A Brief History of Slop: Why Flawless Text Triggers Both Delight and Disgust

I. The Call from the Publisher

A few months ago, a friend of mine fell hard for AI — the way you fall for something new and shiny — and promptly generated a book about it. That’s generated, not wrote, and he was honest with himself about the distinction. He used the most capable, most expensive flagship models, knowing the cheaper alternatives would have produced something worse. He watched carefully to make sure they didn't pad the text, filtered out the obvious nonsense, and ran everything through additional automated tools meant to improve it. Eventually he assembled the whole thing, showed it around, bragged about it a little, and only then spent a long, painful stretch working through the text by hand: editing, filling gaps, cleaning house, arguing with the machine about what it had meant to say. At some point, the obvious AI writing in his own book made him sick.

He could forgive the em dashes. What got to him was the sheer density of idiotic metaphors, like AI as a bridge or a new chapter, and the machine-made triads scattered through the prose like confetti. Anaphoras also dropped in for no reason: three sentences in a row starting the same way, none of them earning it. Not A but B constructions appeared in every paragraph like clockwork. My friend looked at the generated text and saw that it was repulsive.

Last week, he got a call from a small publishing house where the AI trend had finally trickled down. Someone had passed them the draft his machine had produced. They'd read it, liked it, and wanted to put out a print edition. The text is very clean, the editor told him; no heavy editing needed, just light proofreading later.

This editor is good at his job. He’s no less competent than my friend, just at the point where my friend was a few months ago. He didn't recognize the slop for one very simple reason: his editorial detector is calibrated for human errors, and the text had none. He just hasn't gotten up close with machine errors yet, so he hasn't learned to spot them.

The text hasn't changed. The reader has.

II. Word of the Year

"Watching in real time as 'slop' becomes a term of art. the way that 'spam' became the term for unwanted emails, 'slop' is going in the dictionary as the term for unwanted AI generated content."

— @deepfates, quoted in: Simon Willison, "Slop is the new name for unwanted AI-generated content," http://simonwillison.net , May 8, 2024

Two days after this tweet appeared, programmer Simon Willison published a blog post. Not all AI-generated text is slop, he clarified, but if it's been generated thoughtlessly and pushed on someone who didn't ask for it, then “slop” is exactly the right word. Willison later admitted he hadn't coined the term himself; Deepfates, for his part, insisted the credit was mostly Willison's.

The @deepfates account has since been deleted. The tweet that started it all now returns a 404. But the prediction came true, and quickly: less than two years passed between that tweet and the word's entry into Merriam-Webster. “Spam” took fifteen.

"Slop: digital content of low quality that is produced usually in quantity by means of artificial intelligence."

— Merriam-Webster, "2025 Word of the Year: Slop," December 2025

III. The Hit Parade

Over three years, people with appropriately calibrated vision have assembled a ranking of the tells that keep poking through the surface of AI-generated prose.

One important caveat: none of these patterns is an error. Every one of them appears in good human writing. The problem is frequency and the absence of intent. Neural networks gravitate toward these constructions simply because they are statistically safe.

1. Not A but B

The model frames every argument as an unexpected reversal, manufacturing a false sense of depth. Use this device once and purposefully, and it works. Use it ten times and you get structural noise. "This isn't a tool — it's a partner."

2. The Triad

When a number of distinct elements appear, it’s always three: never two, never four. (The Lord, in Monty Python and the Holy Grail, was quite specific on this point: “Four shalt thou not count, neither count thou two.”) The machine applies the rule of three mechanically, regardless of whether three elements are actually necessary. "Write clearly, persuasively, and to the point."

3. Wordiness

AI text runs longer than human text at the same informational payload, producing nominalizations and polysyllabic constructions where simpler language would do. "The implementation of this approach involves..." instead of "Here's how it works:"

4. The Em Dash

This topped the list for a while, and LLMs are gradually filtering it out, but their habit of dropping an em dash into the middle of a sentence with no rhythmic justification hasn't gone anywhere yet. "This matters — and here's why — because..."

