Writing sounds like AI when it has the habits that language models fall into far more often than people do: a few overused words, a few sentence shapes, a way of inflating how much everything matters, and a tidy summary at the end of every section. None of them proves a machine wrote the piece. Together they make a reader stop trusting it, and that’s the cost, whoever wrote it.
The evidence that the habits are real
The clearest study is by Dmitry Kobak and colleagues, published in Science Advances in July 2025. They compared the words used in more than 15 million biomedical abstracts from 2010 to 2024 against what the previous trend predicted. In 2024, 454 words appeared far more often than expected, most of them verbs and adjectives of style rather than new scientific terms. “Delves” appeared about 28 times as often as predicted, “underscores” about 14 times and “showcasing” about 11 times. From the excess alone, the authors estimate at least 13.5% of 2024 abstracts had been through a language model, and up to 40% in some fields. (Science Advances via arXiv, Times Higher Education)
A Stanford team led by Weixin Liang reached a similar figure a different way, estimating across nearly a million papers that up to 17.5% of computer-science text showed signs of model editing by early 2024. (arXiv, Stanford HAI)
The habits readers notice
Wikipedia’s volunteer editors keep the most-used public list, under the title “Signs of AI writing”, built from cleaning up thousands of articles. Reported by TechCrunch and The Decoder, it groups the habits like this.
Inflated significance comes first: every event is a turning point and every fact is said to prove something larger. The writing tells you how important a fact is instead of letting the fact do it.
Then promotional language, the adjectives that would look at home in a brochure and describe nothing a reader could check. Then editorial throat-clearing: a phrase that announces a point is about to be made, and a closing paragraph that restates what you’ve just read.
The editors also list sentence shapes. A sentence that denies one framing before asserting the corrected one. A list of exactly three things, reached for so often it becomes a rhythm the reader can hear. A trailing clause that adds vague analysis to the end of a sentence. Em dashes where a comma, a colon or brackets would do.
And formatting: headings with every word capitalised, heavy bold, bullet points in prose that didn’t need them, and leftovers from the chat window where the assistant hopes its answer was useful.
The editors’ own caveat matters: not all text with these signs is machine-written, and nobody is taking anyone’s em dashes away. Each sign is weak on its own. Density is what gives the game away.
Why the em dash gets singled out
It became the sign people look for. In November 2025 Sam Altman announced that ChatGPT would at last obey a custom instruction not to use em dashes, which Business Insider described as a fix for “a telltale sign that you used ChatGPT”. The change only made ChatGPT follow a custom instruction to skip em dashes; its default behaviour stayed the same. The punctuation itself is fine, and the problem is that enough readers now treat it as a tell that it costs the writer the benefit of the doubt. (Storyboard18)
What people get wrong
The first mistake is trusting a detector. Weixin Liang’s group tested seven popular AI detectors on essays by non-native English speakers and found they wrongly flagged 61% of them as machine-written, while passing essays by US schoolchildren almost every time. (Patterns via arXiv, Times Higher Education) OpenAI withdrew its own classifier in July 2023 for its “low rate of accuracy”: at launch it caught 26% of AI text and wrongly flagged 9% of human text. (OpenAI, Quartz)
The second is trusting your gut. Two separate studies, one from the University of Washington in 2021 and one from Cornell and Stanford in 2023, found people telling AI text from human text at about 50 to 52%, which is a coin toss. Readers in the Cornell study treated contractions and first-person pronouns as signs of a human, when contractions were in fact more common in the machine text. (ACL, PNAS via arXiv)
The third is asking who wrote it. For a business, the useful question is whether the reader keeps reading, and a sentence a person wrote with four of the habits above loses them just as fast.
How BeigeCheck does it
The Beige-ometer reads a draft and scores how authentic it reads, then quotes the exact sentences that pull the score down and suggests a fix for each. It also scores how well the draft fits the voice it was written in, and checks every claim against your company’s approved facts. It tells you what a reader will react to and what to change, without guessing who wrote the piece. Your AI tool makes the changes you agree to, and nothing moves until you do.
It runs inside Claude, ChatGPT and Microsoft 365 Copilot, so the check happens where the writing happens.