Open LinkedIn. Go on, I’ll wait.
Scroll past five posts and count how many open with one punchy line and then, after a blank space, “Here’s what I learned.” All five? Thought so. Five different people in five different industries, one voice. The voice belongs to the model. I mean, you can hear it by now, can’t you? The little rhythm, the three bullet points, the “and that changes everything” at the bottom.
That’s AI slop. Content made in bulk with no one home. Your readers can hear it too, which is the annoying part, because some of what they’re hearing it in is genuinely yours.
I’ve put 750+ videos through my AI clone, onto a channel with 300,000+ subscribers as of October 2026, with zero demonetizations. So I’m the last person who’s going to tell you to put the tools down. I will tell you the 12 places the model’s voice is leaking into yours. Each one comes with a fix you can do this afternoon.
What AI slop means in 2026
AI slop is digital content of low quality, produced with artificial intelligence, usually in bulk. That is close to the dictionary wording. Merriam-Webster made “slop” its 2025 Word of the Year on December 15, 2025, which tells you how many people typed it into a dictionary in 2025 to check what everyone was complaining about.
The useful word in that definition is quality. Your readers don’t care whether a person or a model typed the thing. Neither does Google, as you’ll see in a moment. In May 2026, Graphite ran 55,400 English articles from Common Crawl, published between January 2020 and March 2026, through three detectors. AI written articles passed the human written ones in Q4 2025, at 50.9%. In Q1 2026 the share sat at 49.9%.
Now put that next to a second Graphite study from October 14, 2025. Across 31,493 keywords, 86% of the articles on Google’s first two result pages were human written. ChatGPT and Perplexity cited human written pages 82% of the time. Half of what gets published is mostly machine. About one in seven of what gets found is. The machine pages are getting written and then ignored, which is the whole problem in one sentence. If you want the citation side of that in detail, my generative engine optimization guide covers how answer engines choose a source.
Google gave the problem a name before the dictionary did. Its scaled content abuse policy went into force in March 2024, with the ranking rollout finished on April 19, 2024. Google reported 45% less low quality, unoriginal content in results once it landed. The policy wording covers pages produced at scale to rank, whether a model wrote them or a person did. The question for you, then, is whether a reader would keep the page open. The 12 tells below answer that question one paragraph at a time.
How to run the 12-tell check on your own post
Pick one post you published in the last 30 days. Read it against the 12 tells and give yourself a point each time one lands. Most of you will score somewhere in the middle, which is fine, because every one of them is fixable in a single sitting. The scoring table after tell 12 tells you what your number means and what to fix first. The sentence-level tells, em dashes included, live in my guide to how to humanize AI content. These 12 are the ones that give a whole page away.
The 12 tells of AI slop, with a fix for each
1. The page says the quiet part out loud
You’ve seen the flyer. A barista with six fingers, holding a cup that says “COFEE”, above a line about “your morning, reimagined.” Someone paid for that. Someone looked at it, nodded, and printed 500.
You’d never let it out the door for your own business, because you’d spot the fingers in a second. Text is harder. Text doesn’t have hands.
The written version is a stray “Certainly! Here’s a revised version” at the top of a paragraph, or an [insert company name] that never got inserted. In January 2024, Amazon was selling a recliner whose product title was an entire chatbot refusal, “I’m Unable to Assist with This Request it goes Against OpenAI use Policy and Encourages Unethical Behavior.” Futurism had the screenshots on January 12, 2024. One leftover line marks the whole page. A reader who spots it stops trusting every paragraph around it. So does any editor you ever pitch.
Search every draft for “as an AI” and for square brackets before it goes live. Thirty seconds, every time. If you’re also worried about what detectors make of the rest of the page, I covered how accurate those tools are in my post on whether AI detection is a scam.
2. Words you’d never say at dinner
Dmitry Kobak and his co-authors read 15.1 million PubMed abstracts published between 2010 and 2024. In the 2024 abstracts, “delves” appeared at 28 times the rate the earlier trend predicted. “Underscores” ran at 13.8 times. “Showcasing” at 10.7. Their lower bound estimate was that 13.5% of all 2024 abstracts had been through a language model. Those are scientists writing for other scientists. The model still talked them into “delves”.
