Generative AI in a Consent-Based Subculture
The tech marketed as "AI" is a huge player in our current political, economic, and cultural moment. It's gotten a massive PR push over the last few years, often in misleading terms, to solve any problem or create any content you might want. Any given program or website touts new "AI" functionality, and folks create and share images and text they generated with these systems. Increasingly, I've seen generative AI content shared openly in kink spaces as expressive writing, event marketing images, and so on.
Many people oppose generative AI, from many angles. Often, criticism is expressed hotly in the moment, or else in very broad theoretical terms. My hope here is to sketch out the issues I see with these generative AI systems, focusing on their deeply non-consensual economic basis. I come to this as a data professional, working with and understanding big data models for a living, and having dabbled with the generative systems myself out of personal interest.
This isn't directly targeted to anyone specific, and it's not an attack on people who have used these systems. I'm not here to police behavior or tell people what to do. We all engage with harmful systems, both knowingly and not, and there are complex personal decisions that inform when we choose to do so. However, the case against genAI in the kink context is very straightforward to me, and I hope to at least inform and spur some reflection by laying it out.
How AI models work, briefly
At its core, an AI model is an approximation machine. It contains millions or billions of tiny math problems, which chain together to transform input data to output data. If you knew exactly the relationship you wanted between inputs and desired outputs, you could manually choose the parameters - the numbers involved in those math problems - to fit your expectations perfectly.
The benefit of an AI model, though, is that its parameters can be set through a process called machine learning. First, the parameters are set randomly, and the model will produce gibberish. Then, the model is "trained," a process of showing it a bunch of "correct" inputs and outputs and adjusting the parameters slightly each time, bringing the model's outputs closer to the targets. Each piece of training data leaves a faint trace of itself in the the model, and a well-trained model can reproduce complex patterns found in the training set. In a way, an AI model is a way to compress a huge amount of training data, and prompting the model is analogous to uncompressing that data, semi-randomly, with some overall loss of fidelity.
Generative AI
Most of the AI/ML systems in use before the last few years were focused on math problems, predicting and analyzing complex systems. If a fast food chain trains a model on years of sales and corresponding weather data, it can be used to predict future sales for various potential weather conditions. In data-rich contexts, training a model is often faster and cheaper than human analysts modeling these relationships manually. These analytical models can be critiqued on their own grounds, but they're not my interest here.
More recently, we've seen massive investment in generative AI, use cases where an AI model is trained to create some kind of content. Large Language Models (LLMs) like ChatGPT compress patterns in language, then repeatedly solve the problem of "what word or letter most likely comes next?" based on the prompt and their own output. Similarly, image generators like Midjourney take text prompts and generate images pixel-by-pixel. Audio generation, video deepfakes, and other use cases also fall in the genAI category, but are less mainstream at the moment.
When you prompt a genAI model
It's easy to talk about these models in purely theoretical terms; their marketing encourages thinking of them as magical black boxes for problem solving. But like all software, there's a physical basis that allows them to operate. Many contemporary genAI models are too massive to run (let alone train) on any computer an average person can access. Instead, prompts to ChatGPT or Midjourney are sent to a datacenter, likely operated by the massive tech companies who fund genAI development, to be executed across an array of specialized server computers.
It can't be overstated how expensive this is. These models require incredible amounts of electricity to operate, and even more to train; they are restarting nuclear power plants just to feed genAI datacenters. Cooling the servers requires clean water, competing with communities for potable water supply. A whole industry of specialized chips are produced just to run these models marginally faster. In many cases, the user pays for all this plus profit directly, through subscriptions or paid credits. Some of the services are offered free or cheap (for now), following the familiar tech playbook: operate at a loss to acquire users and kill competition, then later jack up prices when folks are dependent on your service.
There are lots of arguments to oppose the use of genAI just from this overview. How many degrees of global warming is it worth to generate "Colonel Sanders in a ballgag" on demand? How much richer is it worth making the tech billionaires who profit from genAI - in the moment where they're explicitly welcoming hate speech onto their platforms, raiding the federal government, and giving Nazi salutes on live TV? These environmental and economic consequences are direct results of individual people using genAI models, paying for them and sharing their outputs, and can be directly mitigated by those individuals refusing to do so.
Many people find these compelling as they stand, and there's a wealth of AI-critical writing online if you'd like to follow up on these. But for kinky and sex-positive folks, I believe there's a more pointed consent issue to consider.
How genAI training sets are built
Training any "AI" model takes a ton of data, especially for models covering broad subject matter. For general-purpose genAI models, huge datasets of text or images are required, and they're largely sourced by scraping the web. Blog posts, news articles, official publications, and social media are downloaded and used to teach LLMs what text should look like. Photographs, diagrams, paintings, and other images posted online are encoded in the parameters for the visual models. According to the corporations who train these models, it's impossible for them to make broad-purpose genAI models without scraping other people's content indiscriminately.
