TL;DR: AI text reads as AI text for predictable, fixable reasons: uniform sentence length, a small set of tell-tale words, passive constructions, abstraction stacked on abstraction, and em-dashes everywhere. Fix those seven things and most readers — and most detectors — stop flagging it. None of this is about adding "personality." It's about removing the machine's defaults.
I've read a lot of AI-assisted copy this year. Pitch drafts, bylined articles, comment lines for journalists. The good news: AI gets you to a usable first draft fast. The bad news: a first draft is exactly what it looks like, and editors can smell it from across the room.
Here's the thing nobody tells you. "Humanizing" AI text is not a creativity problem. It's a subtraction problem. The model has a set of statistical habits, and those habits are the tell. Strip them out and the prose stops announcing where it came from.
Below are the seven techniques I actually use, in the order I apply them. Every one has a before/after, because abstract advice about "varying your style" helps nobody.
1. Vary your sentence length (burstiness)
This is the single biggest tell. Models produce sentences of remarkably uniform length — usually 15 to 25 words, comma-spliced into a steady rhythm. Human writing is bursty. A long, winding sentence that builds an idea across several clauses gets followed by a short one. Like that.
Before:
The campaign launched in March and generated significant engagement across multiple channels, and the team continued to optimize the messaging throughout the quarter, which resulted in a steady increase in qualified leads over the following weeks.
After:
The campaign launched in March. Engagement climbed across every channel we touched, and we kept tuning the messaging through Q1 — qualified leads rose week over week. Slow at first. Then not slow at all.
Read both out loud. The first one has no pulse. The second one does, because the sentence lengths swing from three words to twenty.
2. Cut the LLM filler vocabulary
There's a vocabulary of words that models reach for far more often than people do. The moment an editor sees three of them in a paragraph, the piece is marked. Delete on sight:
- delve (nobody delves; they look into it)
- tapestry / rich tapestry
- leverage (use "use")
- robust, seamless, streamline
- navigate the complexities of
- in today's fast-paced world / in an ever-evolving landscape
- it's important to note that
- a testament to
- unlock (unlock value, unlock potential)
- realm
Before:
In today's fast-paced world, brands must leverage robust, seamless solutions to navigate the complexities of the modern media landscape.
After:
Brands need media tools that work under deadline pressure. Most don't.
The after-version says more. It just refuses to pad.
3. Prefer active voice
Models default to passive constructions because passive voice is safe and agentless. It's also limp. Name who did what.
Before:
A decision was made to restructure the pitch, and the new angle was approved by the client.
After:
We restructured the pitch. The client approved the new angle the same afternoon.
Passive voice isn't always wrong — sometimes the actor genuinely doesn't matter. But the model uses it as a default, not a choice, and that's the problem.
4. Choose concrete nouns over abstractions
AI text floats. It talks about "solutions," "engagement," "value," "impact," "synergies" — nouns with no edges. Concrete nouns give the reader something to hold.
Before:
Our solution drives meaningful engagement and delivers measurable value across the customer journey.
After:
Reporters reply to our pitches. Two of last month's landed in the FT and Bloomberg. That's the metric the client cares about.
When you can't picture it, it's an abstraction. Swap it for the thing.
5. Mind the rhythm and cadence
Beyond sentence length, prose has a music to it. Models write in a flat 4/4 — every paragraph the same shape, every transition a tidy connective ("Furthermore," "Moreover," "Additionally"). Real writing changes tempo. Start a paragraph with a fragment. End one on a hard, short beat. Drop the connective entirely and let two sentences sit next to each other.
Before:
Furthermore, the data indicates strong performance. Additionally, the metrics suggest continued growth. Moreover, the trend is expected to persist.
After:
The numbers are strong. Growth held through the quarter, and nothing in the data says it stops here.
Three sentences of throat-clearing became one with momentum.
6. Remove the hedges
Models hedge constantly because hedging lowers the odds of being wrong. "Arguably," "it's worth noting," "in many cases," "some might say," "generally speaking." Each one is a small apology. A senior voice doesn't apologize for having a position.
Before:
It's worth noting that this approach is arguably one of the more effective methods, and in many cases it can generally lead to positive outcomes.
After:
This approach works. We've shipped it on six accounts and it held every time.
If you believe the claim, state it. If you don't, cut it. There's no third option worth keeping.
7. Em-dash discipline
Models love the em-dash. They scatter it through every paragraph the way a nervous speaker says "um." One or two per piece is fine — used well, it's a strong tool. Six in three paragraphs is a fingerprint.
When you find a string of em-dashes, ask what each one is doing. Most can become a period, a comma, or a colon.
Before:
The pitch — which we'd been refining for weeks — finally landed, and the journalist — clearly interested — asked for more — a sign the angle worked.
After:
The pitch we'd refined for weeks finally landed. The journalist was interested and asked for more — a sign the angle worked.
One em-dash, doing real work. The rest, gone.
Don't run a global find-and-replace to delete every em-dash. That over-corrects and creates its own flat, choppy tell. The skill is selective: keep the one that earns its place, cut the four that don't.
The workflow, in order
When I edit an AI draft, I run these as passes, not all at once:
- Read it aloud once. Mark every place the rhythm flatlines.
- Search-and-destroy the filler vocabulary (technique 2) and hedges (6).
- Flip the obvious passives (3) and the worst abstractions (4).
- Rewrite for burstiness and cadence (1, 5) — this is where the draft actually becomes yours.
- Last pass: em-dashes (7).
Two passes through a 600-word draft takes me about ten minutes and the result reads like a person wrote it, because by then a person has.
Why this matters beyond the detectors
People focus on AI detectors, but that's the smaller reason. The bigger one: every tell-tale pattern is also just bad writing. Uniform sentences are monotonous. Filler vocabulary is empty. Passive voice is evasive. Abstractions are unconvincing. Hedges are weak.
Strip the AI defaults and you don't just dodge detection — you end up with prose that's clearer, faster, and more persuasive than what most people write from scratch. That's the actual win. The model gives you a draft in seconds; your job is to make it earn the byline.
