Kink Or AI?
If you spend too much time on the ‘booru image boards, you’ll inevitably find yourself squinting at some sex scene with multiple people in it and asking yourself “why is he sticking his eleven-inch horse cock in her ear?” or “Is that goblin fucking her armpit?”
It used to be that the answer was always “Wow, I’m impressed by the artist’s creative perversity, I never would have imagined putting such a penis in such a place under those particular conditions.”
But nowadays? Well over half the time, the answer is “image was AI generated, model doesn’t know that dicks don’t normally go there or bend like that.” Which for me is always a disappointment. Knowing that a sexual idea was important enough to someone for them to intentionally make art about it is, for me, an important precondition for engaging my own erotic imagination. How would that be? What would that be like? Would it be fun to do, or to watch? Under what fantastic or science-fictional conditions would such at thing even be possible? I’m not going to waste any of that mental effort on AI-generated mistakes.
Yes, I do believe this is a special case of the general principle “If you couldn’t be bothered to write it, I can’t be bothered to read it.”
Similar Sex Blogging:
Shorter URL for sharing: https://www.erosblog.com/?p=33655
I strongly appreciate this opinion. When I’m looking for fiction of any kind, I want to find work whose creator has thought about the details and made intentional choices.
Even handcrafted material that is “low quality” on its face is superior to the stiff, generic, averaged output of machine learning content generators.
And that isn’t even getting into the problem of theft inherent to machine learning content generators!
All of this applies to pornography as much as any other type of fiction.
Thanks, J. I’m rapidly coming to despise the “stiff, generic, averaged output” and rather like that phrase.
I still haven’t drank the consensus Koolaid on the input side, though; I’m not yet convinced about “the problem of theft”. In literature, for example, we despise plagiarism but glorify the ancient remixed edifice of tropes and themes and basic plots that are the literary building blocks from which all new works are constructed. Artists do it do; nothing is genuinely novel, every new work is in conversation with the elements of all the other art that the new artist has ever seen. And that’s what generative art looks like to me … more like the cultural remixing that human artists do, and less like plagiarism or theft. There’s even an art form, called collage, that consists of cutting up photos and illustrations and pasting bits of them back together to make new works. It’s not for everyone, but you can find collages in museums, and nobody calls it theft.
I’m not making a final determination or saying this is my final opinion, but everyone seems to have rushed to judgment on the “theft” question and I think it’s very much still open.
NOTE, when I say AI I mean ‘generative AI’.
I think there are a couple of key elements that fuel the theft argument. The first is the lack of control of how an artwork is used by the model or those who use the model. Secondly the absolute arrogance of the tech companies who have never engaged with those whose works they scrape.
In the first case, copyright provides protections for creators from simple replication and exists to ensure the creator is the one rewarded for their work. There are specific allowances for things like satire or modification which exist to allow precisely the inspiration input to new artworks that you describe. Current AI systems can, and do churn out inputs complete and unchanged. This is rare, but frequent enough to be recognised as a problem. Some AI companies have coded their IO layer to prevent this, but these layers are fragile, and every day reveals new examples of the safeguards in these layers being circumvented to allow immoral or illegal usage of AI models. Whether that’s to create political material or to strip the clothes off children or celebrities, or to create a computer virus. It’s not possible to prevent these circumventions partly because they rely on parsing natural language, and it seems likely these layers cannot be resilient or robust, assurances from AI companies aside. Therefore it seems impossible to defend the current copyright privileges of a creator short of excluding their works from the underlying model. Including works in the generation of a model allows for the possibility of them being replicated in outputs to a level of fidelity that breaks copyright as it stands. To be extra clear, while the inputs do not exist in the models in original form, they exist in sufficient description to be replicated as an output. If you specified the prompt adequately, the only possible output would be the input image because every decision about including content would select for the probability associated with that single input.
So far the UK government have already started legislation to allow AI a specific exemption from copyright. Trump rescinded all Biden’s controls on AI already, so the US is not far behind.
As regards the second, AI model creation companies use data sets scraped from the web before anyone knew about the creation of these systems. Any data published before the inception of the Robots directive to prevent AI scraping may be in there. Since then tech savvy creators have been able to tell crawlers to ignore their works on their own sites, but there are no protections on third party sites, like blog providers, social media, community sites etc. It has also been flagged up that some of the data set scrapers ignore the Robots directive, which to be fair is advisory to a crawler, and not a legal requirement (but should be). The data set companies make it hard to remove material from their data sets, usually requiring you supply URLs or works to them for exclusion. And if you believe that means a work is subsequently always excluded, I have a bridge for sale…
The data set companies jealously guard their sources. We can safely assume that they have scraped the big names like Meta’s sites, Amazon and X, but anywhere else is unknown until or unless an input gets chucked out in recognisable form. The Laion 5b data set used by many image based generative AI systems contains Child Sexual Abuse Material, and researchers were able to prove beyond reasonable doubt that the creators had included this, possibly unintentionally.
