On Slop

John David Pressman

Since its introduction as shorthand to refer to unwanted AI content I've been using the word "slop" a lot. I use the word "slop" to refer to a lot of different things, many (most?) of which are not AI. Some samples from my BlueSky account:

Buddhist Bro Slop

Like apparently not everyone else shares my trait of my eyes rolling out of their sockets the minute Claude tries to rizz me with Buddhist bro slop and if I wanted I could just go around hacking peoples brains by saying this kind of stuff without a hint of irony or self awareness. Terrifying.

Psychopathic Slop

What I would have you notice though is that the 2nd wave was in part based on reaction to the first wave, because at some point the algorithm phase transitions edgy humor into jerking off on Shane Dawson's cat and you realize: Wait a minute, this is psychopathic slop this is disgusting what the fuck

ActBlue Slop and Cable Television Slop

The two parties in this country are ActBlue slop and cable television slop and all of the otherwise baffling decisions of their politicians are basically explainable in this framework.

Yuddite Slop

Reading the Arbital corpus and narrating my thoughts into the AI transcription app and guys I'm gonna be honest with you I'm so tired of reading Yuddite slop it's real mud moat hours up in here words and words so many words so hard to turn into a thesis.

Deathist Slop

"Resurrecting the dead with GPUs during the climate emergency is an irresponsible use of resources" is just the most incredible deathist slop and straight out of how I would parody these people but I can't because they are well and truly beyond parody.

Finance Slop

If we're being totally honest this take is clearly driven by finance slop media where shorting is portrayed as a dishonest or Machiavellian practice and most people just hate it on vibes because they think it rewards you for thinking like a social defector.

The Opposite of Slop

I mean in fairness this is top tier video essay content from the guy who does the design for Musecore. It's like, the opposite of slop.

"The opposite of slop" because it inverts two features:

  1. Excellent high quality content.
  2. A primary source, so low Baudrillard simulacrum level.

Slopness is of course mediated by both content quality and simulacrum level but people usually only focus on the first point when arguing about it.

I really like this word. I think it points at an important phenomenon that predates LLMs which we previously discussed in terms like "clickbait" and "yellow journalism" that weren't really fit for purpose. Slop captures the ubiquity of it, and the dehumanization implied by that ubiquity. It conveys an image of this mediocre vaguely disgusting mass produced product filling a feed trough with fat pigs lining up to consume it. It appropriately activates our disgust response and helps us orient to just how bad we've allowed things to become. But slop is undertheorized and frequently conflated with spam, cliche, shoddiness, and kitsch, which are all closely related but not quite the same thing.

What Is Slop?

Let's start with the intuition. What is slop?

Slop is written to pad the word count.
Slop is when you procrastinate on your college essay and crap something out the night it's due.
Slop is the logical conclusion of chasing the algorithm.
Slop is the distilled extruded essence of the Id.
Slop is when you have a formula and stick to it.
Slop is when you can guess the exact minute in a police procedural where they find the killer because it's the same in every episode.
Slop is when the k-complexity of the generator is low enough that you can infer its pattern.
Slop is eating lunchables every day at school until you puke.
Slop is when a measure ceases to be a good target.
Slop is the 12th sequel to a superhero movie.
Slop is generated from the authors prior without new thinking or evidence.
Slop is Gell-Mann amnesia.
Slop is in distribution.
Slop is when the authors purpose for writing is money.
Slop is a failure to say anything interesting.
Slop is what you find at the bottom of the incentive gradient.
Slop is a deeper simulacra level than it purports to be.
Slop is vibes.

Is Slop Spam?

Slop is distinguished from spam by the fact that slop is brought into being by demand rather than supply. Spam happens because 3rd world gray market actors want to sell penis pills and a tiny sliver of recipients buy them. Slop happens because 1st world Facebook boomers want to see emaciated poor people devoured by bugs. Slop is therefore intrinsically populist and emotional in a way that spam usually isn't. Spam usually appeals to greed or insecurity because it's a sales pitch, spam promises free and cheap and easy. It's discount pills and one weird tip and money waiting for you in a swiss bank account. Slop by contrast is usually the product. It's dogs dying of fake illnesses and tear jerking human interest news stories and outrageous political gaffes that never happened.

Probably the most illustrative pre-AI slop incident is Elsagate. If you're unfamiliar basically people started figuring out they could make money from the iPad babies on YouTube by chasing whatever the algorithm wanted them to make. This incentive gradient led them to increasingly depraved Freudian content featuring barely disguised sexual and violent themes. Parents then discovered this rabbit hole and concluded it must be the work of some kind of grooming ring or satanic pedophile cult. Having once been a child on the Internet myself I remember watching stuff like Happy Tree Friends and Madness Combat thinking it was hilarious, so I have no doubt that this content mostly just reflects the whims of children's Id unconstrained by the normal superego of network executives and parental supervision. It's uncomfortable for us to admit that children's revealed preferences lead to an uncanny Freudian nightmare world so we rationalize that part away and blame it on some confabulated Other. In any case if one looks up a screenshot of thumbnails from the relevant channels they're immediately recognizable as having a similar aesthetic to AI slop.

Is Slop Kitsch?