5. The Hedging Opener

The sentence leads with a qualification instead of a claim, creating the illusion of balance where none exists. "It's worth noting that..." / "It's important to understand that..."

6. Lexical Markers

The machine favors some words out of all proportion. Each generation of models brings a fresh set as they become familiar with and correct for the older ones. "Tapestry," "landscape," "certainly."

7. Thin Punctuation

Fewer commas, parentheses, and semicolons appear. “And” chains long sentences together. "This is a complex topic and it demands attention and understanding and patience."

8. Metaphor Overload

The machine reaches for a metaphor where a fact would suffice, drawing generic comparisons to things like mirrors, bridges, and journeys. "AI is a mirror of our society."

9. Anaphora Out of Place

Rhetorical repetition produces rhythm — but without design. It’s f*"We live in an age of change. We stand at a threshold. We choose our path."*

10. The Pseudo-Profound Closing

The final paragraph sounds weighty but says nothing. It could sit beneath any piece of writing on any subject. "Ultimately, it all depends on us. The future is in our hands."

As we noted in the fourth item, em dashes are no longer the giveaway they may once have been. In July 2026, The Economist analyzed 55,940 sentences and 1.2 million words, comparing its own articles against versions written by four AI models given the same prompts. The em dash turned out to be an unreliable marker: ChatGPT now uses it less often than living people do.

The journalists also asked an older version of the model whether AI overuses the em dash. "Ha — great question!" it replied — with an em dash. They asked a newer version the same question:

"[It] soberly says 'they're best used sparingly.' Only a ghost could shapeshift so quickly."

— "How to spot AI writing", The Economist, July 30, 2026

IV. Delight, Fury, Acceptance

An experienced editor opens his first AI-generated text and doesn't find what he usually does. There aren’t any grammatical errors, so the entire first tier of his work has simply vanished. There’s no structural collapse where the third chapter rehashes the first and the conclusion contradicts the introduction. The density is even from start to finish, with no sagging pace. Above all, the editor doesn’t see a single familiar marker of bad writing, the kind his reflexes have been honed to catch over twenty years.

He’s delighted. His detector is silent, and the brain reads silence as quality. It needs time to stop rejoicing in the silence and start hearing the false notes.

After that, fury builds in waves. First comes the recalibration: once you've read enough machine-generated text, the new patterns finally become visible. When a twenty-year veteran realizes he misses something obvious, his professional self-worth takes a hit. Then he can’t unsee the patterns, since they appear everywhere, including in the writing of living people and in his own drafts from three years ago. The LLM picks up his favorite tools and runs them on idle: the anaphora lands not where the text needs emphasis, but wherever the algorithm got bored without structure.

The most painful part is that he can’t explain his anger to an outsider. The editor sees techniques deployed without intent, but the person he's talking to sees only literate, logical, clean prose. They conclude that the editor is nitpicking.

Someone without editorial calibration, however, has their own version of the same irritation.

They don't trip over the triads and they don't know the word "anaphora," but they do know that customer support sends them three paragraphs of impeccably polite text that doesn't mention their question once. They read a letter with their name in the first line — and understand it wasn't written for them. At least the professional knows why he’s annoyed. Outsiders only know that something feels false and they can’t explain exactly what’s wrong.

If you're lucky, a third stage follows. It’s not reconciliation, but instrumentalization.

An LLM simply doesn't notice when it loses the thread of meaning, it won't flag an imprecise prompt, and it will stay silent if the writer hasn't yet decided what they want to say. It certainly won't catch the moment when the task has shifted mid-journey while the brief stayed the same. A human editor stops at exactly that point and asks.

A division of labor emerges. The model gets the draft, the structure, the variant hunting, and the repetition catching. You keep the idea, the voice, the decisions, and the responsibility for what comes out. The order matters: your thoughts first, then the machine’s analysis, then your judgment again. It doesn't work in reverse — machine-made form sets quickly, and breaking it apart is harder than starting from scratch.

My friend, incidentally, eventually set the generated book aside and went off to write his own.

