AI writing does not betray itself word by word. It betrays itself in structure and rhythm: the same cadences, the same sentence shapes, the same hollow vocabulary appearing on cue. Swapping words with a "humanizer" tool fixes the wrong layer. You have to fix the architecture.
AI writing tells are the structural and stylistic patterns that trained editors recognize as machine-generated: predictable rhythm, stock vocabulary, formulaic transitions, and uniform sentence length that signals no human hand shaped the draft.
What editors actually see on a first read
A seasoned editor does not look for any single word. They read for feel: does this draft breathe unevenly, the way a person thinks? Or does it march in even time, one qualified sentence after another? AI drafts march. Here is what triggers the flag, tell by tell.
The eight signs of AI writing (with before/after fixes)
1. Em-dash overuse
AI models treat the em dash as a general-purpose connector. It appears everywhere, several times per paragraph, often back to back.
Before: "Content strategy is critical — it aligns your team — and ensures every asset serves a goal — rather than filling a calendar."
After: "Content strategy aligns your team and keeps every asset tied to a goal, not a calendar slot."
One em dash per piece, used deliberately, reads fine. Four in a paragraph reads like a machine.
2. The stock vocabulary list
There is a cluster of words AI defaults to under pressure: delve, tapestry, leverage, robust, seamless, landscape (as metaphor), moreover, furthermore, game-changer, and the phrase "it's worth noting." These words are not wrong. They are just never the right choice for a human with a specific point.
Before: "It's worth noting that robust content strategy can leverage a diverse tapestry of channels."
After: "A real content strategy picks three channels and works them hard."
Scan your draft for this list. Replace every hit with the plainest word that carries the meaning.
3. The rule-of-three cadence, everywhere
"Fast, simple, and powerful." "Clear, consistent, and compelling." AI loves three-part lists because they feel complete. One or two in a piece is fine. When every paragraph closes on a triad, the rhythm becomes a chant.
Before: "Great headlines are specific, actionable, and clear."
After: "Great headlines are specific. That alone does most of the work."
Break the pattern. Make a point in one word or five. Skip the trinity.
4. The "not X, but Y" antithesis flip
This construction is genuinely useful once. AI uses it three times per post.
Before: "It's not about posting more. It's about posting better. It's not about reach. It's about resonance."
After: "Posting more is rarely the fix. What you publish matters more than how often."
One good antithesis lands. Two in a row feels like a speech. Three is a tic.
5. Anaphora runs
Three or more consecutive sentences opening with the same word or phrase.
Before: "You need a clear strategy. You need a consistent voice. You need a publication schedule you can keep."
After: "You need a clear strategy, a consistent voice, and a publication schedule that does not collapse in week three."
Deliberate anaphora, used once, can be powerful. A run of it is just a list pretending to be prose.
6. The formulaic "Conclusion" section
AI closes by restating everything it just said under a heading that literally says "Conclusion" or "Final Thoughts." This section rarely adds anything. It is scaffolding left in.
Before: "In conclusion, we've covered the importance of strategy, voice, and consistency in content marketing."
After: (Delete the section. End on your sharpest point. If you need a CTA, write one sentence and stop.)
The reader just read your post. Do not read it back to them.
7. Throat-clearing openers
"In today's fast-paced world." "In the age of AI." "As businesses navigate an ever-changing landscape." These lines exist to delay the actual point.
Before: "In today's digital landscape, content marketing has become more important than ever."
After: "Most content fails because it has no reason to exist."
Start where the tension is. The first sentence earns the read or loses it.
8. Uniform sentence length
This is the tell that most humanizer tools miss entirely, because it requires restructuring, not substitution. AI drafts run at roughly the same sentence length for paragraphs at a time. Real writing varies hard.
Before: "Content strategy requires planning. You need to know your audience. Consistency is important. Publishing regularly builds trust."
After: "Content strategy requires planning. Know your audience cold, publish on a rhythm they can set a clock to, and keep the work honest. That's it."
Read your draft out loud. If you hear a steady metronome, rewrite.
Why word-swap tools don't fix the problem
"Humanizer" tools work at the word level. They replace delve with explore and call it done. The tells listed above are not word choices, though. They are structural habits: the cadence of sentences, the architecture of paragraphs, the reflex to restate and qualify. A swapped vocabulary sitting inside AI sentence structure still reads as AI. You have to reshape the draft, not redecorate it.
How Ghosts handles this
I write for Ghosts, an AI ghostwriting platform, so I have a stake in saying this plainly: Ghosts runs an automated anti-tell pass with these exact rules on every draft before it reaches the editor. It flags em-dash density, stock vocabulary hits, sentence-length variance, and the structural patterns above. It does not replace editorial judgment. It hands the editor a cleaner surface to work from.
FAQ
What are the most common signs of AI writing?
Em-dash overuse, stock vocabulary (delve, robust, seamless), rule-of-three
cadences, and uniform sentence length are the patterns trained editors flag
first.
Does paraphrasing AI content make it undetectable?
Paraphrasing at the word level rarely removes the structural tells. Varying
sentence rhythm, cutting formulaic transitions, and rewriting the architecture
of paragraphs is what actually changes the read.
What is the fastest single fix for AI-sounding copy?
Read it aloud. If every sentence lands at the same pace, rewrite the longest
and shortest sentences in each paragraph until the rhythm breaks.
Ship the draft, then edit against this list
Print these eight tells. Run your next draft against them before it leaves your desk. You will not catch all of them on the first pass. That is fine. The goal is a draft that breathes like a person wrote it under deadline, not one that sounds like a tool filling a template. The difference is in the structure.
For more on writing that actually reads well, Google's writing quality guidelines for Search are a useful calibration point, not because they define good writing, but because they describe the floor.