Read this sentence: “In today’s fast-paced world, it’s important to note that a robust, holistic approach plays a pivotal role.” Feel that? That queasy sense that the text is technically English but smells faintly of a data centre. You don’t need an AI detector to spot it — after two years of reading this stuff, we’ve all developed an allergy.
We’ve gathered, in one place, the words and phrases that show up suspiciously often in English text from ChatGPT (and every other model). Treat it as a dictionary: use it to edit your own drafts, to win a bet that you can spot “delve” with your eyes closed, or just to laugh in the group chat.
Before we start, one honest caveat — and we mean it. No phrase on this list proves a text was written by AI. People write this way too. Consultants have for decades, marketers since the 2000s, and students since the essay was invented. These are statistical signals: they appear more often in text from language models, in high density and all at once. A single “it’s important to note” means nothing. Fifteen in one article — that’s a signal. But still not a verdict. Why that distinction matters so much, we cover in our companion piece on whether AI detectors actually work — false accusations hurt real people.
Filler and hedging openers
Phrases that carry no information but pad the text and lend it a false air of gravity. The model drops them in reflexively, like a lecturer saying “so, essentially” every third sentence. Simple test: delete the phrase and check whether the sentence lost anything. It almost never does.
| Phrase | Why it grates | Write instead |
|---|---|---|
| “It’s important to note that…” | If it’s important, just say it. Announcing importance is knocking on an open door. | Delete it and write the sentence. It’ll stand on its own. |
| “It’s worth noting that…” | The twin of the above. Models pair both in one paragraph — and that’s the tell. | Cut it. The reader will note it themselves. |
| “In today’s fast-paced world…” | The number-one opener of every AI essay. It fits everything, so it says nothing — the world has been fast-paced since the wheel. | Open with a scene, a number or a question that stings the reader. |
| “At the end of the day…” | A verbal shrug that pretends to deliver a conclusion while delivering a platitude. | State the actual takeaway, in plain words. |
| “When it comes to X…” | A throat-clear the model uses to ease into a topic. It just delays the point. | Start with the point. “For X, do Y.” |
| “Moreover / Furthermore / Additionally” (on a loop) | Fine once. Three per page and the prose reads like a mechanical conveyor belt of clauses. | “Also,” “on top of that” — or nothing. Sentences can stand side by side. |
The words every model loves
Every writer has pet words. The problem is that all models have the same ones — because they all pick the statistically safest word. One “robust” is normal. Five “robusts” a page is a signature.
| Word | Why it grates | Write instead |
|---|---|---|
| “delve” (delve into) | The single most notorious AI verb. Almost nobody says “delve” out loud; models say it constantly. | “look at,” “dig into,” “let’s take apart.” |
| “tapestry” (a rich tapestry of…) | Decorative nothing. A “rich tapestry of experiences” is a way to avoid naming a single experience. | Name the actual things. Three specifics beat one metaphor. |
| “leverage” | Consultant-speak for “use.” It makes a plain verb wear a suit. | “use,” “make the most of.” |
| “foster” (foster collaboration/innovation) | Nobody fosters anything at work; they help it happen. Pure brochure vocabulary. | “help,” “encourage,” or say what you actually did. |
| “robust” | The all-purpose adjective of strength that describes nothing measurable. A “robust framework” is a framework nobody wanted to specify. | Say what makes it strong: “handles 10k requests a second.” |
| “seamless / seamlessly” | The promise every product page makes. Seams are exactly where software breaks, so the word invites doubt. | Show it: “one click, no re-login.” |
| “pivotal” (plays a pivotal role) | Straight from “plays a key role.” If everything is pivotal, nothing is. | “the main reason,” or show the effect: “without it, sales drop.” |
| “cutting-edge / state-of-the-art” | An adjective-orchestra: loud, and you can’t tell what’s playing. Even a bucket is cutting-edge these days. | Say what’s new and what the competition lacks. |
| “unlock the potential of” | Potential is, by definition, what doesn’t exist yet — so the phrase promises to use something that isn’t there. | Write what you’ll actually do and what comes of it. |
| “navigate the complexities of” | A ten-word way to say “handle.” Nobody is navigating; the text is just circling. | “deal with,” “work through,” or name the hard part. |
Skeleton sentences
Whole sentence-forms the model pours any topic into. You’ll know them because they work identically in a piece about mortgages, compost or urban sociology. They’re why ten AI articles read like one.
