In 2025 Merriam-Webster added a new definition for the word "slop." They're not talking about food anymore. They're talking about the specific brand of hollow, over-polished writing that language models produce by default. AI slop. It's already everywhere.
A 2024 study by Kobach and colleagues found that over 13 percent of biomedical abstracts were AI-assisted, and in some fields the number hit 40 percent. That was two years ago. I would bet my right arm we're well past that now.
Think about the word "underscore." Before large language models, it showed up in scientific writing once in a while, usually when someone wanted to sound a bit clever. By 2024 it appeared in one out of every five publications on PubMed. That's a 5x increase in a single year. Scientists didn't suddenly fall in love with the word. They just all got the same writing assistant.
I write about writing for a living, so I've become the friend everyone sends drafts to. Half of what I read now sounds like the same voice. If I finish a paragraph and can't tell you a single opinion the author actually held, I stop reading. That's slop.
Here's how to spot it in ten seconds, and if you're using AI to help with your own writing (I do), how to stop your work from ending up in that pile. The most useful fix is at the end. Do not skim past it.
The words that give it away
You already know about "delve." It became a meme in 2023, and newer models mostly avoid it now. If your only detector is "look for delve," you'll miss most of what's out there.
The current giveaways in 2026, in rough order of how often they show up.
- Underscore. As in "this study underscores the importance of X." Unless you are literally writing a scientific paper, don't use it.
- Tapestry. Almost always as "the rich tapestry of X." Nothing is a rich tapestry.
- Landscape. Used abstractly, as in "the landscape of digital marketing." Not a landscape. Say "world" or delete the phrase.
- Testament. "Stands as a testament to X." A perfect AI phrase. Kill it.
- Realm. Everything doesn't have to be a realm.
- Foster, showcase, harness, unveil. As verbs, on abstract nouns. "Foster growth", "showcase potential", "harness the power of", "unveil the secrets of". All AI defaults.
- Pivotal, vibrant, meticulous, intricate, groundbreaking, substantial. Fine words. AI uses them as adjectival wallpaper.
None of these words are wrong on their own. The problem is the *frequency* and the *context*. A human writer might use "tapestry" once in a novel. An AI uses it in a LinkedIn post about supply chain management.
The structure that gives it away
Words are the surface. The deeper tell is how the sentences fit together.
Watch for the "not just X, it's Y" construction. Language models are addicted to it.
- "This isn't just motivational. It's behavioral science."
- "Your mindset isn't just personal, it's cultural."
- "This isn't simply a product. It's a paradigm shift."
Every one of those sounds profound and communicates almost nothing. When you see the pattern, you're reading a machine.
Watch for triplets. Rule of three. Beginning, middle, end. AI is obsessed with them. "Our thoughts influence our emotions. Emotions drive behavior. Behavior creates results." Three beats, packaged like insight. Real writing varies the count. Sometimes two, sometimes five, sometimes one.
Watch for the bullet template. A short intro, three or four bullets each starting with a bold phrase and a colon, and a closing sentence that restates the bullets. That's the shape of 90 percent of the AI blog posts on LinkedIn.
Watch for em-dashes. Not hyphens. Actual em-dashes, the long ones. Ask yourself right now if you know how to type one on your keyboard. I didn't. Most people don't. Most keyboards don't even have the key. Language models auto-insert them because they were trained on published prose. If a paragraph has three em-dashes, that paragraph came from a machine.
Watch for words wrapped in quotes for no reason. "AI is 'transforming' work." The quotes signal the writer's discomfort with the word. A human either commits to a word or picks a different one.
The vibe that gives it away
Even if the words and structure are clean, AI writing has a smell.
An Oxford study called it sophistry. Writing that sounds well-reasoned and confident, structured like an argument, but doesn't actually say anything you could agree or disagree with.
You test for it this way. Pick a paragraph. Ask what specific claim it made. If you can't answer, it was sophistry.
The other content tell is confidence. Real writers hedge. They say "I think", "maybe", "from what I can tell", "I could be wrong here." Language models are post-trained to sound authoritative, so they commit hard to bland positions. A confident, unopinionated paragraph is almost always a machine.
Real writers also use judgment. Better. Worse. The best. The worst. Language models avoid these because they're tuned to be safe. Take positions. It's how people talk.
The fix (this is the part everyone skips)
If you're using AI to help with your writing, the trap is asking it to write for you. That's the "write me a post about the responsible use of AI" path, and that path always ends at "transforming the landscape of productivity" style slop.
The workflow that actually works is the opposite. You write the messy first draft yourself. Voice-to-text if it helps. Get the raw thing out. Then use AI for two things only.
First, use AI to prune. Ask it to compress your draft to the essential ideas without rewriting sentences. The prompt is roughly "keep my voice, cut what's redundant, don't rephrase." If it starts rephrasing, stop it and try again.
Second, use AI to critique. Ask "what's the weakest argument here?" or "which section reads generic?" These prompts are ten times more useful than "make it sound professional." The professional-sounding one is the one that gets slopped.
Read the whole thing out loud when you're done. This is the single highest-value two minutes of the process. If a paragraph doesn't sound like *you* saying it, you let the AI do too much. Put the mess back in.
We keep a longer version of all this in the Mindwand style guide, with the full ban-word list and the exact prune prompts we use. Feel free to grab it, or steal from Oxford's paper on epistemic hedging if you want the academic version. Or don't. Twenty other publications are covering this now, though most of them are, ironically, written by AI.
The honest part
I write a lot of drafts. I use Claude. If someone reads this piece and thinks "did an AI help him prune this?" the answer is yes. If they think "did an AI write it?" then I failed.
That's the standard now. Not "don't use AI." Not "use AI for everything." Use it as an editor. Ship what actually sounds like you.
If you got here and you still can't tell whether a piece is AI or not, you're not going to be able to tell. Trust that. Whatever you write from here on will be better if you can hear yourself in it.
We publish new posts on Mindwand most weeks. The income-story pieces are the ones where we apply this discipline most aggressively, so if you want a sample, start there. Feedback welcome. Roasting also welcome.
I hope some of it lands.