You wrote a draft with ChatGPT, read it back, and something was off. So you rewrote it. Still off. You added "write in a natural, conversational tone" to the prompt and got the same thing with contractions. The words are correct. The grammar is spotless. It reads like a press release for a product that doesn't exist.

You're not imagining it, and it isn't a tone problem. It's a mathematical one.

Why AI writing sounds like AI

AI writing sounds like AI because a language model predicts the most probable next word, and averaged across billions of documents the most probable word is the least surprising one. That produces prose sitting permanently at the centre of the distribution: fluent, balanced, hedged, evenly paced, and completely without the friction that signals a specific person wrote this for a specific reason. Human writing gets its texture from choices a model has no reason to make — an odd word, a sentence that stops short, a detail nobody else would have included. "Sounds like AI" is what the statistical middle sounds like. Prompting your way out means deliberately pulling the model off centre.

The centre of the distribution

Think about what "most likely next word" means at scale. Every time the model picks, it's asking what usually comes next. Not what's true, not what's interesting — what's typical.

Do that a few thousand times in a row and you get writing that is the average of all writing. Which is exactly what "generic" means. Nobody writes like this, because the average of a million voices isn't a voice.

This is why "be more creative" and "write with personality" barely move the needle. They're instructions to be unusual, delivered without any information about which unusual. The model does the only thing it can — it reaches for the most typical version of unusual it knows, which is why you get whimsical metaphors about journeys and tapestries. Still the middle. Just a different part of it.

The fix is never an adjective. It's supplying the specific information the average is missing.

The seven tells

Readers clock AI writing in about a second, usually before they could tell you why. These are the things they're registering.

1. Uniform sentence length. Human writing lurches. A long, winding sentence that doubles back on itself, then four words. Models produce 15–20 word sentences almost exclusively. This is the single strongest tell and the easiest to fix.

2. "Not just X, but Y." Along with "It's not about X — it's about Y." The construction is fine once. Models reach for it constantly because it's a cheap way to sound insightful without adding information.

3. Everything comes in threes. Three adjectives, three bullets, three examples. Real arguments have one point, or five, or two and a caveat.

4. Hedging as filler. "It's important to note," "it's worth considering," "while there are many factors." These say nothing. They exist because balanced text is more probable than committed text.

5. Signposting out loud. "In this section, we'll explore…" "Let's dive into…" Confident writing just starts.

6. Abstract nouns doing concrete work. "Streamline your workflow and enhance productivity" versus "stop reopening the same three tabs to send one invoice." The model reaches for the abstraction because it fits every context — which is precisely why it lands in none.

7. The conclusion that restates the intro. Adds nothing, exists because articles usually have one.

Notice that most of these aren't errors. They're the safest available option, chosen repeatedly. That's the whole phenomenon.

What it's costing you

The obvious cost is that it's boring. The less obvious one is that readers now actively penalise it.

Adobe's 2025 survey of more than 16,000 creators found 86% already use generative AI, so the novelty defence is gone — your audience has read a great deal of this. University of Florida research tracking the 2026 backlash found consumer preference for AI-assisted creator content fell sharply from its 2023 level, and reported brand-trust damage roughly doubled year over year.

Worth reading that carefully, because the useful nuance gets lost: audiences aren't punishing AI involvement. They're punishing output that feels mass-produced. Same distinction platforms are drawing — YouTube's monetisation policy doesn't penalise AI use, it penalises "generic or unoriginal templates giving the impression of mass production." Nobody minds the tool. They mind that you didn't show up.

The prompts

Each of these attacks a specific tell. Run them on a draft you already have — they work better as edits than as generation instructions.

1. The rhythm fix

Rewrite the text below with deliberately varied sentence length. At least
one sentence under 5 words. At least one over 35. Never two consecutive
sentences within 5 words of each other in length. Don't change the meaning
or add new claims — only restructure.

[paste your draft]

Why it's built this way: it's a mechanical constraint, not a taste instruction, so the model can actually comply. Rhythm is the strongest tell and this is the highest-leverage single edit you can make. Swap in: nothing. Run it on everything.

2. The specificity pass

Below is my draft. Find every abstract claim and mark it. For each one,
ask me a question that would let me replace it with a concrete, specific
detail. Don't invent the details — ask me for them.

[paste your draft]

Why it's built this way: the model can't invent your specifics, and when it tries, it hallucinates. Making it interrogate you turns it into an editor rather than a fabricator. Swap in: nothing — but actually answer the questions. This one takes ten minutes and does more than any other.

3. The banned-construction filter

Rewrite this, removing every instance of: "not just X but Y", "it's not
about X, it's about Y", "in today's world", "it's important to note",
"let's dive in", "delve", "unlock", "leverage", "elevate", "seamless",
"robust", "landscape", "tapestry", "game-changer", and any sentence that
only restates the previous one. Keep my meaning exactly.

[paste your draft]

Why it's built this way: naming the exact strings works where "avoid clichés" doesn't, because the model has no reliable internal sense of which of its habits read as tells. Swap in: add whatever's overused in your niche. Keep this as a saved snippet.

4. The voice-match prompt

Here are three things I've written: [paste 3 samples, 200+ words each].

First, describe the patterns you see in how I write — sentence length,
punctuation habits, how I open, how I handle transitions, words I avoid,
how formal I am. Be specific and concrete. Don't flatter me.

Wait for me to correct your description before writing anything.

Why it's built this way: style transfer works from examples, never from adjectives. "Friendly and professional" describes ten million writers. Three real samples describe one. The forced pause matters — the model's first read is usually part wrong, and correcting it is where the real briefing happens. Swap in: your actual samples, and pick ones you're proud of rather than ones that were convenient.

5. The read-aloud test, automated

Read this out loud in your head, sentence by sentence. Flag every sentence
a real person would not say out loud in conversation. Don't rewrite —
just list them and say what's wrong with each.

[paste your draft]

Why it's built this way: diagnosis separated from rewriting stops the model laundering the problem into a different phrasing of the same problem. You keep control of the fix. Swap in: nothing. This is the last pass before publishing.

Which tell, which fix

The tell Why the model does it The fix
Uniform sentence length Mid-length sentences are the most probable Prompt #1 — mechanical length constraints
"Not just X, but Y" Cheap insight-shaped construction Prompt #3 — name the exact string
Everything in threes Tricolons are over-represented in training text Ask for 2, 4 or 5 explicitly
Hedging filler Balanced text is more probable than committed text Prompt #3, then cut every "it's worth noting"
Signposting out loud Articles usually announce structure Delete the first sentence of each section
Abstract nouns Abstractions fit every context Prompt #2 — the specificity pass
Restating conclusion Articles usually have one Delete it, or end on the sharpest concrete point

The part no prompt fixes

Every technique here removes something. None of them add the thing that actually makes writing worth reading: a real opinion, an actual experience, a number you got from somewhere, a claim you'd defend if challenged.

A model can't supply that, because it doesn't know anything about your situation, and it has no stake in being right. Strip the tells out of an empty draft and you get cleaner empty writing. Faster, smoother, still nothing.

So the honest workflow isn't "generate, then de-AI it." It's: decide what you actually think, get the specifics down in whatever mess they arrive in, and let the model help with structure, rhythm and compression. That way round, the tells mostly don't appear — because the writing has something in it.

Related guides

Start with the rhythm

If you only take one thing: vary your sentence length hard, then replace one abstraction with a detail only you would know. Those two edits do more than any rewrite prompt, because rhythm and specificity are exactly what the statistical middle strips out.

→ ChatGPT Prompts for Writing — the full editing set, ready to copy.