How to make AI writing sound human - we made it into a Claude skill
Models give themselves away with a handful of sentence shapes, and a word swap leaves every one in place.
Short answer: AI writing sounds like AI because of sentence structure, not word choice. Models lean on a few shapes: a sentence capped with an "-ing" phrase, the "it's not X, it's Y" contrast, the three-part list, and a flat sentence rhythm. Swapping words like "delve" does not touch them. To fix it, edit the structure.
We learned this editing AI drafts for our own clients, until the same shapes turned up in every one.
Why editing the vocabulary does not work
Most people de-AI a draft by hunting words. They delete "delve," swap "robust" for "solid," and strike the em dashes. The draft still reads as machine-written, because the words were never the main tell. The structure was, and structure survives a find-and-replace.
What the research says
Three findings settle it.
The vocabulary shift is real but easy to launder. Dmitry Kobak and colleagues measured "excess words" across more than 14 million PubMed abstracts and found a sharp post-2022 jump in terms like "delve" and "underscore," published in Science Advances. Useful as a signal. Also the first thing a careful editor removes.
The durable tells are syntactic. A Carnegie Mellon team, writing up their PNAS study, found that language models use present-participial clauses (the "..., highlighting the importance of X" shape) at two to five times the human rate, and the gap was largest for instruction-tuned models like ChatGPT. No single word triggers it, so a thesaurus does nothing.
The rhythm gives it away even when the words do not. AI prose sits in a narrow band of sentence lengths, what detection tools call low burstiness. People write a four-word sentence next to a thirty-word one. Models rarely do.
The linguist Colin Gorrie named the cause well: the model is not short on skill, it is short on taste. It runs Churchill's rhetorical machinery, balanced contrasts and three-part lists, on a customer-support email. The device is fine. The saturation is the tell.
Which sentence structures give AI away?
Four shapes account for most of the "a machine wrote this" reaction. For a fuller catalogue, the Wikipedia editors' guide to signs of AI writing is the best public reference.
| The shape | What it looks like | The fix |
|---|---|---|
| The participle tail | "The team shipped in March, marking a turning point." | Cut the tail, or make it a real second sentence. |
| The explicit contrast | "It's not X, it's Y." | State the positive; delete the negated half. |
| The padded triad | Three items where two carry the weight. | Cut to a pair. |
| The flat rhythm | Sentence after sentence at the same length. | Make one short, make one long. |
The "it's not X, it's Y" pattern is now common enough that linguists study it directly. The Conversation argues the negation form actively hurts comprehension, because the reader has to hold the wrong idea in mind before the right one.
Want the method itself? We packaged it as a free Claude skill. Add your email and we'll send you the Notion doc with the full fix table and checklist. Get the skill
How do you fix it?
Fix structure before words, in four moves.
- Decide: rewrite or patch. If a passage trips several of the shapes above and the sentences all run a similar length, rewrite it from the idea. Patching an AI-shaped sentence word by word usually leaves it AI-shaped.
- Run the shapes. Cut the participle tails, kill the explicit contrasts, break the padded triads.
- Break the rhythm. In each paragraph, put at least one sentence under eight words and one over twenty-five.
- Apply two tests a search cannot do. Swap two body paragraphs; if nothing breaks, they were filler. Read it aloud; if it ticks like a clock, vary it. Bloomberry's research on AI sentence cadences catalogues the repeating rhythms in detail.
One guardrail matters more than it sounds: do not over-correct. Every one of these patterns appears in good human writing too. The tell is density, not a single instance. Strip every contrast and every triad and you do not get human prose, only lifeless prose.
Frequently asked questions
Is the em dash a sign of AI writing? On its own, no. Plenty of strong writers use it, as Gorrie argues in a defense of the em dash. It became a tell because models reach for it heavily, and because it is the last formatting habit they keep when told to write plainly. Treat a cluster of them as a flag, not a verdict.
Do AI detectors actually work? Not reliably. They score the patterns above, which also appear in human writing, so they flag clear, structured prose as machine-made. Use them as a hint, then edit by hand.
Does editing the vocabulary fix AI writing? No. Word swaps are necessary but not sufficient. A draft can pass every banned-word check and still read as AI, because the sentence shapes and the rhythm are unchanged.
What is the fastest way to make AI text sound human? Two moves catch most of it. Search for a comma followed by an "-ing" word and fix every participle tail, then scan sentence lengths and break any run of three similar ones.
Get the skill, free
The full method is a Claude skill you can run on any draft: a sentence-by-sentence fix table, the rhythm rule, and a 60-second check. Add your email and we will send the Notion doc with everything in it.
Sources
- Dmitry Kobak, Rita González-Márquez, et al., "Delving into LLM-assisted writing in biomedical publications through excess vocabulary," Science Advances (2025). science.org
- Carnegie Mellon University, "Is it human, or is it AI?" on participial clauses and LLM style, published in PNAS (2025). cmu.edu
- GPTZero, "How do I interpret burstiness or perplexity?" gptzero.me
- Colin Gorrie, "Why ChatGPT writes like that," The Dead Language Society. deadlanguagesociety.com
- Wikipedia, "Signs of AI writing." en.wikipedia.org
- The Conversation, "Slanguage: why AI's stylistic negation 'it's not X, it's Y' is both annoying and doesn't work" (2026). theconversation.com
- Bloomberry Research, "AI Sentence Patterns: 12 Common Cadences in AI-Generated Writing." bloomberry.ai
By Christopher Kliebenstein. We build and edit AI workflows, and test these methods on real client writing.