Variations of Slop
How AI content is reshaping the feed
Hello.
You are reading Understanding TikTok. My name is Marcus. I should be in Berlin at re:publica right now discussing TikTok and Germany’s far-right but I am in bed sick, and my FYP is infested with generative AI content. Hundreds of similar videos: I saw AI bot armies waving German flags but glitching mid-sentence, two-headed women in bikinis and thirst-traps built on racist stereotypes. The genres go darker than the three examples I'll unwrap here.
Slop strategies have been evolving ever since Jesus sat down on a huge shrimp in 2024. AI content production is now sophisticated enough to invent new affective hooks that did not previously exist.
New genres are emerging that exploit cognitive and moral resources in ways that did not have technological pathways before. Let’s do some unwrapping…
Example 1: AI Scam Aesthetics
I’m in my thirties. I live alone in the German countryside. My parents have died. I have no relatives, no friends. I inherited a large fortune and live off field work. I just wish for someone who treats me sincerely. An age difference of twenty to thirty years doesn’t bother me. I only want to have children and build a beautiful family together.
Almost nothing in this video adds up if you watch it the way you’d watch a home video. The visuals show rural China — Chinese New Year couplets on doorways, traditional wooden architecture. The voice insists on rural Germany. The audio describes a life of field work; the body has manicured nails, no strain, no sweat, and poses that real heavy labour wouldn’t permit. The voice claims thirties; the face reads early twenties at most, beauty-filter-smooth, possibly composite or fully synthetic. The voice itself is machine-like text-to-speech.
And yet the video works: as fantasy object, scam bait, meme material, AI curiosity, and disposable entertainment all at once. As Jason Koebler put it, audiences are being trained to value vibes over facts (Koebler, 2024). Compare Tom and my piece on Thirst Trap Propaganda (Bösch & Divon, 2026).
But how are the videos built to work this way?
Steyerl: what the image is
The artist and theorist Hito Steyerl calls what we’re looking at a mean image. The pun does most of her work. Statistical mean: the figure in the frame isn’t a person, she’s an average — a convergence of beauty-filter parameters, ethnic styling cues, Douyin colour grading, possibly a composite or fully synthetic figure draped over scraped rural-lifestyle footage. Asking whether the woman is “real” is the wrong question, because she was never trying to be anyone in particular. She is a type.
The other senses of “mean” pile on. Shabby: cheap content for a demographic the producers code as low-prestige. Nasty: aimed at lonely older men. Mean as in means — somewhere off-screen, workers script the broken German, run the text-to-speech, source or commission the footage, run the composite pipeline. If real rural lifestyle footage has been scraped to set the scene, then the surface plausibility rests on the unconsented visibility of people whose labour is being used as set dressing.
Steyerl places these images in a much longer history. The exoticised, emotionally accessible Asian woman is a recurring figure in Western visual culture, and the operation of producing such types through statistical aggregation goes back to nineteenth-century photographic typologies — Galton’s composite portraits, used to manufacture supposedly objective images of “criminal” or “racial” types. The technologies are different, but the operation is the same (Steyerl, 2023).
Meyer: how the image functions
Roland Meyer argues that AI-generated images are optimised not against the real world but against other images already in circulation. They are trained on billions of platform images and rewarded for matching what viewers already expect to see. Meyer, building on a term from Jacob Birken, calls this platform realism — a realism that has nothing to do with reality and everything to do with image-familiarity.
For our scam video, the visuals aren’t documents of a place; they’re illustrations of a concept. Meyer argues these images traffic in vibes. What’s being copied isn’t Chinese rural life; it’s how Chinese rural life is already represented on platforms.
The artifact operates on familiarity rather than evidence. It sidesteps the entire logic of truth and falsehood because it makes no definitive claims. Because nothing is being asserted, nothing can be debunked. The viewer is left to supply the meaning—which is exactly why visual contradictions don’t ruin the clip. The video cannot be contradicted if it never made a promise to be real in the first place.
Meyer adds one more thing worth keeping. AI-generated images, he points out, are “disposable by design.” Producers generate dozens to find one that engages. Most are never seen. The figure in our video isn’t a stable character whose existence could be checked. She is a test image — one of many, kept because she happened to perform. (Meyer, 2025)
Toister and Zylinska: what happens between image and viewer
So far we have an account of what the image is (Steyerl) and how it functions (Meyer). What’s still missing is what happens when these images meet an audience.
