Brian Ribbon

The world is already in a state of awe, and alarm, over the extraordinary power and potentially terrifying consequences of AI. There are good reasons to be concerned, not least because the power to control social narratives will be concentrated in the hands of those few who own the bots. It is not surprising that politicians, who have made a career out of wielding power or attempting to do so, should seek to act against the new competition in order to maintain their own grip over the populace. In most cases, governments can expect some resistance to their attempts at censoring AI, whether or not the censorship is justifiable. There is one 'issue', however, that nobody is allowed to speak up against: 'pedophilia'. And so governments like the UK, with a long history of government overreach, spying, and censorship, are using the fear of AI 'child pornography' to drive the thin end of the authoritarian wedge under the door.

Possible legal amendments discussed in this article are thought experiments, not endorsements of legal reforms. An article on proposed legal reforms will be published later this year.

AI 'child pornography'

To help with this article, I have received information from an individual familiar with darknet websites. For legal reasons, Mu does not have a presence on such sites.

First of all, a primer on the language we use here at Mu. In this article, I will be referring to criminalized images of children as 'Prohibited Images of Minors' (PIM). This value-neutral language is chosen because a significant amount of material legally prosecuted as 'child pornography' or 'child sexual abuse material' is not even pornographic, often made by adolescents themselves, and can be generated entirely by a computer. The term PIM refers to material for which people may be prosecuted, and does not make any value judgements about whether or not the material constitutes abuse.

The majority of PIM created by AI is made using the programs Stable Diffusion and Flux, which are not specifically designed to generate PIM. There are various ways to generate images with these programs, including using an image as a basis for making further images. Models and LORAs can be created from existing images of real people, and the latter can be fine-tuned to create realistic AI generations of people who actually exist. It is possible for a very skilled person to make a model using thousands of images of naked minors, and a LORA of a young person they know, to make realistic AI images of that person engaging in sexual activity. The skills needed to do this are highly advanced, and the hardware required is very expensive. Even a top-of-the-line consumer GPU would have difficulty in outputting sufficient power to train a full model.

Distributing AI-generated images of real minors involved in sexual situations is clearly ethically wrong. It would be wrong to do so with real adults, too. Organizations such as the UK's Internet Watch Foundation (IWF) have latched onto the very low proportion of people doing this, in order to spread fear and panic. In reality, the majority of examples of AI PIM are made much more crudely, generated entirely by a computer program, no doubt partly due to the high level of skill and cost of equipment required in making sexual images of real young people. In these cases, the artist uses a text prompt, for example '14 year old boy, handsome, shirtless, jeans' to generate an image of a non-existent 14 year old boy wearing only a pair of jeans. Existing models, not trained on PIM and not specifically designed to produce it, are capable of generating PIM using text prompts like these (note that for legal reasons, I provided an example prompt that would not generate an unlawful image). There are reportedly instances of models and LORAs on darknet boards that people can download to help make their own AI PIM, and it is not unreasonable to claim that real people depicted are harmed by having their likeness used all over again, should they find out about it. However, an individual familiar with such matters states that this doesn't represent the majority of examples of AI PIM, and the IWF are lying to the public in claiming that it does.

A better option

One argument used to support the criminalization of AI PIM is that it 'fuels the fantasies' of MAPs, encouraging them to commit contact offenses against real people. This is repeated like a mantra, as if minor-attraction is a weird fetish rather than an inherent biological state. There is no evidence for the claim, but there is some evidence for its opposite: that having a form of sexual release is helpful in reducing the likelihood of the commission of contact offenses. Attraction to minors is an innate trait displayed by a significant number of people, and nobody can be 'turned pedo' or 'turned hebephilic', just as people are not 'turned gay' or 'turned straight' (the perceived morality of acting on minor-attraction versus adult-attraction is not at all relevant to this fact). MAPs remain MAPs no matter what anyone does; barbaric treatments like electroshock therapy or lobotomies have never changed the subject of anyone's sexual feelings, and therapy cannot 'turn' a MAP into a regular straight or gay person. Access to mental health support can help people to not act in harmful ways, but unfortunately safe access to mental healthcare is almost impossible for MAPs due to the incredible stigma surrounding an attraction to minors. Making access to mental health support easier, instead of banning a victimless form of sexual release such as images generated entirely by a computer, would help to prevent some people from committing contact offenses. Unfortunately, it seems that the public and even many professionals are unwilling to have this conversation.

