AI Slop: The 19 Writing Patterns That Scream ChatGPT (and How to Fix Them)
You can usually tell within two sentences when something was written by AI and published without a real edit. You might not be able to name what gives it away, but you feel it — the piece is grammatically perfect, weirdly confident, and somehow says nothing.
That feeling has a name now: AI slop. And the good news is that slop isn’t a vibe, it’s a checklist. The tells are specific, repeatable patterns you can learn to spot in your own drafts and cut in a single editing pass.
This guide walks through 19 of them with before/after examples. The pattern list is adapted from no-ai-slop, an open-source editing skill by Peter Yang (MIT licensed) that I liked enough to build into this site. If you’d rather have AI do the editing pass for you, I turned it into two free prompts: the AI Slop Remover edits your draft while keeping your voice, and the AI Slop Detector flags every pattern with the quoted line so you can fix things yourself.
One note before we start: using AI to draft is fine. I do it constantly. The problem is publishing the first draft, because every model has default habits — and once you know them, you see them everywhere.
Why AI writing sounds like AI
Language models are trained to produce text that sounds authoritative and polished on average. That average is the problem. Punchy contrast structures, dramatic reveals, and importance-signaling phrases show up constantly in the training data’s most engagement-optimized writing, so models reach for them by default — whether or not your content earns them.
Your job as an editor is to notice where the structure is doing fake work: making a plain fact sound like a revelation, or an unsupported claim sound like consensus. Cut the fake work and what’s left is your actual point, stated plainly. That’s the whole method.
The word-level tells
1. The banned words
Some words are so overrepresented in AI output that they function as watermarks. The famous one is delve, but the full list is longer: foster, leverage, utilize, facilitate, empower, streamline, robust, cutting-edge, game changer, tapestry, realm, beacon, multifaceted, meticulous, paramount, transformative, elevate, embark, harness, ever-evolving.
None of these words are wrong. They’re just markers, because real people rarely say them.
Before: “This robust framework empowers teams to streamline their workflow.”
After: “This framework cuts the steps in your workflow from nine to four.”
2. Empty filler phrases
“It’s worth noting,” “at the end of the day,” “in today’s world,” “the reality is,” “let’s dive in.” Each of these delays the point by a clause. Delete the phrase and the sentence almost always gets stronger.
3. Empty adverbs
Just, literally, honestly, simply, actually, truly, fundamentally. Sometimes these carry real emphasis or spoken rhythm — keep those. But AI sprinkles them as texture, and texture is exactly what you should cut.
The structural tells
These are harder to spot than banned words because they hide in the shape of the sentence, not the vocabulary.
4. Binary contrasts
The single most common AI structure: “It’s not X. It’s Y.” Also appears as “The question isn’t X, it’s Y” and “It’s not just X, but Y.”
Before: “This isn’t about writing faster. It’s about writing better.”
After: “The goal is better writing, and faster is a side effect.”
5. Negative listing
Binary contrast’s louder cousin: “Not a tool. Not a platform. A movement.” Just say what the thing is.
6. Throat-clearing openers
“Here’s the thing.” “Let me be clear.” “I’ll be honest.” These promise a point instead of making one. Cut the opener, keep the point.
7. Faux-insight setups
“What nobody tells you…” “The part everyone misses…” “Here’s what most people get wrong…” These frame the writer as the lone truth-teller before saying something everyone actually knows. If the claim is genuinely surprising, it doesn’t need the drumroll.
Before: “Here’s what nobody tells you: consistency matters more than talent.”
After: “Consistency beats talent, and it isn’t close.”
8. Dramatic colon reveals
A noun phrase, a colon, then a punchy reveal: “The best part: it learns.” “The catch: nobody reads it.” Colons are for lists, labels, and quotes. When every third paragraph ends in a colon reveal, the rhythm becomes the tell.
9. Trailing -ing analysis
The clause that pretends to explain significance: “…highlighting the team’s commitment to quality,” “…underscoring the importance of preparation.” These trailing participles never contain information. Replace them with the actual consequence or delete them.
Before: “The update adds offline mode, showcasing the company’s user-first philosophy.”
After: “The update adds offline mode, so the app works on the subway.”