Yanai Toister and Joanna Zylinska (2025) give us a useful concept for that part: cognitive hacking. Every time you see an AI-generated image — even one you know is fake, even one you mock or scroll past fast — the type it represents becomes a tiny bit more familiar to your imagination. Across thousands of variants seen by millions of viewers, this slowly changes what feels normal, what feels imaginable, what counts as a recognisable shape in the world.
The real damage of the scam video isn’t only to the individual viewer, the damage is to the collective imagination. The category being installed — available rural Asian woman seeking older German partner — becomes a more available image, a more familiar shape, regardless of whether any individual viewer believes any individual video.
This also explains why the video works across such different modes of reception. Sincere belief, ironic mockery, erotic interest, critical analysis, scrolling indifference — all of them feed the same engagement metrics and deposit the same type. The artefact doesn’t need any particular response, just exposure.
Example 2: Affect Bait
When Hito Steyerl wants to show what a mean image is, the central illustration in her essay is a Dreamfusion 3D squirrel that the model rendered with three faces — unable to decide which “front” was right, statistically committing to several, rendering all of them. The two-headed woman on the beach is a successor, born out of equally valid completions in latent space, but rendered as a new affective genre.
Where this body comes from
The two-headed beach selfie isn’t a random distortion. The model fused three image traditions that already existed.
The sideshow archive: The two-headed or conjoined woman has been a fixture of Western popular culture since the nineteenth century. Daisy and Violet Hilton, conjoined-twin performers, were major celebrities in 1930s American film. The cultural template — two pretty girls sharing a body, framed for fascinated looking, often coded as sexually available — was already mature a hundred years ago. The disability scholar Rosemarie Garland-Thomson has written about what she calls the staring economy: the gendered, often unequal dynamics of looking at non-normative bodies as spectacle. The model has scraped a hundred and fifty years of that economy from photographs, film stills, posters, tabloid covers.
The beach-influencer selfie: The compositional grammar — golden hour, palm trees, glossy skin, arm extended, hyperreal light — is contemporary Instagram and TikTok in its most mature stock register. The model knows what hot girl at the beach looks like because it has seen tens of millions of variants.
The disability-acceptance discourse: The third tradition, the one the audio activates. Social-media campaigns, grassroots writing, the moral vocabulary of inclusion: real discourse developed by real people doing real political work over decades.
The artefact is legible because all three traditions are recognisable to anyone scrolling a feed. It is manipulative because the audio activates the third over a body assembled from the first two. The moral language built up around real stigmatised bodies is being used to extract empathy-performance from viewers looking at a beach influencer who happens to have two heads.
What the audio does
I just wanted to have a friend who has no prejudices — could you be that friend? Press the red cross on the right if you would accept me.
The two-headed body is being positioned as a stigmatised body asking for acceptance. The artefact performs minoritised-body discourse — the moral language built up around real marginalised people — over a body that doesn’t exist, and instrumentalises that language as an engagement-extraction tactic. Mean in the sharpest sense Steyerl gives the word.
This wasn’t a one-off artefact. The same account has since moved on to a different setup — a child dying in a clinic, parents crying, the same kind of affective pitch with a different body underneath. The two-headed bikini wasn’t a singular experiment; it was a phase. The production logic is now visible: cycle through different moral-emotional registers, test which ones convert viewers into followers, swap the body and audio when engagement flattens. The model can synthesise whatever body the affect-bait requires.
Example 3: A machine in training
None of the people exist. The cathedral doesn’t exist. The crowd doesn’t exist. The clip is six seconds because that is roughly the default output length of current text-to-video tools. And the AI tells are everywhere. The most diagnostic glitch — the foreground woman’s right arm extends up and to the right ending in a hand with a pointing finger, while a second pointing hand floats disembodied in the upper-right corner of the frame, with no arm connecting it. The model held two interpretations of the pointing gesture and rendered both. Steyerl’s Janus problem, applied to political imagery.
One video out of many
This isn’t one artefact. It is one of many. I gathered some in this thread.
Scanning the platform, the same visual register recurs across dozens of accounts: elderly white Germans in decent clothing, pouring rain, Gothic cathedrals, families with prams, people in wheelchairs holding flags, marching crowds, lone figures in foreground addressing the camera. Several languages — German, English, occasionally with the audio glitching mid-sentence as if the model switched mid-generation. Several settings, all converging on the same iconography. This is the visual vocabulary of contemporary right-wing populism — ethnically homogeneous, anchored by Christian-heritage architecture, framed as endangered ordinariness.