The question many will ask is how we can know which images have been generated entirely by a computer. It is technologically possible. In Stable Diffusion, creating an image based on another image (such as a real person) requires a specific input in one of its GUIs, the use of which could be saved in metadata. Likewise, the particular model and LORA used can presently be saved in an image's metadata. It is not at all difficult to implement a system where the metadata contained within an image could be used to prove that an image was created using mainstream models and LORAs that are not based on the likeness of a real young person, and that an existing image was not involved in the generation of the AI image. This would prove, beyond any shadow of a doubt, that no real minor was exploited to produce the AI image. It would also, crucially, solve the alleged problem that law enforcement allegedly have in identifying real versus AI-generated PIM. Lawmakers could simply adjust legislation so that it only criminalized the distribution of PIM without metadata to prove it was entirely AI-generated, and there would be a 100% victimless release for MAPs, with full protection for real children.

The only real argument against MAPs using PIM generated entirely by a computer, and perhaps the main unspoken motivation for the support of its criminalization by the public, is that attraction to children (or teens) is seen as sick and disgusting; that is, prejudice. That kind of attitude should not have a place in the modern world, and should certainly not be used to write laws. It is also a denial of the reality that attraction to underage people, particularly teens, is not at all uncommon and cannot be changed. Seeking to criminalize every possible expression or avenue of sexual release, out of hatred and distaste rather than genuine concern about young people's safety, is not helpful.

The 'icky' question

And now for the difficult part. The distribution of sexual images made with LORAs, that inevitably represent a real person, is something I won't defend. Models are different in that their output represents an amalgam of their input, which is typically comprised of thousands of images, meaning that output does not represent any individual's likeness. If an artist made a model specifically for the purpose of producing PIM, which is not a common thing to do due to the skills and immense processing power required, this would likely involve using real PIM in the dataset. However, there would be no need for the person to take illicit photos to do this; they could simply train the model on old images that are already in existence and widely distributed; some of these go back as far the 1970s. While this will sound gross and icky to most people, it is not wrong from a logical harm reduction perspective. If high quality AI images could be created using such a model, the appetite for the production of real PIM would be reduced. And crucially, images generated using that model, but not a LORA, would bear no direct resemblance to any of the people whose faces were used in the model. It may feel gross, but if protecting young people is the true goal, perhaps the use of specialized models to create PIM ought not to be condemned and criminalized.

The IWF cannot be trusted

The IWF is a government-funded organization that was originally tasked with taking down PIM sites or otherwise blocking access to them from the UK. In a world of darknet distribution and AI, their role is becoming increasingly irrelevant. The German authorities are much more competent at taking down darknet sites; their skillful operation of hundreds of Tor nodes has been quite effective at identifying the real location of such sites and even identifying a small number of users. Meanwhile, the nature of contemporary PIM, being produced by teenagers themselves or by a computer program, is very different to the public's idea of PIM being the violent abuse of a real small child. These realities should lead the public to question whether all the money being thrown at the IWF is a good use of funds, and the IWF is surely aware of this. Their bid to spread panic over AI images is, most likely, a desperate ploy for survival in the face of increasing irrelevance.

It is not the first time the IWF has faced such a challenge to its need to exist. Shortly before the widespread availability of AI image generators, the foundation reacted in horror to the rise in self-generated 'images of abuse'. There were definitely grey areas. Adults or other minors tricking boys into performing sexually online, then spreading the videos for thousands of strangers to see, was far from okay. Again, that kind of behavior is abusive whether the victim is an adult, teen, or child. But just as the IWF now seeks to conflate genuinely abusive use of AI with undeniably victimless examples, they did the same with 'self-generated', by intentionally spreading confusion over genuinely self-generated videos and those where young people were tricked.

