Signs of AI writing: 10 tells and how to fix each
Readers spot machine-written copy in a sentence or two. The ten patterns that give it away, and what to write instead.
The short version
- It is patterns, not single words. One 'delve' is a word. Five tells in a paragraph is a voice.
- Structure gives it away as much as vocabulary. Every list in threes, every paragraph the same length.
- The fix is specifics. Most tells are what a model writes when it has nothing concrete to say.
- Detectors are not the point. A reader deciding whether to trust you is.
The signs of AI writing are patterns more than single words: stock vocabulary such as "delve" and "elevate", the "it is not X, it is Y" turn, lists that always have three items, openers that restate the question, hedging on every claim, paragraphs of identical length, and endings that summarise instead of adding anything. Most of them appear where a writer had nothing specific to say. The fix for nearly all of them is a concrete detail.
Why do readers notice so fast?
Because they have read thousands of these by now. A reader does not run a checklist. They feel the rhythm of a paragraph they have seen before and stop trusting it. That is the real cost: not a detector flag, but a reader who skims the rest.
What are the 10 tells?
1. Stock words
delve, tapestry, testament, elevate, unlock, seamless, robust, landscape, realm, navigate (when nothing is being navigated). Each is a fine word once. Together they are a signature.
Fix: use the plain word. "Look at" instead of "delve into". "Improve" instead of "elevate". If no plain word fits, the sentence probably says nothing.
2. "It's not X, it's Y"
"It's not just a tool, it's a movement." Models love the reversal because it sounds like insight. It usually is not.
Fix: state Y directly. If X was never a belief anyone held, drop it.
3. Everything in threes
"Faster, simpler and more reliable." Three adjectives, three bullets, three examples, every time.
Fix: use the number you actually have. Two reasons is fine. Five is fine. Three because it sounds complete is the tell.
4. The throat-clearing opener
"In today's fast-paced digital world..." "When it comes to pricing..." The first sentence restates the topic instead of saying something about it.
Fix: delete the first sentence and see whether the second one is a better start. It usually is.
5. Hedging everything
"This can potentially help some businesses in certain situations." Every claim softened until it means nothing.
Fix: hedge where you are actually unsure, once. Say the rest plainly, or cut it.
6. Identical rhythm
Every paragraph three sentences. Every sentence about fifteen words. Read aloud, it sounds like a metronome.
Fix: vary it. A one-line paragraph. A long sentence followed by a short one. Like this.
7. Empty enthusiasm
"Exciting", "game-changing", "revolutionary", exclamation marks on ordinary statements. Energy with nothing behind it.
Fix: replace the adjective with the fact that made you want to use it. Not "a game-changing feature" but "it cut our export time from an hour to four minutes", if that is true.
8. Generic examples
"For example, a small business might use this to improve efficiency." An example that could apply to anything.
Fix: a named situation. Who, which number, what happened. If you do not have one, say less.
9. The summary ending
"In conclusion, by following these steps you can..." The last paragraph repeats the first.
Fix: end on the next action, an open question, or the strongest line you have. Then stop.
10. Punctuation habits
Em dashes everywhere, colons before every list, title case on every heading. Not wrong in themselves, but in a page full of them, readers see the pattern.
Fix: use a comma or a full stop where a dash crept in. This blog does not use em dashes at all, for that reason.
How do you check a draft quickly?
| Check | What to look for |
|---|---|
| Read the first line | Does it restate the title? Delete it. |
| Search the stock words | Replace each with the plain version. |
| Count list items | Is every list three items long? |
| Read aloud | Does every paragraph sound the same? |
| Find every claim | Is there a number or example behind it? |
| Read the last line | Does it add anything new? |
How to stop AI content sounding generic covers the step before this one: giving the model enough context that it writes fewer tells in the first place. AI tools for content creation covers which tools to draft with, and how to write a sales page is where these tells cost the most.
Copix, if you want your assistant to catch these itself
A note on fit first. If you write rarely, or you are happy running this checklist by hand, you do not need a tool for it. Copix does not write the final copy, it runs no model of its own for the copy tools, and it is English only.
Copix plugs into the AI you already use, such as Claude, ChatGPT or Cursor, and gives it a pass that flags AI tells in a draft and returns a verdict, plus copywriting frameworks, hook families, headline formulas and story structures. The Complete plan adds SEO checks for queries, drafts and live pages. Your model keeps doing the writing.
Frequently asked questions
What are the most common signs of AI writing?
Stock words such as delve, tapestry and elevate, the 'it is not X, it is Y' construction, lists that always come in threes, generic openers, heavy hedging, and endings that summarise instead of saying something new.
Are AI detectors reliable?
Not reliable enough to judge a single text. They flag human writing as AI and miss edited AI writing. Readers are a better test: if it reads as generic, it does not matter who wrote it.
Does using em dashes mean a text was written by AI?
No. People used them long before models did. Models use them so often that a page full of them now reads as machine-made to many readers, which is why some writers avoid them.
Can I make AI writing undetectable?
The better goal is making it good. Copy with real specifics, a clear point and a voice of its own stops reading as AI, because the tells are mostly what fills the gap when specifics are missing.
Does Google penalise AI writing?
Google's stated position is that it ranks helpful content regardless of how it was made, and acts against content made mainly to manipulate rankings. Generic AI copy tends to fail the helpful part.
The software I build and run
Everything on this blog is tested on these first. Each one was built alone, shipped in public and used every day, by me first.
- Easy Reels easyreels.ioYour clips and any song become a vertical video with every cut on the beat.
- BatchTone batchtone.comOne LUT across a whole folder of photos and video on macOS, with a dial per file.
- bukpress buk.pressA Markdown draft in, a typeset PDF and EPUB out, with your fonts and cover.
- Nexin nexin.siA content department inside Claude or ChatGPT, reading your brand files on every job.
- Copix copix.siCopywriting frameworks, a pass that strips AI tells, and SEO checks for your assistant.
- Valta valta.siOne vault of Markdown notes that every AI you use can save to and read back.
- Zelect zelect.ioAn affiliate storefront for your link in bio: your picks, laid out like a shop.