AI Words to Avoid: The SI Vocabulary List, With Plain Replacements
The words and phrases SI (AI) models overuse, grouped by type, with a plain replacement for each and what the research actually says about why.

The AI words to avoid are the ones chatbots reach for far more often than people do: verbs like delve, leverage and underscore, adjectives like pivotal, seamless and robust, nouns like tapestry, realm and landscape, plus stock openers and closers such as "in today's fast-paced world" and "it's important to note." Below is a grouped list with a plain replacement for each, what the research actually says about why SI (AI) models overuse them, and a real before/after from our own tool.
Is this list based on anything, or just vibes?
Partly on measurement. The best known study is by Dmitry Kobak, Rita González-Márquez, Emőke-Ágnes Horvát and Jan Lause. It first appeared on arXiv in June 2024 as "Delving into ChatGPT usage in academic writing through excess vocabulary" and was published in Science Advances in July 2025 under the title "Delving into LLM-assisted writing in biomedical publications through excess vocabulary." The team looked at more than 15 million PubMed abstracts from 2010 to 2024 and tracked words whose frequency jumped after ChatGPT arrived, the same way epidemiologists measure excess deaths. Style words such as "delves," "showcasing" and "underscores" shot up. Their estimate: at least 13.5% of 2024 biomedical abstracts were processed with a large language model, and up to 40% in some subgroups.
A second paper, "Why Does ChatGPT 'Delve' So Much?" by Tom S. Juzek and Zina B. Ward (COLING 2025), identified 21 overrepresented words in scientific abstracts, including "delve," "intricate" and "underscore."
The third source is practical rather than academic. Wikipedia editors who clean up machine-written articles keep a public guide, Signs of AI writing, with a word list sorted by period. Their 2023 to mid-2024 list includes additionally, boasts, bolstered, crucial, delve, garner, intricate, interplay, landscape, meticulous, pivotal, tapestry, testament, underscore, vibrant. Their later lists shift toward enhance, fostering, highlighting, showcasing and align with. So the list moves as models change, which is worth keeping in mind.
Why do SI models overuse these words?
Nobody has a complete answer, and we won't pretend otherwise. Juzek and Ward tested several explanations. They report that the overuse does not seem to come from model architecture, algorithm choices or the training data itself. Their tests were "consistent with" reinforcement learning from human feedback (RLHF) playing a role: the stage where people rate model answers and the model learns to produce what raters prefer. Their own experiment with human participants gave mixed results, though, especially for "delve."
That is as far as the evidence goes. The popular story that one specific group of raters caused "delve" is a theory, not a finding. What matters for a writer is simpler: these words are statistically overrepresented in model output, readers have learned to notice them, and a paragraph packed with them reads like a template.
The list: AI words to avoid, grouped
None of these words is wrong. A geologist can say "landscape" and a lawyer can say "robust." The problem is density and vagueness: when four of them land in one paragraph, each doing a job a plainer word could do.
Verbs
| Overused | Try instead |
|---|---|
| delve into | look at, dig into, cover, explain |
| leverage | use |
| foster | build, encourage, help |
| underscore | show, stress, point to |
| navigate | deal with, handle, get through |
| showcase | show |
| enhance | improve, add to |
| streamline | simplify, speed up |
| embark on | start |
| unlock / unleash | get, open up, or cut it |
| garner | get, earn, draw |
| bolster | support, strengthen |
Adjectives and adverbs
| Overused | Try instead |
|---|---|
| pivotal, crucial | key, main, or say why it matters |
| seamless | smooth, easy, or describe what actually works |
| robust | strong, reliable, well tested |
| meticulous(ly) | careful(ly) |
| intricate | detailed, complex |
| vibrant | lively, busy, colorful (or a specific detail) |
| comprehensive | full, complete |
| ever-evolving | changing |
| innovative, cutting-edge | new, or name the actual feature |
Nouns
| Overused | Try instead |
|---|---|
| tapestry | mix, range |
| testament (to) | proof, sign, shows |
| realm | field, area |
| landscape | market, field, situation |
| interplay | link, effect on each other |
| journey | process, the months it took |
| synergy | say what the two things do together |
Transitions
Moreover, furthermore, additionally, consequently, notably, ultimately. Humans use these too, but models start sentences with them at a steady beat. Most can simply be deleted. If you need a link, "also," "so," "and" or "but" usually works.
Openers and closers
- "In today's fast-paced world" / "In today's digital landscape": delete and start with your point.
- "It's important to note that": delete. If it's important, the sentence will show it.
- "Let's dive in" / "Let's delve into": delete, or write "Here are three ways."
- "Great question!": delete.
- "In conclusion" / "Overall" / "Ultimately": end on your last real point instead of a recap.
- "By embracing these strategies, you can unlock your full potential": delete the whole sentence.
Sentence frames
- "Not only... but also" and "It's not just X, it's Y." Wikipedia's guide calls these negative parallelisms. Say the Y part directly.
