’s not for making CP. Someone else out there with a different set of fucked-up desires has probably trained a model on gore and snuff images and then used it to create “photos” of people they hate as mutilated messes. There’s sick people of all kinds all over the place, and the jury’s unfortunately still out on whether this new tool actually causes harm when used in such a manner, or if it’s just the newest way they can express their deviance. We don’t know yet.
But this genie is already out of the bottle. Banning the use of this technology for specific, narrow use cases just isn’t going to be effective without banning AI image generation entirely, and we’re past the point where that’s feasible. Image generation is a powerful tool that’s not going away; it’s on us now to figure out what we really believe about harm, health, and personal freedom, and what we want a society with this tool to look like.
Personally, I’m of a mind that if all the data going into the model is legally obtained, anything generated should be considered artistic expression. A person had a thought, then put their thoughts into a tool, which made a picture of those thoughts. No matter how repulsive those thoughts were, I think throwing people in prison for that kind of expression is thought-crime. There’s public obscenity at play, of course, but only once they take the step of showing it to other people. If it’s just for themselves, and nobody else sees it, who is harmed? Even if it does turn out that it harms the person generating the images (which wouldn’t surprise me), that makes it a health issue, like drugs or other addictions, not something to criminalize.
Except if we’re talking about allowing regulated use of AI-generated child porn as a treatment for paedophilia (which is the core discussion in this thread, whether AI-generated child porn could be helpful), then it cannot be left to paedophiles themselves to create their own models based on the honour system of promising that nobody will use photos of real children in the training data. Just like absolutely no aspect of AI can be left unregulated, trusting on everyone to behave honourably… because so far, nobody has been. Not just in the field of AI-generated child porn. All of AI has been developed with an astounding level of unethical behaviour, and it’s nothing short of complete naivety to believe that anyone is going to behave ethically going forward. If AI is to be considered in any way ethical, then it needs regulation beyond simply asking people to pinky swear that they will only use legally and ethically obtained content for training. Regulation requires oversight and enforcement, otherwise regulation is meaningless. And you can’t regulate child porn AI without having innocent human beings subjected to the inputs and outputs to ensure regulations are being followed.
As I said above, though, you don’t need to make a model that’s specifically “for child porn” in order for it to be able to generate child porn. There are already probably plenty of models that know what children look like and also know what porn looks like, made simply by teaching a model about lots of diverse subjects that happened to include both of those subject areas in them. You can even make new models by merging two existing models together or by adding more training to an existing model, so you wouldn’t even need to have those images be part of the same training run.
I obviously haven’t ever tried generating child porn, but I fired up my local Stable Diffusion with the Cyberrealistic model and generated a toddler on the moon and a toddler riding a lion. I’m reasonably confident that the model wasn’t literally trained with images of toddlers in space suits or toddlers riding large wild predators, it was trained on those concepts separately and was able to figure out for itself how to combine them. Notice how it was able to figure out that a toddler on the moon would be in a space suit and re-proportioned the space suit accordingly, and that a saddle used by a toddler would probably have handlebars (I’m guessing it has a bunch of images of toddlers riding ponies that it got that idea from).
@the_third@feddit.de did a very good post above explaining the problem with the idea of thinking AI-generated child porn would be as simple as asking for non-abusive photos of children to be combined with photos of adult porn. The AI needs to know what each component of the image should look like. AI knows what toddlers look like, and it knows what a lion looks like. Where do you propose the photos of child genitals should come from in order to create these ethical AI-generated child abuse images?
Probably fine for child abuse porn in a drawn style, like loli, but probably not sufficient for the photorealistic porn that it is supposedly intended to replace.
This is ultimately the flaw in your argument that an AI can produce “new” works because it doesn’t need to have seen a toddler riding a lion on the moon to be able to produce that image. If you didn’t give AI photos of lions, it would never be able to create a lion. If you never gave it photos of toddlers, it would never be able to do a picture of a toddler. And if there were no photos of the moon in its training data, it would be incapable of producing the moon. It cannot create things it has not seen. It can only arrange things it has seen in combinations that may or may not have been previously thought of (with billions of images in the training data, you can’t say there isn’t a Photoshopped photo of a toddler on the moon in there.)
Without actual child porn in its training data, it would never be able to produce any, because even when it’s capable of piecing individual elements together into a “new” piece (basically an advanced collage), if there’s no images of children being abused in the training data, it’s not going to be able to piece it together by putting a child’s head on an adult’s naked body, and result in anything that’s more satisfactory for paedophiles than actual photos of children. There is therefore no ethical means of producing AI-generated child porn, and therefore it is not an ethical alternative. Somewhere in the chain, there are still children being abused.
I think you haven’t made much use of image-generating AIs. They’re quite capable of reinterpreting images into different styles. A very common use case for me is to draw a sketch of something and then tell an image AI to turn it photorealistic. The “automated collage” approach you describe is simply not how they work, it’s a common misconception. Image AIs very much can create imagery of things that weren’t explicitly in their training set, they’re not just regurgitating pasted-together snippets.
You’re also assuming that there are no literal photographs of children’s genitals in medical literature. Again, I haven’t exactly gone looking, but I’m sure there are some out there. Doctors can’t afford to be prudish.
And finally, you can get plenty pornographic without even specifically showing off genitals.
Oh, so you mean the photos in the training dataset that violate medical privacy, because parents who consented to photos of their child being included in textbooks and medical journals for educational purposes didn’t consent to photos of their child being used for AI-generated child porn for the sexual gratification of paedophiles? The only way photos of children from medical literature would be ethical to use for child porn is if the parents of the child have consented to that usage. Having had surgery last year, where photos were taken, I can confirm there are extensive consent forms to be filled out for what medical professionals can record (photos, videos, livestreaming, etc), and what they can use the visual records for (research, education/training, sharing cosmetic outcomes on social media, etc). Parents that will have checked the “can use photos of my child for research and education” will not have given informed consent for those photos to be inserted into an AI model for child porn, and are unlikely to give consent if “child porn AI” is a separate box on the consent form.
So… yeah, you’re not convincing me using medical photos of vulnerable children in hospital settings to create fapping material for paedophiles is an ethical use of AI technology.
And an AI without photos of lions is never going to be able to produce a photorealistic lion, even if you gave it a sketch of a lion, because it would have no frame of reference for what a lion is supposed to look like. It would make its best guess, which is fine for when it’s something that doesn’t really exist - but when humans know what a lion is meant to look like, they’ll know when an AI botches it.
Oh, so you mean the photos in the training dataset that violate medical privacy
If they’re published in a textbook then they’re not private.
will not have given informed consent for those photos to be inserted into an AI model for child porn
Again, an AI model doesn’t have to be created specifically for the purpose of child porn in order for it to be able to generate child porn. Most of these AI image models are very general purpose, they can create images of all kinds of things.
We’re going in circles here and you’re just getting angrier in your responses, I don’t think this is headed anywhere useful at this point.
You don’t quite seem to understand how easy it is to train these AI models, and because of that, you’re missing a critical point - with open-source technologies like Stable Diffusion, which has models that can be refined and run on a consumer-grade graphics card, the people using models to generate images and the people creating and refining those models are the same people. People who want to generate brand new pokemon sprites can train a model on all the pokemon sprites until it looks good. A few absolute galaxy-brain nerds who want to generate MIDI spectrograms from a text description and convert the output into audio… can apparently do that. And of course, people who want to generate lots of hentai or photorealistic porn can create and fine-tune a model, or multiple models, all by themselves (I won’t link any of these, but hundreds are readily available, and thousands exist in total)
In other words, people who already consume CSAM are the people working on models for generating CP, and a subset of those have definitely been trying to make it work with only legal images so that the model itself can be distributed and used without breaking any laws, maybe even hiding in plain sight pretending it’s not for making CP. Someone else out there with a different set of fucked-up desires has probably trained a model on gore and snuff images and then used it to create “photos” of people they hate as mutilated messes. There’s sick people of all kinds all over the place, and the jury’s unfortunately still out on whether this new tool actually causes harm when used in such a manner, or if it’s just the newest way they can express their deviance. We don’t know yet.
But this genie is already out of the bottle. Banning the use of this technology for specific, narrow use cases just isn’t going to be effective without banning AI image generation entirely, and we’re past the point where that’s feasible. Image generation is a powerful tool that’s not going away; it’s on us now to figure out what we really believe about harm, health, and personal freedom, and what we want a society with this tool to look like.
Personally, I’m of a mind that if all the data going into the model is legally obtained, anything generated should be considered artistic expression. A person had a thought, then put their thoughts into a tool, which made a picture of those thoughts. No matter how repulsive those thoughts were, I think throwing people in prison for that kind of expression is thought-crime. There’s public obscenity at play, of course, but only once they take the step of showing it to other people. If it’s just for themselves, and nobody else sees it, who is harmed? Even if it does turn out that it harms the person generating the images (which wouldn’t surprise me), that makes it a health issue, like drugs or other addictions, not something to criminalize.
Except if we’re talking about allowing regulated use of AI-generated child porn as a treatment for paedophilia (which is the core discussion in this thread, whether AI-generated child porn could be helpful), then it cannot be left to paedophiles themselves to create their own models based on the honour system of promising that nobody will use photos of real children in the training data. Just like absolutely no aspect of AI can be left unregulated, trusting on everyone to behave honourably… because so far, nobody has been. Not just in the field of AI-generated child porn. All of AI has been developed with an astounding level of unethical behaviour, and it’s nothing short of complete naivety to believe that anyone is going to behave ethically going forward. If AI is to be considered in any way ethical, then it needs regulation beyond simply asking people to pinky swear that they will only use legally and ethically obtained content for training. Regulation requires oversight and enforcement, otherwise regulation is meaningless. And you can’t regulate child porn AI without having innocent human beings subjected to the inputs and outputs to ensure regulations are being followed.
