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.
There’s a lot of misconceptions about AI image generators in here.
They can indeed generate “truly new images”, ask an image generator for an image of something that definitely doesn’t exist in its training set and it’ll likely be able to come up with something like that for you. Most importantly for purposes of this discussion, you don’t actually need to have any images specifically of child abuse in a model’s training set in order to train it well enough to produce images of child abuse. Train a model with a bunch of regular porn and a bunch of ordinary images of children and I expect it’ll figure out how to make images of children in sexual situations if you ask it to.
This has been known for years. These AIs are capable of “understanding” the things they’re trained on and creating novel interpretations of those things.
There was an article recently that showed if you trained many generations of AIs on just the outputs of previous generations you got degraded performance over time, but that’s a pretty specific scenario that doesn’t match what’s being done in real life. In real life synthetic training data (ie, AI-generated training data) can be very useful for expanding the capabilities of AI as long as it’s well-curated (humans need to select good outputs and ensure they’re described correctly) and ideally has some of the earlier training set’s original data included as well.
So basically, in order to have AIs that make better child porn, there needs to be humans willing to go through vast quantities of AI generated child porn in order to properly curate the content for the AI. Since this labour is likely to be farmed out to innocent people in developing countries, being paid slave wages, I think it would be fair to add them to the list of potential victims of the creation of child porn AIs.
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.
Like by all means go after people harming others, but man… You realize you have no control right? Black markets THRIVE in this world. Like fuck, they can’t even keep drugs out of prison.
They don’t have a plan they just want to blissfully hand law enforcement more power to spy on citizens and toss non violent offenders in prison. War on drugs called, wants its script back.
The fact that regulation is imperfect doesn’t mean the solution is to do absolutely nothing and let AI be used by bad actors (including those feeding real child porn into it) with impunity.
The entire conversation in the comments here have been “well if AI child porn stops paedophiles from hurting real children, then it has a use”, but in order to prevent real children being used for the training data, it has to be regulated. In fact, you’ll find that the majority of Beehaw users (since I notice you are from a different instance) are in favour of AI being properly regulated. “Some people will break the law anyway” is no excuse not to have laws.
Sorry, I really don’t support your perspective. If you have reasonable suspicion to believe real child porn is being used in these models prove it to a judge and get a warrant. If someone is a bad actor to the point of criminal intent, prove it to a judge, and bring it to court.
There’s a lot of misconceptions about AI image generators in here.
They can indeed generate “truly new images”, ask an image generator for an image of something that definitely doesn’t exist in its training set and it’ll likely be able to come up with something like that for you. Most importantly for purposes of this discussion, you don’t actually need to have any images specifically of child abuse in a model’s training set in order to train it well enough to produce images of child abuse. Train a model with a bunch of regular porn and a bunch of ordinary images of children and I expect it’ll figure out how to make images of children in sexual situations if you ask it to.
This has been known for years. These AIs are capable of “understanding” the things they’re trained on and creating novel interpretations of those things.
There was an article recently that showed if you trained many generations of AIs on just the outputs of previous generations you got degraded performance over time, but that’s a pretty specific scenario that doesn’t match what’s being done in real life. In real life synthetic training data (ie, AI-generated training data) can be very useful for expanding the capabilities of AI as long as it’s well-curated (humans need to select good outputs and ensure they’re described correctly) and ideally has some of the earlier training set’s original data included as well.
So basically, in order to have AIs that make better child porn, there needs to be humans willing to go through vast quantities of AI generated child porn in order to properly curate the content for the AI. Since this labour is likely to be farmed out to innocent people in developing countries, being paid slave wages, I think it would be fair to add them to the list of potential victims of the creation of child porn AIs.
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.
Dude how are you going to regulate AI?!!
Like by all means go after people harming others, but man… You realize you have no control right? Black markets THRIVE in this world. Like fuck, they can’t even keep drugs out of prison.
They don’t have a plan they just want to blissfully hand law enforcement more power to spy on citizens and toss non violent offenders in prison. War on drugs called, wants its script back.
The fact that regulation is imperfect doesn’t mean the solution is to do absolutely nothing and let AI be used by bad actors (including those feeding real child porn into it) with impunity.
The entire conversation in the comments here have been “well if AI child porn stops paedophiles from hurting real children, then it has a use”, but in order to prevent real children being used for the training data, it has to be regulated. In fact, you’ll find that the majority of Beehaw users (since I notice you are from a different instance) are in favour of AI being properly regulated. “Some people will break the law anyway” is no excuse not to have laws.
Sorry, I really don’t support your perspective. If you have reasonable suspicion to believe real child porn is being used in these models prove it to a judge and get a warrant. If someone is a bad actor to the point of criminal intent, prove it to a judge, and bring it to court.
That’s not how the law works.