The AI Slop Field Guide

More than half of what you read online now started as language model output, affectionately dubbed "SLOP". That's not automatically a problem, as AI-assisted writing helps non-native speakers get ideas across clearly, and it gives people without an editor a way to sound coherent under deadline. However, when generated text gets passed off as original thought, which let's be honest is 80% of the time, by people charging for said "original thought," you should be able to tell whether the thinking is theirs. If someone is asking for your attention or your money on the strength of their thinking, you are responsible for your own discernment. This guide teaches you to recognize AI writing by its four core mechanics. It is vital to the effort of reducing SLOP to understand how this SLOP gets made.

Why LLMs Write the Way They Do

When you write a sentence, you start with something you're trying to say and search for the words that get closest to it. A language model does the opposite, as it has no concept or idea to protect, only a running calculation of which word is statistically most likely to follow the words already on the page. If given "In the morning, the rooster ________," it lands on "crows" because that's the highest-probability continuation. Two side effects of that process explain almost every tell in this guide. First, models carry a repetition penalty, so across a long piece they keep swapping in synonyms even when the original word was the correct one, which produces the over-varied, thesaurus-happy word choice seen in long form SLOP. Second, because there's no underlying concept driving the piece or a relationship to words and the real world, models fall back on the most statistically average version of whatever structure they're producing, which is why AI paragraphs, sentences, and arguments all tend to converge on the same shapes and structures.

Word choice, sentence structure, paragraph structure, and argumentation all reveal this core weakness in LLM generated copy.

Word Choice Tells

Four categories of word choice show up disproportionately in generated text. None of them is disqualifying on its own, but concentration of the words is a likely symptom of LLM generated text.

Abstract, non-visual nouns

Words like infrastructure, realm, ecosystem, and landscape describe a combination of things without naming anything material or visual. "Ecosystem" in AI writing rarely means an environmental system; it means "all the stuff related to this thing," which is a concept a model reaches for because it's vague enough to fit almost any context. INFRASTRUCTURE is the worst and most common example of this seen in SLOP. Which is sad, cause that word does work well for a lot of stuff.

AI-favored adverbs and adjectives

Seamlessly, crucially, vibrant, vital, pivotal, and the intensifiers genuinely, truly, fundamentally, and actually. These words import a sense of stakes or importance instead of doing the harder work of finding a more precise noun or a sharper example to justify its importance.

Academic verbs from AI training data

Words like delve, harness, and leverage show up constantly in AI output because models are trained on a large share of academic and formal public writing. Almost nobody says "let's leverage this" out loud, but it's everywhere in generated text because it's common in the corpus and vernacular of SEO optimized and marketing copy that is a large portion of AI training data.

Metaphor-as-verbs

Phrases like earns its place, anchors the idea, illuminates, surfaces, and wrestles with dress up a literal claim as an action, and SLOP is littered with them. "This anchors the idea" usually just means "this is the most important part of the idea," but said in a more dramatic way than the sentence needs.

Sentence Structure Tells

AI sentences follow a tighter set of rules than most human writers do, because AI has no thinking process regardless of the "thinking" it displays before each response. Humans write messy and varied; models write consistently with 6 distinct patterns. A core idea in understanding how AI writes is that its baseline assumption is that the reader is essentially stupid. Consequently, nearly 80% of its output isn't conveying information, but stating it, and then immediately explaining what it just stated, as if you missed it.

Negative parallelism

"It's not X, it's Y." This construction exists in human writing, but AI uses it to manufacture clarity around something the reader was never confused about in the first place. When you see it clarifying a distinction nobody needed, that's the result of an LLM's lack of any understanding.

Dramatic countdowns

"Not X, not Y, but Z." A three-step version of the same move: dismissing wrong conclusions the reader didn't actually have, for dramatic effect.

False-range phrases

"From infrastructure to culture," "From startups to Fortune 500 companies." This sounds like it's establishing scope, but the range usually isn't a real constraint on anything, it's just a way of saying "this is broadly true" with extra flourish.

"The results? Consequential!"

Short questions with dramatic one word answers that don't actually contribute to anything other than transitioning conversation show highly in LLM based writing. This high-drama construction is uncommon in native human writing because it reads as overly staged. It shows up constantly in AI output, and especially in Claude's writing. AI is trained on marketing copy and SEO-optimized content that relies on hyperbolic, melodramatic framing to manufacture tension and faux-stakes where none actually exist.

