Why AI Text Sounds Wrong (Even When It's "Right")
AI can hit grammar, structure, and clarity. Yet it often rings false to a working writer's ear. The tone feels eager, the metaphors feel inflated, and the voice sounds like a brand trying to hug you.
If you've felt that uncanny sameness spreading across emails, press releases, and even essays, you're not imagining it. Models have learned patterns that signal "quality," then overuse them until they turn into tells.
The Tells You Can Spot in Seconds
- Punctuation crutch: A heavy reliance on em dashes to create drama and rhythm.
- Template logic: The not-X, but-Y sentence, used so often it feels like a reflex.
- Rule of threes mania: Everything arrives in neat triplets. It's tidy, predictable, and everywhere.
- Hollow sensory vibes: Obsession with "quiet," "echo," "ghost," "liminal," and a soft "hum," even in loud scenes.
- Overfitted word choice: An unusual fondness for a four-letter verb starting with "d" that rhymes with "shelve," and fancy fabric metaphors that get pasted onto anything.
- Stock names in fiction: Sci-fi leads called Elara Voss or Kael, love interests named Echo, and villains with "Ghost" in the title.
- Self-Q cadence: Mid-paragraph rhetorical questions followed by breathless praise of your "great point."
- Snark template: "An X with Y and Z," where Y or Z are obvious or nonsensical.
- Cultural bleed: Phrases native to one region or institution showing up somewhere they rarely appeared before.
Why Models Keep Writing This Way
Models learn by correlation. If certain punctuation, cadences, and words show up more in "high-quality" texts, the system assumes more of those equals more quality. That's overfitting: a pattern becomes a crutch.
Early systems famously confused outcomes with causes. Ask for "more jokes," and they described tickling. Today's models are stronger, but the core behavior remains. They chase the signals of good writing instead of the experience that produces it.
The Em Dash Problem
Writers love it. Editors indulge it. Models notice. Then they oversupply it. The fix isn't to ban the character. It's to earn emphasis with structure, contrast, and clean syntax before you reach for shortcuts.
Word Drift and Cultural Leakage
Language isn't uniform. That "d-word" is normal in some Englishes, including Nigerian usage, and now shows up far more often across scientific abstracts and business writing. Models amplify these shifts at scale.
Similar bleed happens with phrases like "I rise to speak" crossing contexts. The takeaway for pros: audience and setting matter more than ever. Calibrate tone to the room, not to what the model thinks plays well everywhere.
How to Keep Your Writing Human
- Start with a scene you actually lived. One sensory detail you noticed beats five abstract adjectives.
- Prefer concrete nouns and active verbs. Swap "connection" for "call," "walk," "handshake," "transaction," or whatever actually happened.
- Kill templates on edit. Remove the not-X, but-Y structure unless contrast is the point.
- Limit the em dash. If you see more than one per 400-500 words, you're leaning on it.
- Break the triplets. Keep one. Cut two. Or fuse them into a sharp single line.
- Ban the stock names. If your sci-fi lead is Elara Voss or Kael, rename them.
- Ground metaphors in touch and place. If you can't smell it, hear it, or see it, don't compare it.
- Delete corporate warmth. If a sentence could be taped to a storefront window during a closure, rewrite it.
- Write one intentionally "imperfect" line. A plain sentence resets tone and makes the rest believable.
- Read aloud. If it sounds like a motivational LinkedIn post, you know what to do.
A 5-Minute Anti-AI Edit Pass
- Search for "quiet," "echo," "ghost," "liminal," and "hum." Keep only what serves the scene.
- Find the em dash character (-). Replace half with commas, periods, or reworked sentences.
- Cut one-two triplets per page. Compress or vary the rhythm.
- Swap the "d-word" for verbs like "examine," "look at," or "explore."
- Replace abstract metaphors with one specific, observed detail.
- Remove autopilot empathy lines. Add a concrete fact the reader didn't know.
Where AI Helps (Without Smudging Your Voice)
- Outlines and counterarguments: Generate angles, then choose and reshape.
- Research starting points: Use it to assemble sources and definitions, but verify and rewrite.
- Line-level cleanup: Grammar and clarity are fine. Style is yours. Keep it that way.
Protect Your Signal
We absorb the language we read. If feeds are saturated with machine cadence, that cadence creeps into ours. The antidote is intent: concrete experience, sharp edits, and a few hard rules you enforce on yourself.
Your edge isn't polish. It's presence. Bring scenes you've lived, details you've noticed, and judgments you're willing to own. That's the part a model can't fake.
Helpful Links
Your membership also unlocks:
