What words does ChatGPT use a lot?
Updated · 9 min read · by the HidenGPT team
ChatGPT leans on words most people never say out loud: “delve”, “tapestry”, “a testament to”, “crucial”, “underscore”, “showcasing”, plus stock openers such as “in today's fast-paced world” and “I hope this email finds you well”. A single one means little. Five in one paragraph is what makes a text read as generated.
The list also moves. “Delve” peaked in 2023 and early 2024, and the newest models favor different words, so treat any list, this one included, as a snapshot. The fix stays the same: swap the word for the one you would say, or for a fact.
| Word or phrase | Why it reads as generated | Write instead | Signal |
|---|---|---|---|
| “delve”, “delve into” | Few people say it aloud. One study found “delves” in 2024 medical abstracts about 28 times as often as the earlier trend predicted | “look at”, “cover”, or open with the point itself | High |
| “in today's fast-paced world”, “in today's digital age” | An opening that fits any essay on any subject | Your first real claim, with a date or number in it | High |
| “a rich tapestry of”, “a symphony of”, “a dance of” | A metaphor that could describe anything, so it describes nothing | Name the parts: “Arab, Indian and Italian dishes” | High |
| “a testament to” | Praise with the evidence left out | The evidence: “it has opened every day since 1974” | High |
| “unlock the potential”, “reach your full potential” | A promise with no object and no measure | The actual change: “cuts the weekly report from two hours to forty minutes” | High, Medium |
| “embark on a journey”, “a journey, not a destination” | Turns an ordinary start into an adventure | “start”, “began”, or what you did on day one | High |
| “let that sink in” | Drama borrowed from social posts | Cut it and let the fact stand on its own | High |
| “as an AI language model” | A chatbot disclaimer pasted in by mistake | Delete the sentence | High |
| “it's important to note”, “it's worth noting” | Announces a point instead of making it | Delete the phrase and keep what follows | Medium |
| “plays a crucial role”, “plays a pivotal role” | Importance stated, never shown | What it does: “sets the price”, “signs off the budget” | Medium |
| “deep dive”, “let's dive in” | Promises depth, then starts a list | “Here is how it works”, or simply begin | Medium |
| “in a world where”, “picture this”, “imagine a world” | A run-up before the point | The concrete scene you mean, in one line | Medium |
| “have you ever wondered” | A question the reader did not ask | The answer, stated | Medium |
| “navigate the complexities of”, “in the realm of” | Travel and kingdom words for ordinary tasks | “handle”, “in chemistry”, or the name of the task | Medium |
| “the digital landscape”, “ever-evolving” | Scenery words standing in for specifics | “online shops”, “changes every quarter” | Medium |
| “game-changer”, “revolutionize”, “transformative” | Hype that readers discount on sight | The measured difference | Medium |
| “cutting-edge”, “state-of-the-art” | A label with no fact attached | What is new about it, and since when | Medium |
| “harness the power of”, “unleash”, “elevate” | Verbs from advertising | “use”, “release”, “improve” | Medium |
| “a beacon of”, “serves as a reminder”, “a powerful reminder” | A moral pinned onto a fact | State the lesson once, or trust the reader | Medium |
| “at its core”, “at the heart of” | Signals depth without adding any | Say the central thing directly | Medium |
| “I hope this email finds you well” | The default email opener | The reason you are writing, in the first line | Medium |
| “thrilled to announce”, “excited to share” | The default launch-post opener | The news itself: “We open on 3 March” | Medium |
| “look no further”, “say goodbye to”, “take it to the next level” | Ad copy everyone has already read | The specific benefit, in numbers if you have them | Medium |
| “nestled” | A brochure verb for a location | “sits”, “is”, or the street name | Medium |
| “more important than ever” | Urgency with no date or cause | What changed, and when | Medium |
| “paradigm shift”, “synergy” | Consultant jargon | Say what changed, and for whom | Medium |
| “seamless”, “robust”, “leverage” | Office filler | “smooth”, “sturdy”, “use” | Low (counts when repeated) |
