A reference list of the words ChatGPT overuses, from delve to tapestry, with plain swaps, the research on why it happens, and why swapping words is not enough.
If you've read much AI-generated text, you already know the feeling. A paragraph "delves into" something, calls it "pivotal," and wraps up by saying the whole thing is "a testament to" human ingenuity. None of those words is wrong. There are just far too many of them, packed far too close together.
This is a working reference: the words and phrases chat models lean on most, grouped by type, with a plain swap for each. After the lists, we'll cover what the research actually says about why models write this way, and why a find-and-replace pass won't make a draft sound like you.
What this list is for (and what it isn't)
A list like this is a self-editing aid. It helps you find the places where a draft went generic, so you can say the thing more directly. It isn't a detector. People used "delve" and "robust" long before ChatGPT, and some varieties of English use them a lot. Finding a few of these words in someone's essay proves nothing about how it was written.
Use it on your own drafts, especially ones you started with an AI tool. If your school or employer has rules about AI use, those rules come first.
Verbs ChatGPT overuses
Verbs are the biggest group. When researchers counted the style words whose use jumped in 2024 biomedical abstracts, 66% of them were verbs. Most of these swap for something shorter.
| Overused word | Try instead |
|---|---|
| delve into | look at, dig into, cover |
| leverage | use, build on |
| foster | build, encourage, help |
| navigate | handle, deal with, get through |
| underscore | show, stress, make clear |
| showcase | show, display |
| utilize | use |
| embark on | start, begin |
| streamline | simplify, speed up |
| unlock | open up, make possible |
| boast (as in "the city boasts") | has |
The best swap is often no swap at all. "This report delves into the causes of churn" becomes "This report looks at why customers leave." Better still, skip the announcement and start with the first cause.
Adjectives and nouns
These tend to inflate a claim without adding any information. "A robust and seamless solution" tells the reader nothing they can check.
| Overused word | Try instead |
|---|---|
| pivotal | key, main, or say what it decided |
| crucial | important, needed, or cut it |
| robust | strong, reliable, well tested |
| seamless | smooth, easy |
| comprehensive | full, complete, detailed |
| intricate | complex, detailed |
| meticulous | careful, thorough |
| vibrant | lively, busy, colorful |
| groundbreaking | new, first, or name what changed |
| tapestry | mix, range |
| testament (to) | proof, sign, shows |
| realm | field, area, world |
| landscape | market, field, situation |
A quick test: if you can delete the adjective and the sentence loses nothing, delete it. If it does lose something, replace it with the specific fact. "A pivotal moment for the company" is weaker than "the quarter they nearly ran out of cash."
Transitions and filler phrases
These are the phrases that pad the start and end of paragraphs. Most can simply go.
| Overused phrase | Try instead |
|---|---|
| It's important to note that | Cut it and state the point |
| In today's fast-paced world | Cut it, or name the actual change and the year |
| Moreover / Furthermore | Also, and, or nothing |
| Additionally | Also, plus, or a new sentence |
| Notably | Cut it, or say why it matters |
| When it comes to | For, with, on |
| Let's dive in | Cut it and start |
| Ultimately | Cut it, or use "in the end" |
| In conclusion / In summary | Cut it and end on your last real point |
Sentence patterns that give a draft away
Single words are the easy part. The patterns below are harder to spot because each one is fine on its own. The problem is how often they repeat.
| Pattern | Example | Fix |
|---|---|---|
| "Not just X, but Y" / "It's not X, it's Y" | "It's not just a tool, it's a partner." | Say the Y part plainly: "It drafts your replies." |
| Lists of three, every time | "fast, reliable, and secure" | Keep the one that matters, or use two, or four |
| "X serves as a testament to Y" | "The bridge serves as a testament to the city's resilience." | "The bridge survived two floods." |
| "Despite its X, Y faces challenges" | "Despite its growth, the sector faces several challenges." | Name the challenge in the first sentence |
| Question, then answer | "So what does this mean? It means..." | State what it means |
| A closing paragraph that repeats the intro | "In summary, X is key to Y." | End on a next step, a cost or an open question |
Wikipedia's volunteer editors keep a long guide to these habits, Signs of AI writing, written by WikiProject AI Cleanup. It lists "negative parallelisms" like "not only... but also," the rule of three, and a set of "AI vocabulary" words that includes delve, tapestry, testament and landscape. It's advice, not policy, but it's a good field guide.
The evidence: these words really did spike
This isn't just a hunch. Two large studies measured it.
Biomedical abstracts. Dmitry Kobak and colleagues analyzed about 15 million PubMed abstracts from 2010 to 2024 in "Delving into LLM-assisted writing in biomedical publications through excess vocabulary", published in Science Advances in July 2025. They compared how often each word appeared against what earlier trends predicted. In 2024, "delves" appeared about 28 times more often than expected, "underscores" about 14 times and "showcasing" about 11 times. Common words like "potential," "findings" and "crucial" rose too. From that excess, they estimated that at least 13.5% of 2024 abstracts were processed with an LLM, with some groups of papers reaching 40%. By their measure, the effect on vocabulary was larger than the shift caused by Covid.
