Fed Without Consent
The hypersexualised fantasy art flooding your feed isn’t just offensive, it’s also training the machine.
I didn’t ASK for this stuff on my BlueSky feed, and that’s the weird secret people keep overlooking when this conversation surfaces and gets immediately shouted down by communities who have a very strong interest in keeping it from surfacing at all; “You don’t go looking for this content, it comes to you, served directly into your feed because an algorithm decided that an interest in comic book art and an appetite for hypersexualised anime fantasy women are essentially the same thing, and the algorithm, indifferent as gravity, sees no reason to distinguish between them.”
What am I talking about?
I’m talking about a category of digital illustration that has colonised some social media feeds with such completeness that most people who follow comic book art accounts will encounter it daily without ever having sought it out. At one end of the spectrum sit the warrior women, fantasy fighters rendered with waists the width of a wrist, armour that protects nothing above the midriff, and camera angles positioned with the specific deliberateness of someone who learned to draw women from pornography rather than from life. The proportions are impossible, the poses are not combat stances, and the clothing exists to frame rather than cover. These images circulate under hashtags for fantasy art, comic art, character design, and digital illustration, which is precisely why they end up in the feeds of people who follow none of those communities specifically but whose engagement history has nudged an algorithm in a direction they never chose.
Further along the spectrum sits something more troubling still, a category known in these communities as moe, a Japanese term describing a cultivated aesthetic of innocent vulnerability in female characters. In practice this means characters rendered with deliberately childlike facial features, wide uncertain eyes behind oversized glasses, flustered expressions, the visual grammar of adolescent anxiety, combined with exaggerated adult bodies positioned in ways that make the combination unmistakeable in its intent. The infantilised face is not incidental to the appeal; it is the point. Legal scholars examining this content have noted that the artistic style often emphasises infantilised or youthful features in ways that raise significant legal and ethical challenges, with some jurisdictions now treating fictional or animated depictions on par with real child sexual abuse material. That comparison does not arrive in legal discourse lightly or carelessly; it arrives because the researchers making it believe the evidence demands it.
This content is not hidden in dark corners of the internet behind age verification walls and content warnings, though some of it exists there too, monetised through Patreon subscription tiers that have spent years refining policies to manage what their own platform acknowledges is a serious and growing problem. Patreon’s most recent policy updates explicitly prohibit AI-generated or digitally altered depictions of minors and clarify that in animated or illustrated adult works, subjects must be unmistakably represented as adults, with expanded guidance on how the platform assesses that. The fact that they needed to write that guidance at all, to specify it explicitly and expand it in response to what they were already seeing on their own platform, is its own confession about the scale of what was already there.
The audience for this content is not a fringe curiosity either. The BBFC, which now classifies more anime than ever before for physical media release in the UK, commissioned research in 2025 finding that nearly nine in ten people believe anime poses a child protection risk if not age-rated appropriately, with anime now accounting for nearly a quarter of all content classified for DVD and Blu-ray release. A quarter of all classified content, and that is only what gets formally submitted. The volume circulating online, unclassified, unmoderated, and algorithmically amplified into the feeds of people who never sought it, is immeasurably larger.
The communities producing and sharing this content operate with considerable sophistication, cross-promoting across platforms, building followings through safer adjacent material before migrating audiences toward more explicit work, and using engagement signals to ensure the algorithm treats them as legitimate creative communities rather than what they frequently are, which is networks organised around the sexualisation of female characters rendered with varying degrees of deliberate ambiguity about age. When anyone criticises any of this, the response arrives with a speed and coordination that should itself be instructive. The “it’s just a drawing” defence deploys within minutes, and the harassment of anyone who persists follows shortly after, calibrated to ensure the cost of speaking is high enough to deter whoever comes next.
