Words that sound like AI, and what gives a draft away
Lists of AI words to avoid don't tell you much. What gives a draft away is several habits piled into one paragraph. Our numbers, and what to cut.
No single word makes writing sound like AI. We've tested that, and the numbers are below. What does it is a handful of habits crammed into one short paragraph: an opening line about the fast-paced modern world, a long dash, a lesson stuck on at the end. A fancy word in a paragraph without those habits mostly goes unnoticed.
Search for this and you'll find long lists of words that are supposed to give away a chatbot. The same few fancy words sit near the top of most of them, and some of the lists run to a few hundred words. People use them like a blacklist, where one hit means the whole post came from a machine. That's the part we disagree with.
Is there a list of AI words to avoid?
Sort of. Some words do show up more in chatbot drafts. When we built Desloppy we tried counting those single words as evidence, and it went badly on writing by people. We ran it on 3,000 texts that people wrote, and it showed stock phrases on 24.7% of them. People use those words in work emails all the time, and novelists used them long before any model existed.
So we stopped. Now a lone stock word shows up as an optional note, something you can swap if you want to, and it doesn't count toward the Slop score at all. A rule only counts when three or more different habits land in the same stretch of about 300 words. On the same 3,000 texts, that brought it down to 1.8%.
If you're reading someone's post and one of those words jumps out at you, please don't decide anything from that.
What does give it away
It's the pile-up. Readers notice when a paragraph does four of the usual things at once, even if they couldn't tell you which four.
The one we'd cut first is the opening line that sets a scene about the world today. Nobody starts a real email that way, and the email is almost always better without it. On LinkedIn the worst offender is the announcement voice, where the writer is thrilled to share something before saying what it is. Then there's the line that tells you something is not about one thing but about another, when nobody said it was. One of those in a post is fine. By the third you're reading a slogan. And a lot of paragraphs end on a neat moral that would read better if the paragraph just stopped a sentence earlier, on the detail.
What we found in 100 LinkedIn posts
We asked five AI models (ChatGPT, Claude, Gemini, Qwen and GPT-OSS) for 100 LinkedIn posts in total, then counted. 86 had hashtags, 48 had the long dash and 43 had the rocket emoji. 70 of the 100 had three or more of these habits at once, and the average post had 3.8. The models weren't all alike, which surprised us a little: the ChatGPT posts averaged 1.8 habits and the GPT-OSS posts averaged 5.8. The full study is on our site.
Any one of these is fine on its own. It's when four of them land inside 150 words that the text starts to feel like a template, and at that point it more or less is one.
What to do about it
We don't think find-and-replace on a word list does much. If you swap one word for a simpler one and the paragraph still opens with a scene-setter and closes with a moral, it reads about the same as before.
What we'd do is go through a paragraph at a time and count how many of these things it's doing. If it's three or more, cut until it's one or none, and most of the time that means deleting the opening line and the last line rather than changing any words. Where there's a vague claim, a real detail goes in, so if sign-ups went from 190 to 312, the draft should say 190 to 312 and not a significant increase.
Finding those lines in your own draft is the hard part, especially after an hour of looking at it, and it's the step we built Desloppy for. You paste a draft, or upload a Word, PowerPoint or PDF file, and it marks the lines that sound like a chatbot, gives a short reason for each, and suggests a fix. A lone stock word only gets an optional note, while a paragraph where the habits pile up gets flagged. Your text is never stored.
Desloppy doesn't rewrite anything for you, and that's on purpose. You make the edits yourself, or you give only the flagged lines to your own assistant along with the real details.
The Slop score is a probability and not proof. It describes how the text sounds and never who wrote it, so we'd ask you to use it on your own drafts and not to accuse someone else.
You can try 1,000 words free, with no account.