The IWF is not the upstanding organization it claims to be. Its main role now, rather than catching people who are harming children, is to spread panic over better alternatives to images in which real people may have been abused or exploited. It is an organization causing overall harm in its desperate bid for continued existence.

The thin end of the wedge

A number of countries have laws against AI PIM, but the UK has been the most zealous. Not only are images themselves banned, but there are demands to criminalize models that can be used to make AI PIM, tips on how to make AI PIM, and a proposal to criminally punish companies that fail to prevent their software being used to make AI PIM. While we do not know exactly how the legislation will be phrased, the overall picture is alarming. It is quite possible that those proposing the new laws do not understand how the technology works, but there is also a much more sinister interpretation.

As discussed above, a standard Stable Diffusion model can be used to generate PIM; therefore, any person offering any model for the software would be guilty of a sex crime if the law were phrased in a way that did not require specific intent. And, as the training set for a model is not visible to anyone investigating the model, it really is not possible to decipher which models are intended for making PIM, and which are not. Thus, a proposal to criminalize models that can be used to make AI PIM requires either throwing every model author in prison (because their model could make PIM!), or relying on a suspect admitting that their model's purpose is for making PIM. There is no logical way to enforce such a law without mass arrests and interrogations.

The push to criminalize offering software that can be used to make PIM is technically possible, but it constitutes an outrageous overreach by the UK. Software is distributed globally, to the UK and to countries that do not criminalize computer-generated images, and so any law criminalizing such software represents the UK forcing its values upon the entire world. These 'values' are not universal; respectable countries like Japan take the approach that real people need to be protected, but computer generations of non-existent minors do not. Given the very poor excuses made for criminalizing AI PIM, described above, the UK's actions in this regard represent an act of gross cultural and legal colonialism.

The UK has long used the fear of MAPs and 'child protection' to justify massive governmental overreach. When legal changes were proposed to allow police to demand encryption keys under the threat of imprisonment, the public were told that these powers would be used for the purpose of 'catching pedophiles and terrorists'. Privacy campaigners reacted with alarm, warning that the scope would be much broader. The government proceeded with its meaningless consultation, where it brushed off all of the legitimate concerns, and pushed the law through parliament. Its first use was to incarcerate a paranoid schizophrenic, who was told people would consider him a pedophile or terrorist if he didn't give up his encryption keys.

More recently, the UK's Online Safety Bill, which eventually passed into law as the Online Safety Act, drew sharp criticism from privacy and technology specialists. The original version of the bill contained a provision for the UK's 'super-regulator' Ofcom to force companies to remove 'legal but harmful' content, under the threat of severe financial penalties. Eventually, this proposal was removed from the bill due to backlash, but an alarming power remains. The Act requires service providers (including private messengers such as WhatsApp) to use government-sanctioned technology to scan for 'child abuse material', according to the demands of Ofcom and under the watchful eye of the Secretary of State. This, of course, gives the UK government the power to view all private communications without a warrant, using the excuse of detecting criminalized images of children. Not only does this apply to companies based in the UK, but also to companies that have a 'significant' user base in the UK. The implication is that the UK government would have the ability to snoop on users internationally, rendering end-to-end encryption protections useless, with no oversight from the courts. It is a truly horrific abuse of power, and one that affects people all over the world.

The British establishment's war on AI PIM doesn't make sense from a child protection perspective. However, it does make sense if interpreted as an attempt to protect and expand the powers of its various regulators, and to push through horrible laws that give the UK the power to spy on the global public. The war on AI PIM is 'necessary', for these nefarious purposes, in a world where illegal images of young people are rarely made by adults and therefore no longer suffice as an excuse. Unfortunately, many members of the public support the war out of ignorance and a deep-seated hatred of 'gross pedophiles'. In doing so, they are allowing their privacy and freedom to be heavily compromised by a state that lies its way toward ever more draconian powers. There should be more resistance to governmental overreach, and that can start by questioning whether proposals are really about protecting children, or whether they are excuses to pry.

Thoughts? Feel free to comment in our dedicated forum thread.