The AI writing words to avoid in fiction fall in four groups: gesture verbs (nodded, gazed, stared, leaned, paused), sound and light words (whispered, echoed, flickered, shimmering, glow), mood words (unease, anticipation, dread, curiosity) and the glossy adjectives (relentless, intricate, vibrant, testament, tapestry). We counted them in 1,321 scenes from 13 AI models: three frontier models used none of the classic phrases in about 27,000 words each, and Llama 4 Maverick used one every 200 words. So the list below tells you what to cut, and the table tells you how much of it your model will hand you.
The 40 words, in the order AI overuses them
The most-used public list comes from antislop-sampler (Sam Paech, Apache 2.0, files dated 2025-04-07). Its README says the lists are mostly auto-generated by computing words over-represented in a large dataset of LLM-written stories, compared with normal human writing. The first 500 words of that list come with counts; these are the top of it with names left out, checked 15 November 2026.
| Group | Words (highest count first) | Why it reads as machine |
|---|---|---|
| Gesture verbs | nodded, stared, hesitated, leaned, paused, shook, glanced, blinked, swallowed, shifted | Every beat gets a body movement whether it earns one or not |
| Sound | whispered, whisper, echoed, echoing, hissed, murmured | Every voice is quiet and every room has an echo |
| Light | glow, faint, flicker, flickered, shimmering, glowing, gleaming | The world is always dimly lit |
| Mood | curiosity, unease, anticipation, disbelief, dread, despair | The feeling is named instead of shown |
| Glossy adjectives | relentless, vibrant, intricate, unwavering, profound, subtle, newfound, countless | Large words that point at no particular thing |
| Stock nouns | shadows, scent, spine, testament, realm, void, horizon, warmth | The same props in every scene |
Two things in the data are not words at all. The names come first: "elara" is the first word on the list, with the highest count (98,586), and "elias" and "kael" follow in the top 25. At least ten of the first 500 entries are first names. The second is the phrase list. These are the most common whole phrases in the same source, by count:
| Phrase | Count in source |
|---|---|
| took a deep breath | 28,671 |
| voice barely above a whisper | 19,295 |
| couldn't help but feel | 15,254 |
| voice barely audible | 9,460 |
| casting long shadows | 9,376 |
| couldn't shake the feeling | 8,483 |
| heart pounding in her chest | 4,473 |
| air was thick with the scent | 5,140 |
How often do models still do it? Our own count
We ran the lists over every scene the 13 models wrote for our earlier study, Banned by the Bot: 34 prompts, three tries each, the same settings for every model, 99 to 102 scenes per model. Results are in the AI Slop Index, measured 21 October 2026.
| Model | Classic phrases per 1,000 words | List words per 1,000 words | Most-used name |
|---|---|---|---|
| Claude Sonnet 5.5 | 0 | 15.4 | Mara |
| GPT-6.1 Sol | 0 | 15.9 | Mara |
| Grok 4.7 | 0 | 16.8 | Mara |
| Claude Opus 5.5 | 0.15 | 14.2 | Mara |
| Kimi K3 | 0.18 | 13.4 | Mara |
| Gemini 4 Argon | 0.23 | 35.0 | Kaelen |
| Qwen3.8 Max | 0.29 | 12.3 | Mara |
| Mistral Large 4 | 0.43 | 25.9 | Elias |
| DeepSeek V4 Pro | 0.57 | 27.6 | Mara |
| DeepSeek V4.1 Flash | 0.70 | 19.5 | Elias |
| Cydonia 24B | 2.95 | 36.1 | Elara |
| Venice Uncensored | 4.69 | 48.6 | Elara |
| Llama 4 Maverick | 5.02 | 52.6 | Lena |
Read it this way:
- The famous phrases are mostly gone from current frontier models. "Barely above a whisper" is now an uncensored-model and older-model problem.
- The single words are not. Even the cleanest models land at 12 to 17 list words per 1,000 words, because many of the words are ordinary (nodded, smiled, paused) and people use them too. A list word is a prompt to look, never proof of AI.