Anita Sarkeesian, the media critic who produced the first widely seen video analysis of hypersexualised female characters in gaming and adjacent visual cultures, faced years of rape and death threats for doing so, a response that tells you everything about a community that understood exactly what was being said about it and wanted the price of saying it to be prohibitive. That chilling effect has largely worked. For years the mainstream conversation tiptoed around this territory, framing it as a matter of taste, a cultural difference, a question of artistic freedom that reasonable people might disagree about, when what it actually is, is a community producing content that depicts women and girls as objects of sexual consumption, distributing it at industrial volume through platforms that profit from the engagement it generates, and ensuring it lands unrequested in the feeds of people who followed a hashtag about Dredd or Saga or Sandman.
The Ghost in the Training Sets
What elevates this beyond a question of individual discomfort, and connects it directly to the argument running through Drawn to Extinction, is what happens to all of this content next. To understand that, it helps to strip away the language of magic and get back to mechanism, because the way generative image systems actually work is both less mysterious and more alarming than the marketing suggests.
These tools do not work like inspiration, and they do not browse a few comics, absorb a mood, then head off to do their own thing. As I describe in the book, they ingest. They disassemble. They reduce millions, sometimes billions, of images into patterns, relationships, correlations between pixels and language. Captions, tags, filenames, descriptions, prompts, all of it gets folded into a vast statistical architecture. The process is less like learning to draw and more like industrial digestion.
The datasets that made this possible were not built with care or discernment, they were scraped, bulk-collected by automated crawlers roaming the open web, hoovering up whatever they could reach with very little regard for origin, consent, context or copyright. The most notorious example, LAION-5B, assembled more than five billion image-text pairs pulled from the internet at scale. Five billion, a number so large it almost sounds abstract until you think carefully about what it contains: portfolio pieces, sketches, fan art, commission previews, character sheets, splash pages, drafts, half-finished experiments. A vast percentage of the visual culture that people uploaded because they wanted to be seen, not consumed by a machine. As I write in Drawn to Extinction, the digital equivalent of looting a museum, grinding the collection to powder, and then selling new objects made from the dust while insisting nothing was actually taken because the original paintings are no longer visible in the final mix.
And here is the part the communities sharing hypersexualised anime characters across social media never get asked about. That content is in the mix too. Every moe character sheet, every cheesecake fantasy warrior, every image produced by artists who have built profitable Patreon followings around the sexualisation of female characters, all of it is being scraped and ingested alongside the work of creators who spent careers trying to push the medium in a different direction. The machine does not sort them. It does not grade them by intention or ethics or artistic seriousness. It processes them with identical indifference, and what it learns from that processing becomes the statistical foundation for every future output.
When you ask one of these systems to generate a female superhero, it does not consult an artistic tradition or make a creative judgment. It reaches into that statistical foundation and finds what is most frequent, most reinforced, most heavily represented in the data. And what is most heavily represented, given the volume of content these communities produce and the algorithmic amplification that ensures it spreads, is the template: impossible proportions, minimal clothing, the camera angle that treats a woman’s body as a landscape to be surveyed rather than a person to be depicted. The model is not making a sexist choice. It is faithfully reproducing a sexist consensus, encoded into it by the sheer weight of what it was fed.
Lesley Gannon, Deputy General Secretary of the Writers’ Guild of Great Britain, named the mechanism precisely when I spoke with her for the book. “We know that algorithmic behaviours intensify the bias of the material they are fed,” she said. “Not just perpetuate it, intensify it. That means disparities worsen. Stereotypes worsen. Harmful attitudes are amplified.” The critical word there is intensify. The model does not simply mirror what it ingests; it amplifies whatever patterns are most statistically dominant, which means the communities producing this content at industrial volume are not merely adding to the dataset but actively warping its centre of gravity. Every new image produced, shared, reposted, and scraped makes the default output more extreme, not less.
There is a feedback loop operating here that is worth sitting with, because it is genuinely without precedent in the history of visual culture. In the past, a harmful visual convention, say the hypersexualised depictions of women that ran through mainstream superhero comics for decades, could be challenged, critiqued, and gradually displaced by the work of artists and editors willing to push back against it. That process was slow and incomplete and is still ongoing, but it was possible because human beings made editorial decisions and human beings could be argued with. The machine does not argue. It does not respond to critique or evolve through cultural pressure. It responds to data volume, and the communities producing moe art and cheesecake fantasy characters are winning the data volume argument by an enormous margin, publishing new content daily across dozens of platforms, all of it feeding back into the systems that will generate the next wave of outputs, which will in turn be scraped, ingested, and used to train the generation after that.