Three-part lists

These aren't inherently a tell. Good writers use them because the rhythm works, and most of us were taught rule-of-three structure early. The tell is when every list in a piece follows identical parallel structure with the same rhythm, over and over, suggesting a pattern being applied rather than human intent.

Participle tails, god help us on this one

A sentence ending in a present-participle phrase that just restates the point, such as "contributing to a broader shift," "highlighting the underlying tension," "underscoring the change." This is the LLM not trusting the reader to get the point without a label attached to it.

Paragraph Structure Tells

Models are trained on a lot of five-paragraph-essay structure and rarely stretch past three to five sentences per paragraph before wrapping it up.

Uniform paragraph shape

Every paragraph in the piece is the same length, follows the same topic-example-close arc, and none of them breathe differently from the others. Human writing varies paragraph length based on how much the point needs, while AI writing defaults to the statistically average shape every time.

One or two sentence paragraphs for drama

A standalone sentence, set off with a line break, doing the emotional work an argument should be doing. Human writers reserve this for the rare sentence that's earned it.

Announcement openers

"Here's the thing" or "consider this" at the start of a paragraph, telling you what the paragraph is about to do instead of just doing it. A good editor strikes these because they add words without advancing the idea.

Wrap-up closers

"Ultimately," "In totality," "when we zoom out," or "this is part of a broader shift" at the end of a paragraph. These sentences aren't adding a new idea, they're performing that the paragraph is finished and moving the reader along before one can think about the lack of anything substantial being said.

Triple restatement

A bolded label followed by a colon, then a sentence restating the label in full, then bullets that restate it a third time. Strip the formatting and lay the three versions side by side: it's the same sentence three times. In good writing, every sentence should be a varying building block of the claim or the thesis, while much of LLM writing consists of restated theories.

When arguing or writing in general, we deepen our argument the further we go into the paragraph, whereas LLMs will just expand the same thing over and over again.

Argumentation Tells

Models don't hold beliefs or knowledge in the way a person does, they're just producing the statistically likely continuation of an argument shape. They don't arrive at the position you are at. That produces six recognizable failure patterns.

Hedge-everything balance

"This is true, but it also might not be. There are arguments on both sides." Real nuance narrows a claim, specifying exactly when it holds and when it doesn't. AI hedging just cancels the claim out entirely, because it's an absence of a position.

Restating one point as several

The same underlying claim, rephrased and presented as if it's a second, independent argument. "Media literacy matters" followed by "in an era of declining media literacy, understanding online rhetoric is more important than ever" is one idea acting like two.

Disproportionate stakes

Calling a software update a paradigm shift, calling a simple idea groundbreaking. AI is trained heavily on marketing and SEO copy, which is optimized for drama regardless of what's actually being described. The strongest arguments are usually the most modest in what they claim and heavy in nuance.

Evidence cited only at the label and reference level

"Experts believe" or "research suggests" with no specific study, name, or number attached, or if there is a citation, there is a lack of depth in the citation itself. Real evidence gets particular: whose research, what year, what number. Vague sourcing is a strong signal the writer has no actual citation in mind, aka SLOP.

Institutional, uncommitted voice

No part of the argument carries more emotional weight than any other part. Individuals write with uneven charge, caring more about some parts of their argument than others, having something like a grudge or a blind spot. A model has no such asymmetry, so everything reads at the same temperature.

Self-defeating sub-arguments

A business strategy piece presents "four pillars of sustainable growth": acquisition, retention, monetization, and innovation. In the retention section, it argues that acquisition is actually more cost-effective in the short term, so companies should focus their budget there instead. This is a defensible point on its own. But it undermines the very premise that all four pillars are equally essential to the framework. The paragraph is locally coherent; the broader thesis is dismantled by its own sub-argument.

The Sixty-Second Scan

For a fast read on any piece of writing, whether you're vetting it or checking your own AI-assisted draft:

1. Read the last sentence of every paragraph. If most of them are wrap-up sentences that add nothing when removed, that's a strong signal to others you post SLOP.
2. Scan for the four word-choice categories. A handful is normal. A cluster in every paragraph is not, and is SLOP.
3. Check whether every paragraph is roughly the same length and follows the same internal shape. IT'S EVEN WORSE IF THERE'S TWO SHORT SENTENCES AT THE START as a hook.
4. Find the strongest claim in the piece and ask whether it's backed by a specific source or number, or just "experts" and "research."
5. Ask whether the piece takes an actual position, or resolves everything into "both sides have a point."