The one thing the data set companies, and their customers, the model creators, are really bad at therefore is engaging with the copyright owners of material they wish to use as inputs. Whether their business model is legally theft or not it is parasitical, and like viruses or fleas they do not ask permission. Engaging with creators, allowing clear and easy opt out, and respecting it would go a long way towards removing objections. However like the parasites they model themselves on, they don’t believe in asking, and if they depended on it being given, they would become extinct. The scrapers are now reporting (according to Musk) that all public matter on the web has been scraped and included. To the point where they need fresh input faster than we can create it. Asking permission is going to kill them. Musk also asserts that AI models are now being fed their own output, or the output of other AIs and that is going to kill them just as quick.
AIs ‘make mistakes’ and ‘hallucinate’. To be clear AIs are NOT intelligent by many definitions. They are systems that work based on analysing context and using that to create likely responses to an input prompt. A coin flip is the SAME technology, albeit with very little in the way of model (either heads or tails with even probabilities plus gravity, air resistance, mass) and prompt (spin, impetus). According to all tests so far published if you feed an AI it’s own output, it becomes less reliable and outputs become more erroneous or hallucinatory. In the case of a coin toss maybe edge becoming a significant result, or rabbit (or something else random and neither heads or tails). This ultimately reduces the usefulness of these systems and may become terminal for the companies.
All of this means that those companies involved in creating or peddling AI systems need inputs and need them quickly. Engaging with creators and asking permission, or even just respecting their wishes is going to kill them. They have made some effort to prevent outright copyright violation by duplicating inputs in output but those protections can only be effective where a copyrighted item is NOT used in model creation. Apple are personalising this by feeding everything you do on a device into a local model. Microsoft it seems likely are just feeding everything anyone does in their systems into their Cloud based model. Google are feeding everything they can find in and Meta likewise while probably trying to block other companies from their own systems.
So while feeding a copyrighted item into an AI model is not theft in itself, it facilitates and makes possible copyright violation, which we treat as theft.
There already exist legislations that protect AI companies. You can go out and buy a cassette deck, or a CD burner, or a DVD or BluRay burner. You used to be able to buy an iPod, you can buy a computer, use a network etc. All of these devices are capable of copyright violation, but the companies which produce them are protected because they primarily serve a useful purpose and are not specifically invented for copyright violation. Even though I bet we have all probably used them to retain copyright material even if we have not then distributed it.
So are AI companies thieves for using stuff published on the interweb thingy? Not per se, but their systems can be used for theft. This is not new. Nor is it theft in itself, it comes down to how those systems are used. Are our noses out of joint because we’re not getting a cut of their billions? Yep. Did we get a cut of the revenue from any of the other technologies listed above? Heck no.
AI has the capability to further undermine the safety of creative professions. They change the skill created to produce an artwork. No longer do you need to know how to use a camera, lighting etc to produce a product image for advertising, AI can do it if you can describe the scene, and the model has enough inputs to create what you want. The key here is inputs. AI cannot invent or create, it can only put together things it was fed (usually) in new ways.
The theft argument is unlikely to slow AI companies down, or cause their CEOs to lose sleep. What is going to give them pause is energy. It takes a vast amount of computing power, and electricity to build a model. It takes an appreciable amount to create outputs, though that is getting smaller all the time. However at a time when we are all under pressure to reduce power consumption, to switch to off peak usage, to recycle and to personally take responsibility for our use of resources, AI is unregulated and projections are that AI systems will soon be amongst the heaviest users of energy globally, more than many countries.
So perhaps rather than encouraging new uses and forms of AI, the same politicians that are hair shirting us all over leaving a light on, need to be making AI companies responsible for their power usage, and telling us all ‘learn a new skill, or pay someone skilled or talented, don’t use AI.’ Because otherwise we’ll recommission Three Mile Island.
That help?
Weirdly I started writing that from the perspective of AI as thieves. I don’t want my art in an AI model. But writing it clarified my thoughts and the outcome was not what I expected or intended.