Why every AI draft sounds the same
Who this is for
Anyone who bought an AI writing tool, got drafts that are technically fine and completely generic, and cannot work out what to change.
What you will take away
- The habits readers notice are shapes, not words. We counted seven of them across 79 of our own pieces and 64,995 words.
- The em dash, the most banned marker of AI writing on the internet, showed up in 4% of our corpus. A habit almost nobody names showed up in 44%.
- Banning a shape tends to move it rather than remove it. We have the measurement where that happened to us.
The draft is accurate. The structure follows the brief. Nothing in it is wrong, and you would still never send it to a customer with your name on the bottom. That gap between correct and usable is the most common complaint about AI writing, and it is almost never a vocabulary problem.
We went looking for what it actually is, in the only corpus we had permission to take apart: our own. 79 pieces over 300 words, 64,995 words, generated between April and August 2026. Seven detectors, each one a plain regular expression, each validated against 18 hand-labelled examples pulled from real output before any number below was trusted. One detector failed its own test case on the first attempt and was rebuilt.
What the count found
The em dash is the internet's favourite evidence of machine writing. In our corpus it appeared in three pieces out of 79, at a rate of 0.02 per thousand words. The corrective fragment, a short sentence that opens with the word “Not” and exists to negate the sentence before it, appeared in 35 pieces at 0.89 per thousand words. That is a forty-fold difference in prevalence between the habit everyone polices and the habit almost nobody has a name for.
The four worth knowing by sight
The top four account for most of what a reader registers as sameness. Each one is mechanical enough that you can search for it.
The corrective fragment. A sentence of eight words or fewer beginning with “Not”, placed immediately after the thing it corrects. “We measured arrival. Not interest.” One of these is a choice a writer makes. Eleven of them in the same piece is a fingerprint, and that is roughly what a long draft produces when nothing intervenes.
The antithesis. Any variant of “not just X, but Y”, including “it's not about X, it's about Y” and “this isn't a Z, it's a W”. The shape promises an upgrade on every claim it touches. Two of them in a paragraph and the prose starts to feel like it is arguing with an opponent who never spoke.
The announced count. The word “three” followed by a plural noun, promising a set before delivering it. “Three things decide whether this works.” Nothing is wrong with a list. What gives it away is the announcement, which exists to organise the writer rather than to help the reader.
The withheld payoff. In our corpus this was almost entirely one construction: 21 of the 22 hits were the phrase “here's why”, “here's what”, “here's where” or “here's how”. The sentence announces that an answer is coming instead of giving it. Our detector for this one is deliberately narrow and certainly undercounts, so 23% is a floor rather than an estimate.
Why “write more human” does not fix it
The obvious move is to put the ban in the prompt. We did that, and then we ran the control that almost nobody runs: generate the same briefs with the rules switched off, and compare.
One result from that work is worth more than the rest. Our ruleset told the model to avoid plain-text section headers. The model complied, and then wrote “Number one is sourced pipeline” inline, in the body, as a sentence. The form moved. The behaviour stayed exactly where it was.
We ablated six separate clauses looking for the one responsible for that habit. All six came back negative. The variable turned out to be the brief: a title that named a count produced enumerated output, and briefs that did not name a count produced none of it across eighteen generations. The rule we suspected was innocent, and the instruction we had written to suppress a shape had taught the model a second way to make it.
A prohibition tells a model which shape to avoid. It does not tell it which shape to use instead, so the model finds a neighbouring one. That is the mechanism behind every disappointing “make it sound more human” instruction, and it is why the fix belongs after generation rather than inside the prompt.
What to do with the draft in front of you
Stop rereading it. Rereading is how you lose an afternoon and end up rewriting paragraphs that were fine. Search it instead, for four strings:
- Sentences starting “Not”. Keep at most one per piece. Fold the rest back into the sentence they were correcting.
- “not just”, “it's not”, “isn't about”. Delete the first half. The second half was the claim; the first half was scaffolding.
- “three” followed by a plural noun. Keep the list, cut the announcement, let the reader discover there are three.
- “here's why”, “here's what”, “here's how”. Delete the phrase and start with the answer.
Four searches, a few minutes, and the draft stops announcing itself. The important part is that each fix replaces a shape with a different shape rather than swapping words inside the same one, which is the reason a thesaurus pass never works on this problem.
If you want the other three, they are quicker: a one-word verdict on its own line (“Fine.”), the same two words repeated three times inside one comma-heavy sentence, and any sentence that narrates the article's own structure.
The uncomfortable part
Every number here came from our own output, generated by our own system, under rules we wrote ourselves to prevent exactly these habits. We published it because a corpus you have not measured is a corpus you are guessing about, and because seven of the conclusions we reached during that week were wrong until a control arm proved otherwise.
One of those corrections is worth repeating for anyone about to run a similar check. Corpus data cannot tell you what a rule does, because everything in the corpus was generated under that rule. We spent a day confidently reading effects out of prevalence figures before adding the arm that showed us otherwise.
Common questions
Why does AI-generated content sound generic even when the facts are right?
Because the repetition readers notice happens in sentence shape rather than in word choice. Across 79 of our own pieces and 64,995 words, one habit appeared in 44% of everything we published: a short sentence beginning with the word Not, used to correct the sentence before it. Facts can be perfect while every paragraph turns the same corner.
What are the most common AI writing patterns?
In our corpus: the corrective fragment at 44% of pieces, then the antithesis, the announced count and the withheld payoff at 23% each, then the anaphoric triple at 11%, structural narration at 8% and the one-word verdict at 5%. The em dash, the most commonly banned marker of AI writing, appeared in 4%.
Does telling the model to write more human help?
Not reliably, and it can move a habit rather than remove it. When we banned plain-text section headers, the model stopped writing Number one: as a heading and started writing Number one is inline instead. The form changed and the behaviour did not. A prohibition tells a model what shape to avoid, not what shape to use.
How do I fix a generic AI draft?
Search the draft for the specific shapes rather than reading it again. Sentences starting with Not, the phrase not just, the word three followed by a plural noun, and the phrase here is why. Each takes seconds to find and about a minute to rewrite, and the rewrite is a different sentence shape rather than different words in the same shape.
Is a banned-words list useful for AI content?
Our 46-item list produced four hits across 67 pieces, and a control with the rules switched off found zero banned terms in the same conditions. The list was policing a base rate the model was not reaching. Vocabulary rules operate a level below the thing readers actually register.
More on that last one in why a banned-words list won't fix AI-sounding drafts.