How to make AI writing sound human
The specific tells that make generated text obvious — em dashes, a fixed vocabulary, sentences that are all the same length — and what to do about each one.
Updated 2026-08-03
Text that reads as machine-written does so for a small number of countable reasons, and none of them is mysterious. It is punctuation no keyboard produces easily, a vocabulary of about forty words used far more often than people use them, sentences that are all roughly the same length, and a handful of structural habits. Fix those four and the writing reads as yours. One thing this guide will not tell you is how to beat a detector, because detectors do not work — and understanding why is the most useful thing in it.
The tells, in order of how much they give away
| Tell | Why it stands out | What to do |
|---|---|---|
| Em dashes | Used several times more often than in human prose, and clustered in the same three places | Replace with commas or periods; read the result aloud |
| A fixed vocabulary | delve, tapestry, realm, testament, seamless, robust, myriad, pivotal | Rewrite the sentence, not the word — the phrasing around it is the problem |
| Uniform sentence length | Human writing swings between five words and thirty; generated text settles into a rhythm | Cut one sentence in half. Join two others. Vary deliberately |
| "Not just X — it's Y" | The single most recognizable generated sentence shape | Say the thing you actually mean and delete the setup |
| Opening connectives | Furthermore, Moreover, Additionally, In conclusion | Delete them. A paragraph that needs one is not following from the last |
| Curly quotes and ellipses | Typographic characters most keyboards do not produce | Straighten them, especially in anything a machine will read |
| Invisible characters | Zero-width spaces arrive with copied text and survive everywhere | Strip them — they also break search and comparison |
| A tidy three-item list | Every list has exactly three items, every time | Use two. Use five. Use a sentence |
Why "undetectable" is a lie you can safely ignore
Detection does not work, in either direction, and this is settled rather than contested. OpenAI withdrew its own AI classifier in 2023 citing low accuracy. Studies since have found high false-positive rates on human writing, with a documented bias against non-native English speakers, whose more regular sentence construction reads to a classifier exactly like generated text. Real students have been accused on this evidence.
- A tool promising undetectable output is selling you protection from something that cannot reliably detect you anyway.
- A tool promising 99% accurate detection is claiming an accuracy that no published evaluation supports.
- The useful goal is not evasion. It is text that reads well — which is worth doing whoever wrote the first draft.
The em dash question, answered properly
Em dashes are not wrong. Good writers use them. The problem is density and placement: generated text uses roughly four to six per thousand words against one or two in ordinary prose, and puts them in the same positions every time. If you like them, keep one or two per page in places you chose deliberately. Replace the rest.
| Character | Looks like | What it is for |
|---|---|---|
| Em dash | — | Setting off a clause: "the result — all of it — held up" |
| En dash | – | A range: 1939–45, pages 30–34 |
| Hyphen | - | Joining words: well-known, twenty-one |
| Minus | − | Arithmetic. A different character again |
Measure before you edit
Editing by feel means you stop when it feels different rather than when it is different. Counting gives you a place to start and a way to tell whether the second pass helped. Sentence-length variance is the most useful single number: below about 0.4 the rhythm is flat regardless of who wrote it.
The edit that actually matters
Everything above is cosmetic. The thing that genuinely separates writing worth reading from writing that merely parses is specificity — a real number, a named example, a concrete consequence. Generated prose is fluent and general. It says a strategy is robust rather than saying it cut onboarding from nine steps to three. Replace two abstractions per paragraph with something specific and the piece stops sounding like anything but itself.
Questions
Will removing em dashes get my work past an AI checker?
Possibly, and that is not a good reason to do it. Those checkers produce false positives on human writing constantly and miss edited model output constantly. Edit because the text reads better, not to game a number that is close to a coin flip on any single document.
Is it wrong to use AI to write a first draft?
That is a question about your assignment, your employer or your publisher — not about the writing. What this guide addresses is the separate, universal problem that generated prose has recognizable habits that make it duller to read.
Which single change makes the biggest difference?
Sentence-length variance, by a distance. Em dashes are the most talked about because they are the most visible, but a page where every sentence is eighteen words long reads as machine-made even with perfect punctuation.
Do these tells apply to every model?
The vocabulary shifts between model families and over time — "delve" became a marker only around 2024. The structural tells are more durable: uniform rhythm, tidy three-item lists, and the "not just X" construction have persisted across every generation so far.