Kitsch by contrast is more like a subset of slop than a distinct category. Kitsch is comfortable, nostalgic, familiar, nonchallenging. Slop is however frequently if not usually bizarre, disgusting, surreal, fear inducing, sensationalist. Slop has a distinct negative tilt for the same reasons why cable TV news has a distinct negative tilt. Slop is also focused on the appearance of novelty while in fact being low entropy extrusions spun from the authors prior. Kitsch wants to seem familiar and unthreatening, wholesome, there's a reason why Thomas Kincaid paintings look like the abstract spirit of Christmas vomited onto the canvas. AI slop is a hallucinogenic fever dream because it's based on the janky world models of neural nets. The term 'slop' refers not just to a feed trough but also the idea that the creations are sloppy, they're full of AI jank that gestures in the direction of something but doesn't quite get the details right. A woman's water breaks during pregnancy and it spills out across her belly as thick white milk. A fat woman with a boulder falls over on a glass bridge and her butt breaks the glass, the rest of the bridge just falls apart for some reason. AI models make mistakes that no human artist ever would, rendered in exquisite detail that would be way too expensive to do with a pen or brush.

At the same time the definition of kitsch given by Hermann Broch, "mimics its immediate predecessor with no regard to ethics", is a clear precursor to the concept of slop. Perhaps the most obvious example of this is the way that AI models make it easy to mimic the style of an artist like Hayao Miyazaki, which is associated with Miyazaki's worldview, and apply it in ways contrary to that worldview. Consider the time that Trump's staff shared a Studio Ghibli meme of ICE arresting a crying woman. Miyazaki's work usually evokes a kind of whimsical misanthropic naturalism that would avoid both this subject matter and portrayal of this subject matter in this way. If I give the model a picture of ICE arresting a crying woman and tell it to render the subject "in the style of Studio Ghibli", what I get back is not how Miyazaki might choose to render this subject if he felt he had to for some reason. It's not for example a metaphor for this kind of arrest which would be more family friendly. It is a literal translation of the concept with a thin veneer of the Studio Ghibli animation style applied over it, creating a bizarre simulacrum that has never existed before. Suddenly art styles and art materials like claymation or embroidery are mere aesthetics that can be swapped around onto any subject or combined in photorealistic ways that would be absurdly difficult if not impossible to create in real life.

The Ghibli ICE incident is emblematic of a core feature of AI slop: Poor taste or boring use of mediums and styles that were previously gatekept by skill barriers. Slop is Hollywood level visuals paired with sub-amateur storyboarding and writing. Style and medium are increasingly no longer a costly signal of effort. It used to be that if you wanted a plastic figurine of something your best bet was to 3D print and paint it, which was already a huge upgrade over the previous process of injection molding or similar. If you wanted a carving, plushie, etching, wax figurine, crochet or chalk drawing of something you had to actually make it and take a photo. Different art mediums require different levels of effort and that acted as a useful way to show someone put thought and care into sending a particular message. That going away creates a much flatter media environment where it's harder to tell who is putting deep effort into what they're doing. On the other hand it's now possible to create beautiful things without them costing a fortune.

Do AI Outputs Have To Be Slop?

To me the most interesting question is to what extent slop is a necessary feature of AI output. We've seen AI image and video generators get massively better over the last few years, but they still have the slop nature most of the time. This raises the important question of what the slop nature is and what it would take to transcend it. I think "slop nature" is created by a few core features. One of these is jank, like the classic problem where image generators couldn't draw good hands for a while. Jank is the kind of thing you can fix by just making the models better. But a subtler problem is outputs feeling too regular or low entropy. The various ChatGPT styles are nauseating to read after you've waded through hundreds of pages of output in those styles, especially when they're associated with web pages or social media posts that waste your time. I feel like answers before every chat service incorporated search were slop-ier because they were written in an authoritative tone from the models prior without reference to any external sources.

Ultimately I think what gives the impression of slop is when you can fully infer the generator. When you type in "fat cat rolling around on carpet 4k digital" the video you get back from that is entirely defined by the prompt. It doesn't matter how good the model is, if I type "fantasy movie 4k young boy takes up the sword and defeats evil" and the model dutifully creates me an entire animated film with that premise there's still a sense in which the movie's fundamental information content is summed up by "fantasy movie 4k young boy takes up the sword and defeats evil". Or at the very least models right now interpret such prompts in very literal and unimaginative ways. Certainly unless you have a thing that goes out of its way to be original I have to imagine that what you got out of that prompt would have the slop nature even if it was well made. In the limit then overcoming slop is about active learning and novelty search.

Because these AI models are trained on such vast corpora outputs which are novel to them should mostly be actually novel, the question is how to get those outputs. My theory is that models will perform a kind of iterated tuning to extend their training distribution in a particular direction. Similar to how new music genres are made when a band cultivates excellence at the edge of the distribution of existing music, I think models will perform novelty search by going out to the edge of what is known and applying relatively objective standards to what they find there. People aren't really doing this right now because the best AI models are proprietary and expensive to tune. We don't really have online learning as a service yet, though Thinking Machines is about to change that. I think AI models will probably have the slop nature until online learning and agency get deep enough to let them make high effort artifacts outside the original training distribution, at which point they'll become the core drivers of media and culture.