Why Your Writing Sounds Like AI (and How to Fix It)
Someone replies to your message with "did you write this yourself?" and you did. Or you read a paragraph you wrote at midnight and it sounds like nobody wrote it.
The tells are real, and the em dash is only the most famous one. The rest are rhythm, specificity, and a small set of sentence shapes that models produce constantly because they are safe. This guide covers what actually gives writing away, six rewrites, and the one fix that works when nothing else does.
Start with the em dash
There is a good argument that the em dash is innocent. It is ordinary punctuation, every phone keyboard inserts it automatically, and writers have used it for three hundred years.
That argument has lost. Readers now treat a run of em dashes as a signature, and a reader who suspects the text stops reading it on its merits. So cut them. A comma, a colon, or a full stop does the same work, and the sentence is usually shorter for it. This is the cheapest fix in the article and the only one that is purely mechanical.
What the em dash is standing in for is the rate. Three of anything in four paragraphs, each doing the identical job of bolting a qualifying aside onto a sentence that was already finished, is a pattern, and a pattern is what gets noticed. The same goes for the semicolon, the colon, and the bulleted list. Any device becomes conspicuous when it is the only device.
Curly quotes, ellipses, and non-breaking spaces are not tells. They are what text editors produce.
What actually reads as machine-written
Five things, roughly in order of how quickly a reader notices.
Every paragraph is the same length. Model output settles into three-to-four-sentence blocks, one idea each, evenly spaced down the page. Human writing lurches. A paragraph runs eleven lines because the thought needed eleven, then the next one is four words long. That unevenness is the single strongest signal of a person, and it is the thing most editing tools quietly remove.
Nothing in it could only have been written by you. No date, no number, no name, no thing that happened. "Improving your onboarding flow can significantly increase user retention" is true, checkable by nobody, and could be pasted into any document in the world. "We moved the email step to the end and week-one retention went from 31% to 44%" can only come from someone who was there.
The contrast frame, on repeat. "It's not just X, it's Y." "This isn't about X. It's about Y." One is rhetoric. Three in a page is a template. The related one is the false-correction opener: "The problem isn't that people don't read. It's that..."
Threes everywhere. Three adjectives, three bullets, three examples, three clauses. The rule of three is genuinely effective, which is why it shows up in training data constantly and why models reach for it when two items or five would be the honest number.
The summarising close. "In conclusion", "By following these steps, you can", "Ultimately, the key is": a final paragraph that restates the piece instead of ending it. Also the opening version: a first paragraph explaining what the piece will cover, before covering it.
Vocabulary matters less than any of these, but it is worth knowing which words have become markers through overuse: delve, tapestry, testament, underscore, navigate, landscape, realm, leverage, robust, seamless, multifaceted, pivotal, and the phrase it's worth noting that. None is a bad word. All are now conspicuous, and none of them was going to be your first choice anyway.
Why it happens
Models are trained to produce answers that feel balanced and complete. Hedged claims are safer than specific ones, symmetrical structure reads as competent, and a summary at the end feels finished. Every one of those instincts is also what a tired human does at 11pm, which is why your own writing drifts in the same direction under deadline. The fix is the same for both.
Six rewrites
1. Replace the general claim with the thing that happened
Before
Clear communication is essential for effective team collaboration, and investing in it pays dividends over time.
After
We were re-explaining the same deploy process in standup twice a week. Writing it down once took an afternoon and the question stopped.
The first sentence is unfalsifiable. The second has a frequency, a cost, and an outcome, and no model could have produced it without you.
2. Break the rhythm
Before
The redesign improved the checkout flow. Users found the new layout easier to navigate. Conversion increased across all segments. The team was pleased with the results.
After
The redesign moved payment to a single screen, which was the part everyone argued about for two weeks and which turned out to be the only change that mattered. Conversion went up 8%. Mostly on mobile.
Four sentences of near-identical length read as generated. One long, one short, one fragment reads as someone thinking.
3. Cut the frame, keep the claim
Before
It's worth noting that this isn't just a design problem, it's a fundamental issue of how teams communicate across functions.
After
Design can't fix this. The handoff between support and engineering is where it breaks.
The contrast frame was doing no work the plain sentence does not do, and it cost a line and a half.
4. Say the number
Before
Response times improved significantly following the migration.
After
Median response time went from 840ms to 210ms after the migration. The p99 barely moved, which is the part we are still working on.
"Significantly" is the word people use when they do not have the number or do not want to give it. Give the number, including the one that is not flattering. That second sentence is the reason the first is believed.
5. Delete the opening summary
Before
In this section, we'll explore the various factors that contribute to email response rates and examine some strategies for improving them.
After
Delete it and start with the first factor.
Headings already tell the reader what is coming. A paragraph that repeats the heading is a paragraph nobody reads.
6. Let the ending stop
Before
In conclusion, by applying these principles consistently, you can transform your writing and communicate with greater clarity and confidence.
After
Nothing. End on the last real point.
Summaries at the end of short pieces exist because the writer felt the piece needed a shape. It already had one.
The test that works
Read the piece and ask of each paragraph: could someone else have written this exact paragraph?
If yes, it is filler, whoever produced it. If no, because it contains a date, a number, a name, a mistake you made, or an opinion you would defend, it belongs there. Most drafts lose 30% of their length to this test and read better for it.
The second test is reading aloud. A sentence you run out of breath in is too long. Three sentences that sound the same in a row need one of them broken.
A note on being accused
This cuts the other way too. People are being told their writing "sounds like AI" on the basis of a detector score or a vibe, and the writers most often accused are the ones writing carefully in a second language. Short, regular sentences and common vocabulary are exactly what classifiers read as machine-written. AI text watermarking and detection covers what those tools actually measure and what a score does and does not prove.
If you write in English as a second language, the answer is not to add irregularity on purpose. It is specificity: the details of your own work are not something a model or a classifier can imitate. Writing work emails in English as a second language covers the rest.
Editing a paragraph without flattening it
Handing a whole document to an AI tool and taking back what it returns is how writing ends up sounding generic. The tool rewrites your structure along with your sentences, and the structure was the part that was yours.
BlinkWrite works one paragraph at a time, in whatever app you are already writing in:
- Put your cursor in the paragraph you want to change.
- Double-tap Tab to open rephrasing.
- Choose Shorten to cut length, Improve for clarity, Formal or Friendly for register, or Custom for an instruction like "keep the numbers, make this two sentences".
- Review it and press Tab to apply, or Esc to keep what you wrote.
Because it changes the current paragraph only, the shape of the piece stays yours, including the uneven paragraph lengths that make it read as written by a person. The rephrasing guide covers each mode.
Common questions
Should I stop using em dashes?
In work writing, yes. Nothing is lost: a comma, a colon, or a full stop replaces every one of them, and you stop paying for a suspicion you cannot argue your way out of. Keep them where the register is your own and the reader knows your voice.
Do AI detectors actually work?
Style-based detectors are unreliable on arbitrary text and misfire most on second-language writers and formulaic technical prose. Watermark detection, run by the company whose model generated the text, is a different and much more accurate technique, and it only works on that company's output.
Does using an AI tool at all make my writing sound generated?
Only if you let it replace your structure and your specifics. Fixing grammar in a sentence you wrote does not change whose sentence it is. Generating a whole page from a one-line prompt does.
Is there a word list I should avoid?
Not really. A list turns into its own tell as soon as everyone applies it, and the words are a symptom of vague writing rather than its cause. Fix the vagueness and the vocabulary follows.