Read your post out loud. Any word you’d feel silly saying to a friend over coffee comes out. A plainer one goes in. The sentence usually gets shorter in the process.
3. Every paragraph is the same size
Scroll your post with your eyes half closed. If every block is two sentences long and the same height on the screen, a model wrote the rhythm. People speed up and slow down when they write. A long paragraph sits next to a short one. Sometimes a single line stands alone because it earned the space.
The fix is mechanical. Merge two paragraphs that belong together and split the one that runs on. Then let one short line per section stand on its own.
4. The sentence that argues with a claim you never made
You’ve read this one a thousand times. “AI isn’t replacing writers. It’s replacing writers who don’t use AI.” The first half picks a fight with a person who isn’t in the room. The second half is the only part with any content in it. Models produce this shape because it sounds decisive without the writer having decided anything.
Delete the negated half every time you find it, then check whether the sentence that’s left still holds up. Usually it does. It also reads twice as sure of itself.
5. Everything comes in threes
Fast, reliable and affordable. Plan, build, measure. Models love a group of three because three feels finished, so a draft fills up with them until the reader can hear the drumbeat. One per page is plenty. Yes, I counted mine.
Cut each list of three to the two items that carry weight. Where the list is real, let it run to four or five. A reader can feel a list that was padded or trimmed to hit a rhythm.
6. The opening paragraph could sit on any website
Try this. Paste your first paragraph into a competitor’s post on the same topic and ask whether anyone would notice the join. If there’s no name and no date in it that belong to you, a model could have written it for anyone. It probably did.
Mine opens with your LinkedIn feed this time. Yours should open with the one thing you know about this topic that your competitor doesn’t. A customer conversation, or a number from your own dashboard that cost you something to learn. The reader decides in the first few lines whether a person is in the room.
7. No “I” anywhere and nothing with a date on it
Scan the page for the word “I”. If it never appears in a post published under your name, the reader has no reason to believe you were involved. Dates work the same way. “Many businesses have found” is a model phrase. “In March I cut our review time in half” is the kind of sentence only you can write. Readers forgive a lot of plainness for one of those.
My own version on this page is the 750+ clone videos line in the intro. Dated, and checkable on the channel. Put one of yours in every post, with the month attached.
8. Statistics with no address
“Studies show that 73% of consumers prefer personalized content.” Which study. Which year. Which consumers. A model will hand you that sentence with a straight face. It’ll hand your competitor the same one. Every number on your page needs a primary source with a date on it, linked where the number lands, the way the Kobak paper is linked under tell 2.
If you can’t find the original, the number comes out. Four sourced figures will carry a page. Fourteen orphans will sink it.
9. Headings that describe the furniture
“Tools that help you write faster.” “Benefits of AI content.” A heading like that tells the reader what the section is about and nothing about what they’ll get from it. Models default to labels. Rewrite each heading so it carries the payoff. “Benefits of AI content” becomes “What 90 hours a week of content looked like before the clone,” which is a heading I could put on a page tomorrow, because that was my week before the clone took the filming.
One rule I use on my own posts. If a heading could go on a competitor’s page unchanged, it isn’t finished.
10. The same article already exists 40 times
Search your target keyword and open the first page of results. If your draft covers the same ten points in the same order, you’ve written the 41st copy. Google already had 40. The model built your outline from those pages, which is why it matches them so neatly.
The fix starts before you prompt. Decide on the one point the top results get wrong or skip, then build the post around it. For the related question of whether a draft that follows those pages too closely counts as copying, I wrote a full answer on whether AI content is plagiarism.
11. Nothing in it a reader could disagree with
Read your post looking for one sentence somebody could argue with. If there isn’t one, the model hedged every claim into mush. “It depends on your goals.” “Results may vary.” Readers remember pages that pick a side, even the ones who pick the other side.
Say the thing you believe, in plain words. Then let the comments do their job. Tell 11 is the one most of you will score a point on, because hedging feels safe while you’re writing it. Readers skim straight past it.