| Phrase | Why it grates | Write instead |
|---|---|---|
| “X is a testament to Y.” | A ceremonial way to say “X shows Y.” Models reach for it whenever they want to sound moved. | “X shows Y,” and then give the evidence. |
| “In the realm of X…” | “Realm” is a fantasy word doing office work. It’s just “in X.” | “In X,” or delete the opener entirely. |
| “This underscores the importance of…” | Meta-commentary that tells the reader what to feel instead of showing them. | Let the fact make the point. Then draw the one conclusion. |
| “Not only… but also…” | Grammatically fine, but AI runs it in series because it loves symmetry. Three of these and the rhythm turns mechanical. | Sometimes keep it. More often: two separate sentences. |
| “Embark on a journey…” | Nobody is going anywhere. The reader clicked a link; they’re not Frodo. | Say what they’ll do or learn, step by step. |
| “In conclusion, …” | A meta-announcement: you declare the act instead of doing it. Like saying “now I’ll tell a joke” before the joke. | Just conclude. The last paragraph already reads as the last one. |
| “I hope this helps!” | The auto-generated sign-off, polite as a hold-music recording. Hope is not a call to action. | Invite something concrete: a reply, a trial, the next article. |
Punctuation and formatting tics
This isn’t about single words but about presentation. Models have visual habits — and those often give a text away faster than the vocabulary, because you see them before you read a word.
| Habit | Why it grates | Do this instead |
|---|---|---|
| An em dash in every other sentence — like this one | Models love parenthetical em dashes — they fake a bit of flair — but in bulk it becomes a nervous tic. | Turn some into commas, brackets or separate sentences. |
| A bulleted list for everything | Three thoughts become three bullets. An argument that should flow as a paragraph gets sliced like deli meat. | Bullet only what’s truly a list. Argue in prose. |
| Bold on every third word | When everything is bold, nothing is emphasised. The page looks like a first-year’s pre-exam notes. | One bold phrase per section. Or none — a good sentence carries itself. |
| Emoji as bullet points | Rocket, lightbulb, check mark — the starter kit of every AI LinkedIn post. Readers see them before the content. | In serious writing: none. On social: one, if you truly must. |
Before and after: two rewrites
Spotting clichés is half the job; replacing them is the other half. Here’s the method in action — same meaning, human rhythm.
Before: “In today’s fast-paced world, it’s important to note that leveraging cutting-edge tools can foster a seamless and robust workflow that plays a pivotal role in unlocking your team’s potential.”
After: “Good tools save your team time. The right one cut our release process from five steps to two — here’s how we set it up.”
Before: “This underscores the importance of a holistic approach as we navigate the complexities of the modern landscape.”
After: “So the pieces have to work together. Here are the three that most people get wrong.”
Notice what changed: the filler is gone, the sentences vary in length, and each one now carries a specific — a number, a concrete next step. That last part is the bit a model can’t fake for you.
You found these phrases — now what?
First, the thing that matters most, because someone’s reputation can ride on it: this list is not an AI detector. A single phrase means nothing — people wrote “it’s important to note” long before ChatGPT. Even a heavy cluster is only a hint: maybe AI, maybe a human trained by school, a corporate style guide or a marketing team. Before you accuse anyone of handing in machine output, read how often AI detectors get it wrong — and remember that a human with a checklist is at least as error-prone. A cluster of phrases is a reason for a conversation, not a verdict.
If, on the other hand, you found these phrases in your own draft — congratulations, you just did your first editing pass. One tip unique to this dictionary: work the “write instead” column. Don’t rewrite from scratch — go through the text phrase by phrase like a proof reader with a marker, and at each hit swap in a specific.
Some of that work you can hand to a machine: humai does exactly this — it spots the clichés and rewrites the text into natural prose, with your meaning intact. We don’t promise the result will slip past a detector, because no one can honestly promise that. Curious how patterned your draft looks first? Our free AI style detector gives you an estimated score — an estimate, not proof. What we do promise is that the text will read without making anyone grind their teeth.
And finally: if a phrase made you laugh because you see it on your work Slack every day, send this dictionary along. The more people who recognise AI boilerplate, the less of it there is in circulation — and that’s good for everyone, models included. They just don’t know it yet.