Peer reviews. A Stanford-led team studying reviews written for AI conferences found that in ICLR 2024 reviews, "meticulous" was about 34.7 times more likely to appear in a sentence than in earlier reviews, "intricate" 11.2 times and "commendable" 9.8 times (Liang et al., ICML 2024).
Why does ChatGPT say "delve" so much?
Short answer: nobody knows for certain, but the evidence points to how chat models are tuned after their main training.
The tuning step rewards a certain style
Chat models aren't only trained to predict text. They're also tuned on human judgments. In OpenAI's InstructGPT paper, labelers ranked several model answers to the same prompt, a reward model learned to predict which answers people preferred, and the language model was then trained to score well on that reward. This is reinforcement learning from human feedback, or RLHF. If raters lean even slightly toward answers that sound formal and thorough, the model can learn to produce that sound everywhere.
Linguists Tom Juzek and Zina Ward tested this in "Why Does ChatGPT 'Delve' So Much?" (COLING 2025). They picked out 21 overused words, including delve, intricate, underscore, showcasing, realm and groundbreaking, and found no evidence that model architecture or training data explained the overuse. Comparing Llama 2 before and after chat tuning was consistent with RLHF playing a part, though they called their human experiment inconclusive. In a follow-up study that mimicked the feedback process, participants picked the version with more tuning-favored words 52.4% of the time. That's a small edge, but a small, consistent preference is exactly what this kind of training is built to pick up.
The annotator theory, and why it's shaky
In April 2024, The Guardian's Alex Hern suggested that "delve" might come from the people doing the rating, pointing out that the word is more common in Nigerian business English and that some of this work is outsourced to Africa (The Guardian). The idea spread fast. But when Juzek and Ward checked their 21 words against the International Corpus of English, they found no variety of English where those words were especially common, so their analysis did not support it. Treat it as an interesting idea, not a settled fact.
Safe words are a safe bet
There's also a simpler pull. A model picks likely words, and formal, upbeat words fit almost any topic. "Pivotal" works for a battle, a product launch or a soccer match. Specific words only fit one situation, so a model hedging its bets drifts toward the general ones.
Why swapping words isn't enough
Here's the catch. You can replace every word in the tables above and still end up with a draft that reads like a machine wrote it. The words are a symptom. The deeper habits are structural: sentences of similar length, every paragraph built the same way, claims with no examples, a tidy summary at the end. We go through those in why AI writing sounds robotic.
Blind swapping can also backfire. A thesaurus might turn "utilize" into "employ," which is just as stiff. And if you replace "crucial" with "important" fifty times, you've only moved the repetition.
What works better:
- Cut before you swap. Most filler phrases and many adjectives can go entirely.
- Replace vague with specific. A number, a name or an example does more than any adjective.
- Read it out loud. If you wouldn't say a sentence to a colleague, rewrite it.
- Vary the rhythm. Drop in a short sentence after two long ones.
- Add what only you know. Your own experience, opinion or source is the part no model can supply.
Tools that help with the cleanup
Chat output often carries formatting leftovers too: em dashes, curly quotes, stray markdown asterisks and hidden characters. Our free AI text cleaner strips those out without changing your words.
For the wording itself, the WriteHumanly humanizer rewrites stiff AI-assisted drafts so they read naturally and keep every fact you put in. Think of it as a starting point, not a replacement for your own edit. Read the result, add your own details, and make sure what you hand in follows the rules you're working under.
Frequently asked questions
Why does ChatGPT say "delve" so much?
The leading explanation is the human feedback tuning step, where small rater preferences for formal, polished answers can get amplified. Research by Juzek and Ward is consistent with this, though they describe the evidence as mixed. The popular theory that it comes from Nigerian English wasn't supported by their corpus check.
Does using these words mean my writing will be flagged as AI?
Not on its own. These words were common in human writing long before chatbots, and no single word proves anything. Heavy, repeated use alongside uniform sentence structure is what makes text read as machine-made.
Is there an official list of AI words?
No. The best evidence comes from studies like the PubMed excess vocabulary paper. Wikipedia's "Signs of AI writing" guide is a useful practical list.
Will replacing these words make AI text sound human?
Only partly. Word swaps fix the most obvious tells, but structure, rhythm and a lack of specific detail matter more. Use the list to find weak spots, then rewrite those sentences in your own words.
Written by
WriteHumanly Team
The team behind WriteHumanly has spent thousands of hours studying how AI detectors actually score text, building tools used by students and professionals worldwide. We publish what we learn so other writers can make better decisions.
Ready to humanize your AI text?
Paste your content and get human-sounding output in seconds.
Try WriteHumanly Free