TikTok’s interface logic shows you one video at a time. The recommendation system sequences them. Each clip looks like a curiosity, an outlier, a one-off. Stepping outside that logic and inspecting the accounts behind the clips reveals what the interface obscures: a cluster of low-effort accounts with random-string usernames — user9181718, kissable011 — running similar generations through similar templates, all converging on the same affective registers and the same political iconography.
What the cluster is doing
The toolkit from the earlier examples still applies, but the new thing this example shows isn’t visible at the level of any single artefact. It only becomes visible across the cluster.
The glitches are not just AI tells. They are the visible residue of testing. Different prompts, different parameters, different audio overlays — different operators producing similar variants of the same emotional package, with engagement metrics selecting which variants survive and propagate. What you are watching, when you scroll through the cluster, is a system iterating. A machine in training.
This is the same logic Example 2 showed at the level of one account cycling through affect-baits, but scaled up. Where one operator cycles through grief and acceptance and arousal in series, the cluster cycles through nationalist tableaux in parallel.
The infrastructural layer
The frameworks above treat the videos as images. There’s a layer they don’t reach, and this case makes it visible.
The media theorist Jussi Parikka calls this layer in-visuality: the image-work that happens before images become visible to humans at all. Moderation pipelines, recommendation algorithms, AI-detection systems, labelling infrastructures. The decisions about what gets shown to whom, with what warning, under what conditions.
The metadata for this video contains a field called aigc_label_type set to zero. That field is TikTok’s marker for “is this AI-generated content.” The zero means no AI label appears to the viewer. Whether the system actually checked the video and decided no, or whether zero is just the default that gets written when nothing has been checked, can’t be determined from outside. Either way, the user sees no label.
TikTok’s policy nominally requires creators to label realistic AI content themselves. The accounts in this cluster do not. The platform’s auto-labelling, designed for content carrying embedded AI metadata, also does not fire. Neither the policy nor the infrastructure brings the label into existence.
The recommendation algorithm operates on engagement signals — completion rates, shares, sound reuse — and TikTok has not documented whether AI-content status acts as a downranking signal in that system. What is documented is that the labelling pipeline runs separately, on different inputs. Whatever the recommender weighs, it doesn't appear to be slowing this content down. A six-second AI-generated piece of right-wing iconography that performs well gets amplified by the same platform that fails to mark it as synthetic.
A six-second AI-generated artefact ends up circulating alongside vacation footage and dance clips, with no infrastructural marking to distinguish them. Hundreds of similar clips across dozens of accounts, all carrying the same dormant field, all moving through the same amplification logic. The platform isn’t just hosting the cluster. It’s shaping the conditions under which the cluster becomes visible, legible, and ordinary.
Scam in the System
The three examples now line up.
The scam video showed slop at the level of reception — a hybrid composite holding together by familiarity, with the viewer supplying whatever absent claim made the contradictions matter less.
The trial-and-error engagement-bait showed slop at the level of production — a single operator cycling affect-baits, testing what converts viewers into followers, the body and audio swapping while the follow button stays in the same place.
This third case shows slop at the level of the system. Many operators, many accounts, many variants, all converging on the same affective and political registers because the platform’s selection pressure rewards convergence. Engagement metrics do the aligning work.
This is the operation I have been describing as postdigital propaganda. Not a campaign by an identifiable actor with an identifiable message, but an alignment produced inside the system itself: platform infrastructure surfacing what fits the feed, platformized affect rewarding what hits the feeling, participatory recursion across accounts that aren’t coordinating with each other but are producing the same thing anyway, and synthetic plausibility accumulating into a shared imaginary that doesn’t need any individual viewer’s belief to operate.
We are still at the beginning of a transformative technology whose effects we can barely measure — partly because the changes sit just below the threshold of attention. A weird video here, a strange video there, and yet each one is part of a process that is shifting how we consume media and what we treat as something ordinary.
What else?
I have been discussing Iranian Lego propaganda for ZDF Heute and ARD Weltspiegel if you like.
Thanks for reading. If you’re interested in tailored insights, workshops, consulting, or policy support, or just want to discuss stuff get in touch. Here is Linkedin. Here is Bluesky. Ciao







Thank you for your awesome work and your perspecitves!
I hadn’t read 'Mean Images' - thanks for that - Hito Steyerl never misses.