Ever pasted an AI draft into an email and thought, "that doesn't sound like me"?
A developer named Erik Kannike just trained a small AI model to stop writing like AI.
And he measured what changed.
Em dashes went from 20.5 per 1,000 words to zero.
The "not X, but Y" lines and the lists of three nearly vanished too: 7.8 down to 0.8, and 17.3 down to 1.1.
That's the useful part for the rest of us. The tells are countable, so you can find them and fix them yourself, in whatever AI you already use.
(Go on, open your last LinkedIn post and count the em dashes. I'll wait.)
Right after this: the seven tells, plus a prompt you can paste into any AI today.
Now a quick word from today's sponsor.
Leave Granola and get up to 12 months free of Wispr Flow Notetaker + Dictation
If you have paid time left on an individual Granola plan, we'll match it with a Wispr Flow subscription that includes Notetaker and dictation, and add bonus time, up to 12 months total. Sign in or create a Wispr account and submit proof of your plan to check eligibility.
Right. Back to it.
The seven tells
Em dashes everywhere.
"It's not X, it's Y." Arguing with a claim nobody made.
Lists of three. "Fast, flexible and scalable."
Announcing instead of saying. "Here's the thing." "It's important to note that."
Short fragments for drama. "Simple. Fast. Done."
Words nobody says out loud: delve, seamless, robust, leverage.
Bold labels on every bullet, and headings over sections only a few lines long.
The one prompt
Copy this, and fill in the two brackets:
Rewrite the text below so it sounds like a person wrote it.
Here is a paragraph I wrote myself. Match its voice:
[paste a paragraph you wrote, ideally an email]
Rules:
- No em dashes.
- No "not X, but Y" lines.
- Only list as many items as the text can back up.
- Cut any sentence that announces what's coming instead of saying it.
- Join short, choppy sentences where they belong together.
- Swap words like delve, seamless, robust and leverage for plain ones.
- Keep every fact, number and name exactly as it is.
Text:
[paste the AI draft]The paragraph you paste matters more than the rules. A guide from Anaconda puts it well: "Models imitate examples more reliably than adjectives like 'concise' or 'natural'."
So skip your best-ever writing and paste a real email you sent. That's the voice people already hear from you.
What makes it work better
Cut before you rewrite. Most AI filler is extra words, and deleting them is safer than asking for yet another version.
Don't rely on a list of banned words. Models route around them, so "delve" turns into "dig into."
Read it out loud. Lists of three and repeated points are easier to hear than to see.
(Yes, that's a list of three. They're fine when there really are three things to say.)
Want the model itself?
It's free and runs on your own computer. Search Hugging Face for "qwen3-4b-deslop-v4".
Fair warning from the author: the writing is "plain to the point of flat," and it still makes up details like studies and names in about one answer in thirty. Fine to play with. Your usual AI plus the prompt above is the easier win.
Make it stick
Save the prompt with your own paragraph already pasted in.
Next time a draft sounds off, it's one paste instead of ten minutes of fixing.
Getting AI to write like you starts with giving it your voice once, so you're not fixing every draft. That's the whole idea behind The One-Person Company Playbook.
It's the complete setup for running a business where AI does the work of a full team. Inside:
A company folder with your brief, pipeline tracker and invoice log already laid out, so any AI picks up where you left off
Thirty AI roles across six departments, from sales to finance, ready to use
Six jobs that run on a clock: the morning briefing, follow-ups, invoice checks and more
The step-by-step path from picking a business model to your first paying customer
Works on Claude, ChatGPT, Gemini or Grok, whichever you already pay for.
The catch: it takes about eight hours over a weekend to set up. After that, it's yours. $97, one-time (down from $297).
Talk tomorrow,
Zephyr
P.S. Run the prompt on your last LinkedIn post, then hit reply: which of the seven tells did you have the most of?