| “streamline”, “empower”, “foster” | Vague benefit verbs | “speed up”, “let”, “build” | Low (counts when repeated) |
| “meticulous”, “intricate”, “vibrant”, “bustling” | Decorative adjectives | One concrete detail | Low (counts when repeated) |
| “embrace”, “boasts”, “thrive” | Attitude verbs, and “boasts” standing in for “has” | “accept”, “has”, “grow” | Low (counts when repeated) |
| “moreover”, “furthermore”, “additionally” at the start of a paragraph | Connectors stacked where none is needed | “also”, “and”, or nothing | Low (counts when repeated) |
| “at the end of the day”, “the key takeaway”, “the real question is” | Spoken filler used as a signpost | Cut it, then say the point | Low (counts when repeated) |
Why ChatGPT reaches for these words
The clearest evidence comes from scientific abstracts, because millions of them are published every year and the wording before 2022 is a clean baseline. Dmitry Kobak and colleagues counted words in 15.1 million PubMed abstracts from 2010 to 2024 (Science Advances, July 2025). In 2024 “delves” appeared about 28 times as often as the 2021 and 2022 trend predicted, “underscores” about 14 times and “showcasing” about 11 times. Earlier jumps in vocabulary were content words such as the names of new diseases. The 2024 jump was almost all style words: of 379 of them, 66% were verbs and 14% adjectives. From those words the authors estimate that at least 13.5% of 2024 abstracts went through a language model, a lower bound.
Why a model overuses them is less settled. Tom Juzek and Zina Ward (COLING 2025) found 21 such words, among them “delve”, “intricate”, “underscore”, “showcasing”, “realm”, “garnered” and “groundbreaking”. They found no evidence that model design, the training algorithm or the training data explains the overuse. Their comparisons fit the idea that training on human ratings plays a part: if raters liked answers that sounded polished, the model learned to sound polished. They present that as a likely contributor, not a proven cause.
The list changes with each model
Wikipedia's editors, who track this for their own cleanup work, note that “delve” was heavily overused in 2023 and early 2024, became rarer later in 2024 and dropped off sharply in 2025. Their page groups the words by period: “boasts”, “tapestry”, “testament” and “vibrant” for 2023 output, “align with”, “fostering” and “showcasing” for mid-2024 to mid-2025, “emphasizing”, “enhance” and “highlighting” after that.
A September 2026 study by Graphite, a growth agency, compared 10,000 human web articles with 90,000 articles from nine models on the same topics. OpenAI's GPT-6 Astra used marketing words such as “unlock”, “elevate” and “streamline” 73% less often than GPT-4.1, but leaned on evaluative adjectives instead: “dependable” at 59 times the human rate, “practical” at 26 times and “meaningful” at 13 times. The study also found that 65% of the patterns belonged to one model family only.
So a scan for “delve” will miss a 2026 draft. The table above covers the long tail that still turns up, especially in drafts from older models and in copied templates. For newer output, read for words that rate something without describing it.
How to use the table without making it worse
Swap toward plain speech, not toward another fancy word. “Delve into” becomes “look at”, not “dig deep into”. Wikipedia's guide makes the same point from the other side: a word being overused does not mean its synonyms are, so a thesaurus swap only moves the problem.
When you cut a phrase, check what is left. If “plays a crucial role” goes, the sentence needs to say what the thing does. Often the honest fix is a fact, and a word list can't supply facts.
A worked example. Before, 51 words with nine flagged phrases: “In today's fast-paced world, our bakery plays a crucial role in the neighborhood. It's important to note that we leverage locally sourced flour to create a rich tapestry of flavors. Our meticulous bakers embrace tradition, and our new rye loaf is a testament to their passion. Say goodbye to bland bread.”
After, 45 words with none: “Our bakery has sold bread on Elm Street since 2009. The flour comes from a mill 40 miles away. The new rye loaf took Ana four months to get right: it uses a starter she feeds twice a day, and it keeps for five days.”