V. Who Invented This

Nobody ever asks where LLMs’ stress on form over function came from in the first place.

It was the same place as press releases, meeting minutes, and diplomatic notes. Rote phrasing originated in small talk, where every remark is appropriate and nothing gets said. Humans invented this genre for themselves — and took pride in it. It required skill. People spent years learning it. We built careers on the ability to say nothing beautifully.

Now AI has mastered this entire elaborate science in a single prompt — and practices it better and cheaper than humans ever did.

Form without substance is not, in itself, a vice. A polite smile or a standard “congratulations!” does its job. A condolence assembled from ready-made formulas works better than an honest attempt to say something personal at a moment when there is nothing personal to say. Platitudes are the grease without which the social machinery creaks. The problem is only that we stopped distinguishing them from the real thing.

The machine has removed the last remaining difference — human effort. Before, a living person stood behind the empty shell. Their fatigue, their tone, their typo all betrayed a presence. There was no substance, but someone had spent fifteen minutes of their life reproducing the form. Now a model can create the identical structure in no time. Suddenly the emptiness became visible.

Science attempted to describe this phenomenon in formal terms. A group of researchers led by Cody Kommers of the Turing Institute proposed three identifying properties of slop: surface competence, asymmetry of effort, and mass producibility. The first is the appearance of quality with no real meaning behind it. The second is that content costs the producer almost nothing while it costs the reader dearly. The third is scale: the economics of such content rest entirely on mass production of text.

The authors themselves honestly admit that slop continues to resist formal definition and remains at the stage of "you know it when you see it." After three years of discussion, the best science has managed is the same as editorial instinct.

"AI slop is cheap to generate and expensive to review, and the review layer is already thin."

— Sebastian Baltes, Marc Cheong, Christoph Treude, "AI Slop and the Software Commons," arXiv:2604.16754, April 2026

VI. Adaptation

A trained eye extends beyond text.

Once you learn to recognize form without substance in a paragraph, you start seeing it everywhere. It appears in the meeting where everyone weighed in and nothing got decided, the statement where the structure holds but the meaning doesn't, and birthday-party conversation where every remark lands perfectly and nobody has said anything at all. The situations haven't changed — the person looking at them has.

Slop gave a name to something that always existed.

The new lens comes at a cost. The editor, now spooked, starts running his old drafts through the same checklist. He hits a triad and crosses out the third item, even when it belongs there. He stumbles on a "not A but B" construction and rewrites a sentence that was perfectly fine. The detector, calibrated on other people's generated text, doesn't make exceptions for his own.

Worse, detectors are always running late. A pattern disappears the moment people start noticing it: the em dash has vanished from ChatGPT and delve has faded out. The next generation will bring something new, starting the whole phenomenon over. The detector is chasing a target that keeps moving precisely because it's being chased.

"The detector punishes the human for writing like a machine and rewards the machine for writing like a human."

— Joseph Smith, "Are AI Detectors Accurate? How They Get It Wrong and Who Pays", SilentRoom Journal, June 2026

In the 18th century, philologists declared they an error — a plural pronoun can't stand for a single person. People kept talking the way they always had. In 2019, Merriam-Webster conceded and recognized the usage as standard.

The norm followed the statistics, the way it always does, but this particular they had a reason behind it. People needed a way to refer to someone without specifying their gender. The pattern was doing real work.

People adapt to a great deal. Olives are bitter the first time, unpleasant and baffling. Then you get used to them. Then you start telling the varieties apart. The olives haven't changed — your relationship with them has.

If humanity managed to make its peace with taxes, vertical video, and the phrase "fast-paced, dynamic company," the triad shouldn't pose much of a problem over the long run. Even if it never gets scrubbed from every last piece of text, at some point it will stop grating and recede into background noise. The text won't change. The reader will.

The em dash has already been through this. It dropped out of machine-generated text and out of human text too. The bots stopped using it because it annoyed people, and people stopped because they were being mistaken for bots.

Norms always follow statistics.

Nobody specified whose statistics those were.