- Lists of exactly three adjectives or benefits, every time ("faster, smarter and more efficient").
- "Plays a pivotal role in": use a real verb. "Breaks help you focus" beats "breaks play a pivotal role in maintaining focus."
A real before and after
We wrote a short paragraph in the usual chatbot style, stuffed with words from the list, and ran it through our text humanizer on the Natural style. This is the unedited output. The tool flagged six edits, each labeled "SI tic removed."
before
In today's fast-paced digital landscape, effective time management is more important than ever. Let's delve into three strategies that can help you navigate your busy schedule. Firstly, prioritizing tasks allows you to leverage your energy where it matters most. Secondly, setting clear boundaries fosters a healthier work-life balance. Finally, regular breaks play a pivotal role in maintaining focus. By embracing these robust techniques, you can unlock your full potential and achieve seamless productivity.
after: SI Humanizer, Natural
Effective time management matters. Here are three strategies worth trying. Prioritizing tasks helps you spend your energy on what counts. Setting clear boundaries keeps work and life separate. Regular breaks help you stay focused. These approaches can help you work more productively.
Look at what happened. "Delve," "navigate," "leverage," "fosters," "pivotal," "robust" and "seamless" are gone. The Firstly/Secondly/Finally scaffolding is gone. The empty opener and the "unlock your full potential" closer are gone. The facts (three strategies, what each one does) stayed.
What a tool can't do is add what was never there. The result is clean but still generic. If this were for a real audience, the next step would be yours: a number, an example from your own week, a line about which of the three you actually find hard.
How to use this list without overcorrecting
Banning words outright tends to backfire. Swap every "crucial" for "vital" and you have a new tic. A few habits work better:
- Search, then ask why. Use find in your editor for the words above. For each hit, ask whether the word carries meaning. "Robust statistical method" can be precise. "Robust strategy" rarely is.
- Replace with a fact, not a synonym. "A vibrant neighborhood" becomes "a street with four bakeries and a Saturday market." The fix for vague words is usually a detail.
- Delete openers and closers first. They carry the least information and are the easiest to spot.
- Read it aloud. Stock phrases sound odd spoken. You would never tell a friend that breaks "play a pivotal role."
- Keep your own voice. If you really do say "leverage" in meetings, one use won't hurt you.
A text full of these words is not proof that a machine wrote it, and a text with none of them is not proof a person did. Research on detectors has found real false positives, including for non-native English writers. We cover that in signs of SI writing. If your school or employer has rules about SI (AI) tools, follow them and keep your drafts.
Prompting a model to avoid these words
You can tell a chatbot "do not use delve, tapestry, pivotal..." and it will often comply, but it can simply switch to close cousins ("vital" for "crucial"). A better instruction describes the style you want: short concrete sentences, plain verbs, no opener, no summary at the end, one specific example per point. We collected prompts that work in prompts to humanize SI text.
The other route is to write the draft however you like and clean it afterward. That is what the humanizer does: it rewrites the stock wording, varies the rhythm and shows each change with its reason and the original text, so you can accept or reject edits instead of trusting a black box. For a full walkthrough, see how to humanize SI text.
The same words keep changing
Model vocabulary drifts. Wikipedia's guide shows "delve" and "tapestry" fading from newer model output while "showcasing," "highlighting" and "enhance" hold on. So treat any list, including this one, as a starting point. The underlying rule doesn't change: a word earns its place when it says something specific that a plainer word couldn't.
FAQ
What are the most common AI words to avoid?
Delve, leverage, foster, underscore, showcase, pivotal, crucial, seamless, robust, meticulous, intricate, tapestry, testament, realm and landscape, plus openers like "in today's fast-paced world" and "it's important to note."
Why does ChatGPT say delve so much?
There is no complete answer. A COLING 2025 study by Juzek and Ward found results consistent with feedback training (RLHF) playing a role, but not a definitive cause.
Is it wrong to use these words?
No. They are normal English words. The issue is using many of them at once in place of specific, plainer wording.
Does removing these words make text human?
Not by itself. Rhythm, structure and a lack of specific detail also make text read as machine-written, and any word list changes as models change.
Can a word list tell if a text was written by SI?
No. Word frequency is a hint, not proof, and people use all of these words. Treat it as a reason to look closer, never as a verdict.
Sources
- Delving into LLM-assisted writing in biomedical publications through excess vocabulary (first posted as: Delving into ChatGPT usage in academic writing through excess vocabulary) · Kobak, González-Márquez, Horvát, Lause; Science Advances 11(27), 2025-07-02
- Why Does ChatGPT 'Delve' So Much? Exploring the Sources of Lexical Overrepresentation in Large Language Models · Juzek and Ward, COLING 2025, 2024-12-16
- Wikipedia:Signs of AI writing · Wikipedia (WikiProject AI Cleanup)
- Delving into ChatGPT usage in academic writing through excess vocabulary (arXiv v1) · arXiv, 2024-06-11