Every fact in the second version came from the owner. That is the real change, and it is shorter too.
When these words are fine
Context decides. “Crucial” in a clinical note, “landscape” for a painting, “navigate” for a boat and “underscore” for the keyboard character are ordinary uses. Wikipedia's guide asks editors to read the list literally for the same reason.
People who learned English from textbooks, and people who write formally for work, use some of these words naturally. One “moreover” in a 2,000-word report is a habit, not a sign of anything. What reads as generated is density: several of these in one paragraph, next to sentences with no names, numbers or dates.
The reverse also holds. Graphite's study found that human writers use the first person, casual qualifiers like “pretty” and “sort of”, small anecdotes and parenthetical asides far more often than current models. Those are worth keeping in your own drafts.
Where this list comes from
The table is drawn from the English pattern list behind HidenGPT's red pen, which has more than 50 English entries as of 10 October 2026. Each entry carries a weight: High and Medium entries lower the draft's naturalness score on a single use, Low entries only when they appear twice or more, because one “moreover” is normal. The same list runs in the AI cliché checker on hidengpt.com, in your browser, with no sign-up. It is an editing score for readability, not an AI detector, and it can't tell who wrote a text.
Common mistakes
- Swapping one fancy word for another: replace “delve into” with “look at”, not with “dig deep into”.
- Deleting the phrase and leaving an empty sentence: if “plays a crucial role” goes, say what the thing actually does.
- Hunting only for “delve”: newer models lean on evaluative words such as “meaningful” and “dependable”, so read for praise with nothing behind it.
- Treating one hit as proof that someone used AI: look for clusters of these words and for missing names, numbers and dates.
- Pasting a banned-word list into ChatGPT and stopping there: it often moves to a near neighbor, so read the new draft line by line.
Questions
What words does ChatGPT use the most?
It depends on the model version. Older ChatGPT output leaned on “delve”, “tapestry”, “testament”, “crucial”, “pivotal” and “vibrant”; output from 2024 and 2025 favored “showcasing”, “underscore”, “enhance” and “fostering”; a September 2026 study found GPT-6 Astra overusing “dependable”, “practical” and “meaningful”.
Why does ChatGPT say “delve” so much?
Nobody has proven why. A 2025 study by Juzek and Ward found no evidence that model design or training data explains it, and their comparisons fit the idea that training on human ratings rewarded polished-sounding words.
Does ChatGPT still use “delve”?
Much less than it did. Wikipedia's editors report that “delve” became rarer in late 2024 and dropped off sharply in 2025, while newer models picked up other favorite words.
What words should I avoid so my writing doesn't sound like AI?
Start with the High rows in the table: “delve”, “a testament to”, “tapestry”, “unlock the potential” and stock openers like “in today's fast-paced world”. Then cut words that rate something without describing it, and replace them with a fact.
What words does ChatGPT never use?
There is no published list of words it never uses. One study measured the reverse instead: in Graphite's 2026 comparison, human articles used first person, casual qualifiers like “pretty” and “sort of”, short anecdotes and parenthetical asides far more often than any current model.
Can I tell ChatGPT not to use certain words?
Yes, in your prompt or in Custom Instructions, and it usually listens. Expect it to reach for a close substitute instead, so a word ban works best next to a request for plain, specific language and your own facts.
Is using these words proof that something was written by AI?
No. People use every word on this list, and Wikipedia's own guide warns that these are signs to look closer, not proof. Several in one paragraph, with no specifics around them, is a reason to edit, not to accuse.
Sources
- Kobak et al., Delving into LLM-assisted writing in biomedical publications through excess vocabulary, Science Advances (2025) (checked October 10, 2026)
- Juzek and Ward, Why Does ChatGPT “Delve” So Much? Exploring the Sources of Lexical Overrepresentation in Large Language Models, COLING 2025 (checked October 10, 2026)
- Wikipedia: Signs of AI writing (editors' field guide) (checked October 10, 2026)
- Graphite research: AI Tells (16 September 2026) (checked October 10, 2026)