How to spot AI-written text (and stop your own from getting caught)
4 signs of "plastic" text, one myth about the em dash, and a humanizing checklist. Plus a mini-test: does your own writing give itself away?
01 — The question
Why is AI text recognizable in three seconds?
ChatGPT is a giant model trained on billions of texts. It sees your request and then, essentially, guesses which words are most likely to come next so that the result holds together.
It doesn't think "this is more beautiful" or "this makes me sound more human." It does what looks right in its mathematical picture of the world. And "the most likely thing" is, most of the time, exactly cliché and template.
That's why all out-of-the-box texts share the same tics. A reader may not articulate what's off — but their nose doesn't lie: "plastic." The good news: the tics are countable. There aren't many, and each one can be removed. Let's go through them.
02 — The signs
Which 4 signs give the AI away?
Sign 1. Choppy. Short. Sentences. The classic: "You want to start writing. But you don't know how. That's okay. We'll figure it out." Real people write in compound sentences, with tangents. (Doubly painful for me — I wrote in a punchy rhythm for years, and now it reads as AI 😂)
Sign 2. Stock transitions. "Sound familiar?", "And you know what?", "Now pay attention." The model adores these bridges — they're the "most likely" connectors across billions of texts.
Sign 3. Sterile correctness. Every comma in place, no "well," no "who knows." You read it and everything is polished to a squeak. A live human doesn't write like that even in a CV.
Sign 4. The "not just… but…" template. Left unchecked, the model produces: "This isn't just a tool. It's your new assistant." Or worse: "This isn't just a course. It's a step toward the life of your dreams."
You're not just an expert. You're a brand. Sound familiar — that
feeling when content doesn't work? That's okay. We'll figure it
out. And you know what? It's simpler than it seems.One paragraph like that is a coincidence. Three in a row is a diagnosis.
03 — The test
Does your text give itself away? Check in 30 seconds
Open the last text you wrote with AI and answer five questions honestly. Most people are sure "mine doesn't show" — which is exactly why it's worth checking:
04 — The myth
Is the em dash a sign of AI?
There's a theory going around Threads: if a text uses a proper long em dash instead of a hyphen — it's definitely AI.
Oh come on. It's just the correct punctuation mark)) Copywriters and editors used em dashes long before ChatGPT. If paragraphs get declared "an AI sign" tomorrow, do we stop breaking up text?
Em dashes, correct commas, subheadings, lists. These are signs of editing — human or machine, doesn't matter.
Repeating patterns: the same transitions, the same rhythm, the same constructions — in every paragraph, every post, across thousands of authors at once.
The difference is fundamental: AI output isn't "too correct," it's "too identical." Correctness is the norm. Uniformity is the diagnosis.
05 — The fix
How do you remove the "bot" from a text in 10 minutes?
How I handle it: I always come back to the text by hand. I walk through it and add humanity — my favorite words, little phrases, funny asides. Sometimes I deliberately leave small rough edges, because that's what makes text alive. Here's the whole pass as a checklist — tick as you go, and download the file:
- Find and rewrite "Sound familiar?", "And you know what?", "Now pay attention" with your own transitions
- Remove every "not just… but…" — say the thought directly
- Merge some choppy sentences into normal compound ones
- Add 2–3 words or phrases you actually say in real life
- Insert one live detail: an incident, a number, a name
- Leave one rough edge — a "well," a parenthesis, some self-irony
- Read it aloud: do you stumble where you'd stumble in real speech
And the systemic fix, so there's less to edit: put your speaking style into the prompts — give the model your texts as references, and the template tone breaks at the input stage. I covered the one-time setup in the article on making ChatGPT write in your voice.
06 — The stakes
Why does this matter if AI content gets views anyway?
Fair question: template AI texts with trigger headlines rack up thousands of views — why bother?
Because reach and trust are different currencies. A template text can go viral, but it leaves no trace of "a real person with real experience wrote this." And trust is the only thing that converts into clients, deals and a relationship with your audience. The more machine text there is around, the more a live voice costs: "sounds like a human" is becoming a luxury that both algorithms and readers reward separately.
Rule: the AI writes the draft — the final touch is always yours.
A 10-minute manual pass (checklist above) is the difference
between "another AI text" and "oh, this is definitely you."AI text is betrayed not by dashes and commas but by repeating patterns: choppy rhythm, stock transitions, sterility, "not just… but…". Run your text through the mini-test, walk the checklist — and keep writing with AI in a way that stays you.
FAQ
Does the em dash really give away AI?
No — that's a myth. The em dash is simply correct typography that copywriters and editors used long before ChatGPT. AI is betrayed by repeating patterns — identical transitions, rhythm and constructions in every paragraph — not by literate punctuation.
What's the most common sign of AI text?
Choppy short sentences ("That's okay. We'll figure it out.") and the "not just… but…" construction. If a text has both, plus stock transitions like "And you know what?", readers will almost certainly sense the plastic — even without articulating why.
How do I write with AI so the text doesn't get caught?
Two layers: at the input — give the model your speaking style and reference texts, breaking the template tone early; at the output — return to the text by hand: your words, live details, a couple of rough edges. The final touch is always human — that's a rule, not a wish.
Should I deliberately leave mistakes in the text?
Not mistakes — rough edges: a "well," a self-ironic parenthesis, a conversational turn. A perfectly polished text with no living trace is itself a sign of generation. One or two natural bumps make text human — as long as they're yours, not random.