12. Thirty posts a month with no reader in mind
Volume is the tell Google wrote a policy about. Its scaled content abuse rule, in force since March 2024, targets pages made in bulk to rank. In April 2025, Ahrefs crawled 900,000 newly published English pages and found AI content in 74.2% of them, with only 2.5% pure AI and 71.7% a mix. The mix is where most of you live, which is fine. The batch of 30 a month from a keyword export and a single prompt is where the trouble starts.
Publish fewer pieces. Give each one a reason to exist that you could say out loud to a customer. Then measure whether anyone read past the first screen. The pages that survive this are the ones readers return to. Those are also the ones answer engines cite.
Your score and what to fix first
Add up your points from the 12 tells. The order you fix things in is the useful part, so the table carries both.
| Points | Where you stand | What to do first |
|---|---|---|
| 0 to 2 | Clean. The page reads as yours. | Keep the 12-tell check in your pre-publish routine and move on. |
| 3 to 5 | Showing. A careful reader can tell a model did the typing. | Fix tells 1, 2 and 4 today. They take minutes each. |
| 6 to 8 | Obvious. Readers leave before the second heading. | Rebuild the brief before the next post, using the section below. |
| 9 to 12 | Google’s scaled content definition fits the page. | Stop the calendar. Start again with one post you’d defend in person. |
The cure for AI slop is a brief written before the prompt
Every one of the 12 tells traces back to the same moment. A prompt went into a model with nothing of you in it, so nothing of you came out. I spent 18 months building Claude projects that wrote like a polite stranger before I understood that, and a 15+ page document is what fixed it.
My clone works from a Voice DNA document that runs 15+ pages, built over 3 years from my own writing and 6,000+ chats. It holds how I phrase things, what I believe about AI, which words I’d never use, and the stories I tell on stage. That document, plus the review loop behind every video, is why 750+ clone videos have gone out with zero demonetizations.
Yours can start small. It needs your opinion on the topic, two numbers from your own business, the words you refuse to publish, and one real reader you’re writing to. Put that in front of the model before it writes a word. I walk through the full version, step by step, in how to humanize AI content. The 12 tells on this page are the test you run on the output.
Do that and the model stops producing the 41st copy. It starts producing a draft of you, which is the only draft worth publishing.
FAQ on AI slop
What is AI slop?
AI slop is low quality digital content produced with artificial intelligence, usually in bulk, with no person checking whether a reader needs it. Merriam-Webster made “slop” its 2025 Word of the Year on December 15, 2025. The term covers written pages and video alike. The defining feature is the quality of the output.
Is all AI generated content AI slop?
No. AI generated content becomes AI slop when it is produced at volume with nothing of the publisher in it and no reader in mind. Julia McCoy’s AI clone has published 750+ videos to a channel with 300,000+ subscribers as of October 2026, with zero demonetizations. Each one works from a 15+ page Voice DNA document.
Does Google penalize AI slop?
Yes, when it is produced at scale to rank. Google’s scaled content abuse policy has been in force since March 2024 and covers pages made in bulk whether a model or a person wrote them. Google reported 45% less low quality, unoriginal content in its results after the rollout finished on April 19, 2024. A single AI assisted post with real sources and a point of view is not what that policy targets.
How do I tell if my content is AI slop?
AI slop shows up in 12 tells you can check on any published page in a few minutes. The fastest ones to spot are leftover chatbot text such as “as an AI”, words you would never say aloud such as “delves”, paragraphs that are all the same length, and statistics with no linked source. Score one recent post against all 12 tells, then fix the cheapest ones first.
Can AI slop be fixed, or do I start over?
Usually it can be fixed. A page that scores up to 5 on the 12 tells needs an editing pass and one first person line with a date on it. A page that scores 6 to 8 needs the brief rebuilt before the next post. A page that scores 9 or more is faster to rewrite from a proper brief that holds your opinion, your own numbers, the words you refuse to publish and one real reader you’re writing to, because the problem started before the prompt.
Where to learn the whole system
AI Labs is where I teach the brief and the review loop behind every video my clone publishes, so your next post comes out sounding like you. Inside there are 60+ courses and two live builds every week, with 5,000+ students trained so far.
Join AI Labs and run your next post through the system before you hit publish.
Go back to that LinkedIn feed one more time. Five posts, one voice. Now you can hear exactly what’s missing from each of them. You’ve got 12 ways to make sure it’s never missing from yours.
