This is the catalog version of a longer post we did called How To Humanize AI Writing in 2026. That one is the playbook. This one is the cheat sheet. If you only have ten minutes and you want to scrub your draft of the obvious AI flavor, this is what you scan for.
Words and phrases
The "delve" cluster
This is the most famous one at this point and it's still everywhere. Words the model reaches for when it wants to sound thoughtful. None of them are bad words. They're just statistically over-represented in AI text compared to how humans actually write.
- delve, delve into
- navigate, navigating (the metaphorical use, as in "navigating the challenges of")
- tapestry, rich tapestry
- ever-evolving, ever-changing
- landscape (as in "the AI landscape," "the cybersecurity landscape")
- realm
- journey (especially "embark on a journey")
- showcase (as a verb)
- foster, fostering
- multifaceted
- nuanced
- seamless, seamlessly
- robust
- leverage, leveraging
- utilize (just say "use")
- furthermore, moreover
- in conclusion
The hedge cluster
Models hedge constantly. They were trained to. Cut these out and your writing sounds twice as confident.
- "It's important to note that"
- "It's worth mentioning"
- "While this approach has many benefits"
- "That said," (overused)
- "However, it's worth considering"
- "It's crucial to remember"
- "In today's fast-paced world"
- "At the end of the day"
The fake balance
Constructions that pretend to weigh both sides but are really just filler. These are the deadliest AI tells because they sound like real prose to a casual reader and they read like cardboard to anyone paying attention.
- "It's not just X, it's Y"
- "More than just X"
- "X is more than X"
- "On the one hand... on the other hand..." (used in short pieces where the actual on-the-other-hand never lands)
- "While X may seem Y, in reality it's Z"
Punctuation patterns
Em dash overuse
The em dash is the punctuation mark most associated with AI writing in 2026 and the association is partially deserved. Plenty of human writers use em dashes (the writer of this post used to use them constantly). The model uses three per paragraph. That's the tell. Volume.
Practical fix. Open your draft. Count the em dashes. If you have more than one every two paragraphs, you have too many. Replace most of them with commas, periods, or just delete the clause they were separating. The clause is usually a hedge that doesn't need to be there.
The Oxford triplet
"Designed, built, and tested." "Fast, reliable, and affordable." "Discover, explore, transform." The model loves these. So much so that you'll find them in almost every AI-written paragraph that lists anything. Real writers use this rhythm sometimes. Models use it constantly.
Practical fix. Cut every other triplet down to a pair, or expand it into a sentence. "Fast, reliable, and affordable" becomes "fast and reliable, and the price hasn't moved in three years."
The bullet reflex
Models convert prose to bullets whenever they're explaining something. Bullets are fine in documentation. They are a tell in personal writing, essays, marketing copy, and anywhere with a voice.
Practical fix. If a passage could be bullets or prose, choose prose at least half the time. The connective tissue between ideas (because, however, since, after) is part of what makes writing feel like writing.
Sentence shapes
Uniform paragraph length
If every paragraph in your draft is four to six sentences long, you're inside the model's default. Real writing has paragraphs of one sentence, two sentences, eight sentences. The variance is what makes it feel like a person.
The sandwich opening
Models often open sections with the same structure. A short setup sentence, a meatier middle, a wrap-up that restates. By paragraph three you can guess the shape of paragraph four. Break the pattern by starting paragraphs with fragments, questions, or single-word lines.
The numbered narrative
"There are three things to know." "I'm going to walk you through five steps." "Five lessons we learned." When every section announces its own count, you're reading a template. Real writers sometimes use numbered structure, but they don't open every section with "here are the X things."
Specific structural tells
The intro that summarizes the post you're about to read
"In this article, we'll explore X, look at Y, and conclude with Z." Nobody writing for actual humans does this. It's a content-marketing artifact that the models picked up from training data. Cut these intros entirely and start with the first real sentence of the piece.
The "conclusion" header
If your final section is labeled "Conclusion" and starts with "In conclusion," you have an essay-template tell. Either cut the header and the opener, or rename the section something specific. "What we learned." "Where this goes next." Anything that's about the content rather than the form.
The CTA at the end
"If you found this helpful, please share it." "Feel free to reach out with any questions." These are CTA fossils from blog templates the model trained on. They feel automatic because they are. Either say something specific (a real link to a real resource) or skip it.
What to do with this list
You don't need to memorize it. Just do one pass through your draft with the list open. Mark every instance you find. Then decide which ones to cut and which to keep. Some of these patterns are fine in moderation. The volume is what makes a draft read as AI.
If you want a tool that does this scan automatically, we built Cloak for exactly this. It looks at the same pattern list (and a lot more) and gives you a score from 0 to 100 for how AI-flavored your draft reads, then rewrites it with more natural rhythm. Use it as a first pass. Do the final edit yourself.
For the longer playbook on what to actually do once you've found these tells, see How To Humanize AI Writing in 2026. And for why detectors care about these patterns at all, see How AI Detectors Actually Work.