Signs of AI writing, counted
Everyone has a list of AI tells. We counted them in 600 human texts and 900 AI drafts to see which ones hold up, and found they depend on the kind of writing.
Updated October 1, 2026
The ones that hold up
- "Utilize"
- How much more often in AI draftsAbout 33 times
- Sentences opening with Moreover or Furthermore
- How much more often in AI draftsAbout 18 times
- Sentences opening with Additionally
- How much more often in AI draftsAbout 18 times
- Em dashes in LinkedIn posts
- How much more often in AI draftsAbout 10 to 15 times
- Em dashes in explainers and answers
- How much more often in AI draftsAbout 5 times
- "Robust"
- How much more often in AI draftsAbout 6 times
- "X isn't just Y, it's Z" in LinkedIn posts
- How much more often in AI drafts32 times in 600 AI drafts, never in 300 real posts
| Sign | How much more often in AI drafts |
|---|---|
| "Utilize" | About 33 times |
| Sentences opening with Moreover or Furthermore | About 18 times |
| Sentences opening with Additionally | About 18 times |
| Em dashes in LinkedIn posts | About 10 to 15 times |
| Em dashes in explainers and answers | About 5 times |
| "Robust" | About 6 times |
| "X isn't just Y, it's Z" in LinkedIn posts | 32 times in 600 AI drafts, never in 300 real posts |
Tells depend on the genre
- "Utilize"
- Human explainers0.4
- AI explainers12.3
- Human LinkedIn0
- AI LinkedIn (Gemma / Qwen)0.6 / 0.6
- Moreover, Furthermore
- Human explainers0.6
- AI explainers11.2
- Human LinkedIn0
- AI LinkedIn (Gemma / Qwen)0.9 / 0.3
- Em dashes
- Human explainers7.7
- AI explainers41.8
- Human LinkedIn3.6
- AI LinkedIn (Gemma / Qwen)55.6 / 35.7
- "Isn't just X, it's Y"
- Human explainers0
- AI explainers0.3
- Human LinkedIn0
- AI LinkedIn (Gemma / Qwen)3.3 / 6.4
| Per 10,000 words | Human explainers | AI explainers | Human LinkedIn | AI LinkedIn (Gemma / Qwen) |
|---|---|---|---|---|
| "Utilize" | 0.4 | 12.3 | 0 | 0.6 / 0.6 |
| Moreover, Furthermore | 0.6 | 11.2 | 0 | 0.9 / 0.3 |
| Em dashes | 7.7 | 41.8 | 3.6 | 55.6 / 35.7 |
| "Isn't just X, it's Y" | 0 | 0.3 | 0 | 3.3 / 6.4 |
AI explainers sound like a textbook trying too hard: utilize, moreover, robust. AI LinkedIn posts barely use those words. Their tells are em dashes, the "isn't just X, it's Y" frame, and a question answered in the next breath ("The result? A team that…"), which appeared 10 times in 600 AI posts and once in 300 real ones.
The ones that don't
- Lists of three. In explainers they appeared 23.6 times per 10,000 words in human writing and 23.8 in AI drafts. People write in threes too. In LinkedIn posts AI used somewhat more (about 10–11 against 6), still not enough to judge a single post.
- "Delve". It appeared twice in the human explainers and never in the AI drafts. The famous tells go out of date as models are tuned away from them.
- "Not just X, but Y". Human explainers used it 1.1 times per 10,000 words, AI drafts 0.6.
- "In order to". Human explainers used it 1.7 times per 10,000 words; the AI drafts never did. Wordiness is human too.
Most AI text has none of these
Only 19–22% of the AI LinkedIn drafts contained a single stock word or frame from our list, or about 28% if you count heavy em dash use. Yet a detector caught around 90% of them at a 1-in-20 line. The signs you can spot by eye are a small part of what makes text read as AI. The rest is rhythm and word choice spread across every sentence.
The same goes the other way: 16% of the human explainers and 8% of the real LinkedIn posts had at least one stock word or frame. Finding one in a piece proves nothing.
How to use the signs
- 01Look for density, not presence. One em dash is nothing; four in a 100-word post is a pattern.
- 02Compare against the genre. A formal explainer and a LinkedIn post have different normal habits.
- 03Look for what's missing: specific names, dates, numbers, and anything that only the writer could know.
- 04Treat signs as a reason to read closely, never as proof someone used AI.
How we counted
Human texts: 300 LinkedIn posts from 2021 and 300 well-edited pieces from before 2022 (educational pages, forum answers, Wikipedia openings). AI drafts: Gemma 4 writing or rewriting the same 600 pieces, and Qwen3.5 doing the same for the 300 LinkedIn posts. The tell list is the one in Macaron's prose de-slop. Counts are raw matches per 10,000 words, except lists of three, which count only past a few per 1,000 words.
Questions
Is the em dash a sign of AI writing?+
In bulk, yes. AI drafts used 5 to 15 times as many as people writing the same kind of piece. One em dash proves nothing; four in a short post is a pattern.
Are lists of three a sign of AI?+
Not in explainers: human and AI writing used them at the same rate. In LinkedIn posts AI used somewhat more, but not enough to judge one post by.
Is "delve" still a sign of AI?+
Not in our data. It appeared twice in 300 human explainers and never in the AI drafts. Model makers tune away from words that become famous tells.
What are the most common signs of AI writing on LinkedIn?+
Heavy em dash use, "it isn't just X, it's Y", and a rhetorical question answered right away. "Utilize" and "Moreover", the classic explainer tells, barely show up there.
Can I prove something was written by AI from these signs?+
No. 16% of well-edited human explainers had at least one stock word or frame, and most AI posts had none. Signs are a reason to look closer, not evidence.
Is there a tool that checks for these signs?+
Yes. The free AI phrase checker on this site highlights them and fixes the safe ones, in your browser.