- AI writing gives itself away with em-dashes, tricolons, hedging intros, and the "it's not just X, it's Y" construction.
- Run three passes: cut the intro, break the rhythm, add one specific (a number, a handle, a timestamp).
- Read it aloud. If it sounds like a press release, delete it and start again from a sentence you would text a friend.
Most "written by AI" threads fail the same way: they are grammatically perfect and completely forgettable. The model reaches for the safest phrasing every time, and safe reads as generic. Your reader's brain runs a pattern-match in the first three seconds. Here is what it checks, and how to break the match.
The 8 tells that out a bot#
None of these are wrong on their own. Stacked together, they are a fingerprint. If your draft has four or more, it reads as machine-made no matter how good the idea is.
- Em-dash stuffing. One per paragraph, maximum. Models use them like commas because they let a sentence keep hedging without committing to a full stop.
- Tricolons everywhere. "Faster, cheaper, and more reliable." Real people do not list in threes constantly. It is the rhetorical equivalent of stock photography.
- "It's not just X, it's Y." The single most recognisable AI sentence shape on the platform. Delete it on sight.
- Hedging intros. "In today's fast-paced world." "Let's dive in." "Here's the thing." The first sentence exists to warm up the model, not to inform the reader.
- Zero specifics. No numbers, no names, no dates. Everything is "many" and "often" and "teams". Vagueness is what an average over the whole internet looks like.
- Symmetrical rhythm. Every sentence lands between 12 and 18 words. Humans vary wildly: a 40-word ramble, then three words.
- Title-case earnestness. Capitalised Important Concepts that do not need it. "The Power of Consistency."
- The bow at the end. A neat summarising conclusion nobody asked for. Threads should end on the strongest point, not a recap of it.
If it could have been written about any product, by anyone, it will be scrolled past by everyone.
The 3 edit passes#
You do not need to internalise all eight tells. Three passes over the draft catch most of them.
1. Cut the intro#
Delete your first sentence. Then delete the new first sentence. The post almost always gets better, because openers are where hedging lives. Start on the claim, not the runway to it.
2. Break the rhythm#
Find your longest sentence and your shortest. If they are within five words of each other, you have a metronome problem. Split one long sentence in two. Merge two short ones. Read the result aloud and listen for the beat: if you can nod along to it, keep editing.
3. Add one specific#
One number, one handle, one timestamp. "We shipped 412 evals last quarter and 36% failed silently" beats "teams often struggle with evaluation" every single time. Specificity is the cheapest credibility you can buy, and it is the one thing a model cannot invent for you honestly.
Before and after#
Same idea, run through the three passes.
Before (AI slop):
In today's fast-paced world, shipping AI features is more important than ever. It's not just about speed, it's about reliability, scalability, and trust.
After:
Shipped evals to prod on Friday. 36% of our checks were failing silently until we added a sweep step. Here is the exact config we used.
Nothing about the second version is clever. It is just concrete, first-person, and rhythmically uneven. That is what human sounds like.
Why models write like this#
A language model optimises for the most probable next word given everything it has read. The most probable phrasing is, by definition, the most average phrasing. Averageness is not a bug you can prompt away with "write casually"; the model will simply produce average-casual, which is its own recognisable flavour of slop. See 12 AI phrases that instantly out you as a bot for the vocabulary version of the same failure.
The fix is constraint, not vibes. A model given your actual sentence-length distribution, your vocabulary, your stance, and your formatting habits has far less room to regress to the mean. That is the entire idea behind a voice profile: describe the shape of your writing precisely enough that a tool can draft inside it.
Voxly builds that profile from your real timeline and runs these edit passes at draft time. If you would rather run them by hand, the checklist above is the whole trick. Either way, the bar is the same: read it aloud, and if you would not say it to a friend, do not post it to ten thousand strangers.
FAQ
Does removing em-dashes make a post sound human?
Not on its own. Em-dash stuffing is one of a few hundred signals. The bigger tells are rhythm (every sentence the same length) and the total absence of specifics. Fix those two and the punctuation matters much less.
Will these edit passes work if I have never posted much?
Yes. The three passes only need a draft, not a posting history. What a history gives you is a reference for your own voice, so you can tell the difference between generic-good and you-good.
Can I use these passes without Voxly?
Absolutely, that is the point. The three passes are free and manual. Voxly just runs the same kind of checks at draft time, against a voice profile built from your own timeline, so you don't have to.