10. Importance puffery
“Marks a pivotal moment.” “Stands as a testament.” “Plays a vital role.” “Solidifies its position.” These phrases announce that something matters instead of showing why. State the fact and let the reader decide.
Before: “The launch marks a pivotal moment for the company.”
After: “The launch is the company’s first paid product.”
11. Weasel attribution
“Experts agree.” “Studies show.” “Widely regarded as.” Attribution without a source is decoration. Name the study, name the expert, or cut the claim. If you’re using AI to draft, be extra careful here — models will happily generate the shape of a citation with nothing behind it.
12. Fake-strong verbs
“Serves as a centralized hub.” “Acts as a catalyst.” “Functions as a bridge.” These verb phrases sound muscular and mean “is.” Plain is and has are usually clearer, and clearer wins.
13. Synonym cycling
AI rotates synonyms to avoid repeating a word: the agent, then the assistant, then the tool, then the system — all the same thing. Human writers repeat the clear word. If it’s an agent, call it an agent every time.
The rhythm tells
14. Dramatic fragmentation
“That’s it. That’s the whole thing.” “One tool. Zero setup. Infinite possibilities.” Stacked fragments read as profound for exactly one sentence, and AI writes them by the paragraph.
15. Robotic rhythm
Every sentence the same length. Every paragraph the same three-beat structure. Every section opening the same way. Real writing speeds up and slows down; when the shape of paragraph four predicts the shape of paragraph five, readers feel the template even if they can’t name it.
16. Rhetorical setups
“What if I told you…?” “Plot twist:” “Think about it:” And the self-answered question: “Does this work? Absolutely.” All of these simulate a conversation the reader never agreed to have.
The ending tells
Endings are where AI drafts fall apart most reliably, and most people stop editing before they get there.
17. The fake-profound kicker
The final line that turns your point into a cute metaphor: “In the end, the best prompt is the one you never have to write.” It sounds deep and commits to nothing. Delete it and end on your last concrete point — a real takeaway or a next action beats a mic drop.
18. The summary-recap ending
“In conclusion…” “Ultimately…” or a full paragraph restating what the reader just read. The reader was there. End on the last thing that’s actually new.
19. Formatting slop
Emoji in headings, bold sprinkled mid-sentence for fake emphasis, bullet lists where two sentences of prose would flow better, and a header for every two-sentence section. Formatting should follow the content. When the formatting is doing the emphasis, the words aren’t.
Em dashes belong in this bucket too. One or two in a long piece is fine when they beat commas. AI uses them as a default rhythm crutch — if your draft has five in a paragraph, that’s the model talking, not you.
How to run the edit without losing your voice
Knowing the patterns is half the job. The other half is not overcorrecting, because a draft scrubbed of everything distinctive is its own kind of slop — technically clean and completely anonymous.
The approach that works, borrowed straight from Peter Yang’s skill, is the minimum effective edit:
- Read the full draft first. Note your actual point and the three or four things that sound like you — a blunt opinion, a joke, a digression, a spoken rhythm. Those survive the edit untouched.
- Cut the word-level tells. Banned words, filler phrases, empty adverbs. This pass is mechanical and takes five minutes.
- Fix the structures. Binary contrasts, colon reveals, puffery, weasel claims. For each one, ask what the sentence is actually saying and say that instead.
- Rewrite the ending. Delete the kicker or recap, end on your last concrete point.
- Read it out loud. Anything you wouldn’t say to a sharp colleague, change. Anything that sounds like you, keep — even if it’s imperfect. Especially if it’s imperfect.
If you want this whole process done in one shot, paste your draft into the AI Slop Remover prompt. It applies these exact rules, makes the smallest edit that fixes the problem, and reports what it changed so you can veto anything that went too far. For a read-only audit — say, checking a freelancer’s submission or a teammate’s draft — the AI Slop Detector prompt names each pattern with the quoted line and leaves the text alone.
For the writing-quality fundamentals underneath all of this — active voice, concrete detail, front-loading the point — my Prompt Engineering 101 guide covers how to get better first drafts out of the model in the first place, which means less slop to remove later.
Credit where due: the pattern taxonomy in this guide comes from Peter Yang’s no-ai-slop project, released under the MIT license. The examples and commentary here are my own. If you use Claude Code or another agent harness, his repo installs directly as a skill and is worth having.