Why some words give AI away
Language models favor certain words far more than people do, and readers have learned to notice. The clearest evidence comes from science. A study of more than 15 million PubMed abstracts from 2010 to 2024 found that a handful of style words, among them “delves”, “showcasing” and “underscores”, became sharply more common after chatbots arrived. From that excess vocabulary, the authors estimate that at least 13.5% of 2024 abstracts were processed with a language model (Kobak and colleagues, Science Advances, 2025).
None of these words is wrong on its own. The problem is density: when several turn up in one email, the reader stops reading for meaning and starts wondering who wrote it.
The words, and what to say instead
Most of these words stand in for a plainer one, or for a specific fact the writer did not supply. The fix is usually the plain word, or the fact.
| Instead of | Try |
|---|---|
| delve into, dive into | look at, cover, go through |
| leverage, utilize, harness | use |
| seamless, seamlessly | say what works without effort, or cut it |
| robust, comprehensive | say what it covers or survives |
| pivotal, crucial, vital | key, or explain why it matters |
| elevate, enhance, empower | improve, raise, let, or the actual change |
| navigate the complexities of | deal with, handle |
| underscore, showcase | show |
| foster, cultivate | build, encourage |
| a testament to | shows |
| tapestry, realm, landscape | name the actual thing: the market, the field, the options |
| game-changer, revolutionize, unlock | the specific result, with a number if you have one |
| in today’s fast-paced world | cut it; start with your point |
| it’s important to note that | cut it; say the thing |
The habits
Words are the easy part. These habits give a draft away even when every word is ordinary:
- The long dash, or em dash, used two or three times a paragraph, where a person would use a comma, a colon or a full stop.
- A contrast set up only to be knocked down: “This isn’t just a tool, it’s a partner.”
- Lists of three: three adjectives, three benefits, three short sentences in a row, whether or not there are really three things.
- Filler at either end: “Certainly!” or “Great question!” at the start, “I hope this helps” or an offer of more help at the end.
- Praise for the reader’s idea or question before the answer.
- A bold label on every bullet, and headings on an answer short enough to need none.
- Emoji in work writing, often used as bullet points.
- Every sentence about the same length, which reads flat even when each sentence is fine.
- A closing summary that repeats the answer, beginning “In conclusion” or “Overall”.
Signs, not proof
People wrote “delve” and used dashes long before chatbots, and plenty still do. Wikipedia’s editors, who keep a long field guide called Signs of AI writing, say the same: the patterns are reasons to look closer, not proof. Automated AI detectors are unreliable for the same reason. Use the list to edit your own drafts, not to accuse a colleague or a student.
A prompt that stops them
Fixing every draft by hand is slow. Put a standing instruction in your custom instructions, a Claude Project or a GPT, so the model avoids the habits from the start:
Write in plain, specific language, as a person would in a work email. - Do not use these words: delve, leverage, utilize, seamless, robust, pivotal, crucial, elevate, foster, underscore, showcase, tapestry, realm, landscape, testament, game-changer, unlock. - Do not use long dashes. Use commas, colons or full stops. - Do not open with praise or filler, and do not end with a summary or an offer of more help. - No "not X but Y" contrasts. Use a list of three only when there are exactly three things. - Vary sentence length. No emoji. Bold nothing. - Where you would write a vague claim, write the specific fact, or ask me for it.
The last line matters most. Most AI filler fills a gap where a fact should be, so the cure is to supply the facts. The prompt to make a draft sound human applies all of this to a draft you already have.
Key takeaways
- A few words, such as delve, seamless and pivotal, appear far more in AI writing than in human writing, and readers notice.
- Swap them for the plain word, or better, for the specific fact.
- The habits give writing away more than the words: long dashes, “not X but Y”, lists of three, filler and flattery.
- These are signs, not proof. Use them to edit, not to accuse.
- A standing instruction stops the habits at the source, and supplying facts removes most of the filler.