As I said above, though, you don’t need to make a model that’s specifically “for child porn” in order for it to be able to generate child porn. There are already probably plenty of models that know what children look like and also know what porn looks like, made simply by teaching a model about lots of diverse subjects that happened to include both of those subject areas in them. You can even make new models by merging two existing models together or by adding more training to an existing model, so you wouldn’t even need to have those images be part of the same training run.
I obviously haven’t ever tried generating child porn, but I fired up my local Stable Diffusion with the Cyberrealistic model and generated a toddler on the moon and a toddler riding a lion. I’m reasonably confident that the model wasn’t literally trained with images of toddlers in space suits or toddlers riding large wild predators, it was trained on those concepts separately and was able to figure out for itself how to combine them. Notice how it was able to figure out that a toddler on the moon would be in a space suit and re-proportioned the space suit accordingly, and that a saddle used by a toddler would probably have handlebars (I’m guessing it has a bunch of images of toddlers riding ponies that it got that idea from).
@the_third@feddit.de did a very good post above explaining the problem with the idea of thinking AI-generated child porn would be as simple as asking for non-abusive photos of children to be combined with photos of adult porn. The AI needs to know what each component of the image should look like. AI knows what toddlers look like, and it knows what a lion looks like. Where do you propose the photos of child genitals should come from in order to create these ethical AI-generated child abuse images?
Medical textbooks.
Probably fine for child abuse porn in a drawn style, like loli, but probably not sufficient for the photorealistic porn that it is supposedly intended to replace.
This is ultimately the flaw in your argument that an AI can produce “new” works because it doesn’t need to have seen a toddler riding a lion on the moon to be able to produce that image. If you didn’t give AI photos of lions, it would never be able to create a lion. If you never gave it photos of toddlers, it would never be able to do a picture of a toddler. And if there were no photos of the moon in its training data, it would be incapable of producing the moon. It cannot create things it has not seen. It can only arrange things it has seen in combinations that may or may not have been previously thought of (with billions of images in the training data, you can’t say there isn’t a Photoshopped photo of a toddler on the moon in there.)
Without actual child porn in its training data, it would never be able to produce any, because even when it’s capable of piecing individual elements together into a “new” piece (basically an advanced collage), if there’s no images of children being abused in the training data, it’s not going to be able to piece it together by putting a child’s head on an adult’s naked body, and result in anything that’s more satisfactory for paedophiles than actual photos of children. There is therefore no ethical means of producing AI-generated child porn, and therefore it is not an ethical alternative. Somewhere in the chain, there are still children being abused.
I think you haven’t made much use of image-generating AIs. They’re quite capable of reinterpreting images into different styles. A very common use case for me is to draw a sketch of something and then tell an image AI to turn it photorealistic. The “automated collage” approach you describe is simply not how they work, it’s a common misconception. Image AIs very much can create imagery of things that weren’t explicitly in their training set, they’re not just regurgitating pasted-together snippets.
You’re also assuming that there are no literal photographs of children’s genitals in medical literature. Again, I haven’t exactly gone looking, but I’m sure there are some out there. Doctors can’t afford to be prudish.
And finally, you can get plenty pornographic without even specifically showing off genitals.
Oh, so you mean the photos in the training dataset that violate medical privacy, because parents who consented to photos of their child being included in textbooks and medical journals for educational purposes didn’t consent to photos of their child being used for AI-generated child porn for the sexual gratification of paedophiles? The only way photos of children from medical literature would be ethical to use for child porn is if the parents of the child have consented to that usage. Having had surgery last year, where photos were taken, I can confirm there are extensive consent forms to be filled out for what medical professionals can record (photos, videos, livestreaming, etc), and what they can use the visual records for (research, education/training, sharing cosmetic outcomes on social media, etc). Parents that will have checked the “can use photos of my child for research and education” will not have given informed consent for those photos to be inserted into an AI model for child porn, and are unlikely to give consent if “child porn AI” is a separate box on the consent form.
So… yeah, you’re not convincing me using medical photos of vulnerable children in hospital settings to create fapping material for paedophiles is an ethical use of AI technology.
And an AI without photos of lions is never going to be able to produce a photorealistic lion, even if you gave it a sketch of a lion, because it would have no frame of reference for what a lion is supposed to look like. It would make its best guess, which is fine for when it’s something that doesn’t really exist - but when humans know what a lion is meant to look like, they’ll know when an AI botches it.
If they’re published in a textbook then they’re not private.
Again, an AI model doesn’t have to be created specifically for the purpose of child porn in order for it to be able to generate child porn. Most of these AI image models are very general purpose, they can create images of all kinds of things.
We’re going in circles here and you’re just getting angrier in your responses, I don’t think this is headed anywhere useful at this point.