At best, genAI corporations are indifferent to the rights and wishes of the humans creating this scraped content. In many cases, they willfully ignore them and outright lie about their methods. GenAI scrapers impersonate organic users, publish misleading information on how they can be identified, and outright ignore the "robots.txt" file (a thirty-year-old standard for communicating limits on automated web traffic). And when platforms make deals to sell user data to genAI corporations, they ban users who attempt to delete or edit their own posts to avoid scraping. While there are some niche genAI models trained with permission, they don't include any of the major players.
Why care about scraping
So genAI tools more or less universally rely on other people's content, openly taken without permission. You might think, as some genAI advocates argue, that this isn't fundamentally a consent issue - because genAI is sufficiently transformative of its input data, or because the publicly posted content is "fair game," or that the benefits of genAI justify the practice.
The most obvious issue is that genAI models plagiarize constantly. From the perspective of their designers, plagiarism a feature, not a bug - your LLM is at its most "convincing" when it reproduces real human writing verbatim. Because of this, there is no major genAI system with meaningful safeguards against plagiarism, and the models regularly output other people's content directly. This is at its most concerning with photorealistic genAI image models, which may reproduce features or whole images of real people from photographs in the training data, without their knowledge or permission - plagiarizing human bodies. Reproducing other people's content or bodies directly, without attribution or compensation, for profit, is an undeniable consent issue.
You can also look at the direct impacts of genAI on the human creators it steals from. Artists and writers are now directly competing with the corporate models trained on their stolen work. This affects already-precarious fields such as journalism and erotica, but also disproportionately targets early-career and amateur artists who rely on commissions or patronage to develop their skills and careers.
When forum discussions and social media are scraped, the goal is to capture freely-accessible communities of knowledge into a paywalled corporate service. Instead of asking a community of people for travel recommendations or help with a DIY home repair, genAI companies would have you prompt an LLM - one trained on those same community discussions, turning spaces of free communication and knowledge sharing into a corporate product with a hefty price tag (even if the user isn't paying full price yet).
If you're not compelled by these issues, there's another way I think of it. By the developers' own admission, these systems require mass scraping to be trained for broad use. If scraped content is so integral to genAI products, then the humans who create it are necessary participants in the system. And in the same way that folks need to consent to their involvement in our play, the people whose labor is necessary for genAI systems have a right to opt out. Fundamentally, genAI as an industry only works by steamrolling that right at a massive scale.
The consequences of genAI in the scene
What does this mean for the average kinkster who might prompt a genAI model? Again, start with the material impact: every time you prompt a genAI system, a datacenter computer spins up to process it; coal is burned, fresh water is used, billionaires are enriched. The connection of consumption and production is much more direct than a mass-produced physical product, where your purchasing choices have to ripple up the supply chain to have a material effect; there is a one-to-one and near-instant correlation between using the system and its harmful outputs, and choosing not to use it has an immediate positive impact.
The consent issues are exacerbated by how you use genAI content. If you're running a business, selling a product, or even just advertising a paid event using genAI, you're generating value from the plagiarism and non-consensual extraction of other people's creative labor. And if you paid to generate that content, your business is now subsidizing that extraction directly. Anyone doing this in the kink or sex-positive spheres should re-evaluate their business practices to focus on more consensual and ethical ways of sourcing content.
But even if you aren't making money, posting genAI material sends a message. It normalizes the use and (frankly ugly) aesthetics of these harmful systems in our subculture. It competes with the content it plagiarizes for attention and interest, while often smuggling in bias and stereotypes from its training data. Intentionally or not, it signals an alignment with the corporate plundering of the sorts of artistic and educational labor that make our scene function. Sharing genAI content communicates that it has a place here, and I think it's important not to do so.
A final thought
If anything should be apparent from the last decade, it's that the big tech companies will not be on the side of the sex-positive or kink subcultures in the long run. The work that keeps the scene going is in communities of knowledge, reflective and educational writing, art that inspires and turns us on. Much of this happens in person, but most people encounter some form of it online first. The heavy push to normalize genAI in our mainstream culture is a calculated business strategy to consolidate this type of online production under corporate umbrellas, and there's no reason to think it will continue to serve our subculture long-term.
We owe it to each other, and ourselves, to push back. To celebrate human creativity, even at its most amateurish. To develop our own skills at writing and art. To continue to highlight our educational communities of shared knowledge. To show the folks among us doing the work that we'll stand with them against big tech plagiarism. It's an easy choice to make, at any time, and I hope we make it together.