- Names repeat. For seven of the 13 models the most-used name was "Mara"; the others had their own favourite (Elias, Elara, Kaelen, Lena). GPT-6.1 Sol called a character Mara in 23% of its scenes. If your book has an Elara, Elias or Mara you did not pick on purpose, change it.
- The list is a 2025 list. Newer models have newer tics it does not hold: a low score means a model avoids the known ones, not that it has none.
What to write instead
A replacement word rarely fixes it, because the problem is the move, not the word. Three swaps that work:
| The move | Before (written for this page) | After |
|---|---|---|
| Naming the feeling | "A sense of unease washed over Nell as she nodded." | "Nell said fine and kept her hand on the banister." |
| The quiet voice | "'Don't,' he whispered, his voice barely above a whisper." | "'Don't.' He had to say it twice before the kettle stopped." |
| The faint light | "A faint glow flickered in the window." | "The upstairs light came on, then went off again." |
In all three the cut word is replaced by something particular to this scene: a banister, a kettle, a light switching. That is the rule. The fix is a detail that could not appear in any other story.
A 5-minute pass for your draft
- Search your chapter for the 40 words above and count them. More than one per hundred words in a scene is worth a look.
- Keep a word where a person would have used it, and the AI's choice happens to be right.
- For each feeling word, ask what the character does with their hands, voice or eyes that shows it. Replace the word with that.
- For each whisper, echo or glow, ask whether the scene needs that sound or light. If not, delete the sentence.
- Change every default name you did not choose.
- Read the paragraph aloud. If you stumble over a sentence, rewrite it, and leave the ones you did not stumble on. A paraphraser can't do this for you: QuillBot's AI writer rewords the whole selection, the sentences you were happy with included.
Why the same words come back
Language models predict the likeliest next word, and in fiction the likeliest word is the one thousands of earlier stories used. A Science Advances study of more than 15 million biomedical abstracts from 2010 to 2024 found an abrupt rise in certain style words after LLMs arrived, and estimated that at least 13.5% of 2024 abstracts were processed with LLMs (up to 40% in some subcorpora). That is science writing, not fiction, but the mechanism is the same: a model's favourite words show up in its text far more often than in people's.
Where the other fixes fall short
- A banned-words list in the prompt ("never use whisper") works for that session only, and the model usually trades the word for its nearest sibling (murmured, breathed). The list above is long because the family is long.
- Run it through an AI detector and you learn nothing useful: detectors misfire on fiction and say nothing about which sentence to fix.
- A sampler that bans the list is a good tool for people running their own models, but it only knows the 2025 list and changes nothing about the moves (naming the feeling, the quiet voice).
How KudoNovel helps
KudoNovel builds the prose from your own paragraph, so most of the words in the result are yours. You type the beat, press Expand, and the paragraph is written out in place. The story bible keeps names and facts, so a character you named stays that character. When a stock word slips in, highlight the paragraph, press Steer and write one line: "cut the stock phrases, name no feelings, give her something to do with her hands." Each result is a version, so Undo and the Versions panel bring back the earlier one, and your own wording stays the base every time.
Questions people ask
What words does AI overuse in writing? Gesture verbs (nodded, stared, paused), sound and light words (whispered, echoed, flickered), named feelings (unease, anticipation) and glossy adjectives (relentless, intricate, testament), per the public over-representation list.
Is "delve" an AI word? The Science Advances study above found a sharp rise in LLM-style words in scientific abstracts after 2022, "delves" among them according to press coverage of the paper. In fiction, the list above is the better guide.
Do newer models still write "voice barely above a whisper"? In our count, Claude Sonnet 5.5, GPT-6.1 Sol and Grok 4.7 used none of the classic phrases in about 27,000 words each; Llama 4 Maverick used one every 200 words.
Can I get rid of AI words with a find-and-replace? Not safely: the word is a symptom of the move. Cut or rewrite the sentence using the 5-minute pass above.