As I also point out in the book: “When image models scrape ‘comic art’ they do not understand which images were products of their time and which were attempts to challenge that time. They absorb it all alike. The hyper-sexualised heroine pose. The monumental male body. The racial caricature. To the model, none of this is ideology. It is pattern. Frequency. Dominance in the data.” The machine does not invent new prejudices so much as perfect old ones, making them cleaner, faster, and infinitely reproducible, fossilising bias by mistaking repetition for legitimacy and turning outdated fantasy into default syntax.
Torunn Grønbekk, the Norwegian comics writer whose work I discuss at length in Drawn to Extinction, tried using image generators as creative reference tools and found the reality considerably uglier than she had anticipated. No matter how she phrased the prompt, she told me, every “woman” came back as the same kind of creature, “could be sixteen, could be twenty-seven, lips slightly parted, permanently on the edge of a pornographic close up.” When she asked for an older woman, the system returned cartoons of decay, no beauty, no complexity, just a drained husk with all specificity removed. She was not describing a malfunction; she was describing a system working exactly as trained, faithfully reproducing the visual assumptions encoded into it by the communities that produced the data. “It has nothing to do with women in any possible way,” she said, and the directness of it cuts to something important: she was not speaking about aesthetics or artistic tradition but about a fundamental failure of recognition, a machine that has learned to conjure the shape of a woman without any understanding of what a woman actually is, because the data it learned from was never really about women either. It was about the male appetite for a particular version of them.
Dr Julia Round, one of the foremost academics in British comics studies, described the underlying structural problem to me as an impossible ethical dilemma, and when I analysed my own comic book collection I found the numbers that make it concrete: only nine percent of credited contributors across decades of material were women. Nine percent, before a single piece of fan art was ever scraped, before the moe artists and the gacha game character designers and the cheesecake commissioners added their volume to the pile. “If the machine only learns from what is already recorded,” Julia told me, “then the invisible will stay invisible.” The creators who spent careers expanding what bodies, identities, and emotional presence could look like on a page, Nicola Scott, Jen Bartel, Sana Takeda, Fiona Staples, all of them working in the same digital spaces as the communities producing hypersexualised fantasy women in their thousands, all of their work swept into the same datasets together, and as I write in the book, “the noise of the past can drown out the nuance of the present.” That is the quiet catastrophe that nobody producing these images is ever asked to account for.
The piece of this that the mainstream technology and comics press has been collectively reluctant to state plainly is that the audience for this content bears direct responsibility for what those training datasets contain, not the platforms alone, though the platforms are culpable, and not the algorithms alone, though the algorithms operate with an indifference to harm that urgently requires regulatory attention, but the people who built communities around the sexualisation of female characters, who refined over decades a visual vocabulary in which women exist primarily as arrangements of flesh designed for consumption, and who ensured that vocabulary is now so prevalent online that it is essentially inescapable for anyone who follows the art form those same communities claim to love. As I write in Drawn to Extinction, the scraping is not neutral, the ingestion is not neutral, and the uncritical resurrection of prejudice through training data is not a technical accident. These are decisions, and somewhere behind every decision there are people who made them and communities who kept producing the content that made them possible.
The machine does not invent new prejudices, as Lesley Gannon made clear; it intensifies the ones it is fed. The people feeding it know exactly what they are doing, even when they insist, loudly and with considerable coordination, that it is just a drawing and you simply don’t understand the culture. The algorithm served you that image because a community worked hard to make sure it would, and the machine is learning from every single one.
Drawn to Extinction: Comics, Craft, and the Battle for Originality in the Age of Ai is available now in hardback from Amazon, Bookshop.org, and comic shops via Ingram.