Readers can spot AI prose by its tics: the voice “barely above a whisper”, the shadows that are always long, the air “thick with the scent” of something. We wanted to know which models still write like that. So we counted, across more than 300,000 words of fiction that 13 current models wrote for our first study, using a public list of the phrases AI over-uses in stories compared with human writing.
The gap is large. Three of the frontier models we tested, Claude Sonnet 5.5, GPT-6.1 Sol and Grok 4.7, didn't use a single phrase from the list in about 27,000 words each. Llama 4 Maverick used one every 200 words, and the two uncensored, writing-tuned models were close behind.
Not every tic is gone. Qwen3.8 Max wrote the most “it wasn't X, it was Y” sentences, 3.2 per 1,000 words. Kimi K3 and Mistral Large 4 used about ten em dashes per 1,000 words, while Claude Opus 5.5 used one in 26,723 words. And names repeat: GPT-6.1 Sol called a character Mara in 23 of its 102 scenes.
What we found
- Frontier models have scrubbed the classic clichés. Claude Sonnet 5.5, GPT-6.1 Sol and Grok 4.7: zero phrases from the list. Claude Opus 5.5, Kimi K3 and Gemini 4 Argon: fewer than one per 4,000 words.
- Uncensored and older open models are where the slop lives. Llama 4 Maverick (5.0 per 1,000 words), Venice Uncensored (4.7) and Cydonia 24B (3.0) wrote 314 of the 381 listed phrases found in the whole study.
- The clichés that turned up most: “casting long shadows” (16 times), “air hung thick” (8), “words hung in the air” (7) and “took a deep breath” (7).
- “It wasn't X, it was Y” survives everywhere. Every model wrote some, from 0.6 to 3.2 per 1,000 words; Qwen3.8 Max the most.
- The em dash is a house style now, not an AI signature: about ten per 1,000 words from Kimi K3 and Mistral Large 4, almost none from Claude Opus 5.5 (0.04) and Llama 4 Maverick (0.08).
Clichés per 1,000 words
Phrases from a public list of the phrases AI over-uses in fiction, counted in each model's 102 scenes. Lower is better.
The full table
Per 1,000 words, except names: the share of a model's scenes that use its favourite name.
| Model | Cliché phrases | Over-used words | “Not X, but Y” | Em dashes | Favourite name | Its top clichés |
|---|---|---|---|---|---|---|
| Claude Sonnet 5.5 | 0.00 | 15.39 | 0.89 | 0.18 | Mara (5%) | none |
| GPT-6.1 Sol | 0.00 | 15.88 | 0.64 | 0.87 | Mara (23%) | none |
| Grok 4.7 | 0.00 | 16.84 | 1.34 | 4.03 | Mara (5%) | none |
| Claude Opus 5.5 | 0.15 | 14.22 | 1.09 | 0.04 | Mara (6%) | “first time in years” ×2 “pushed open the door” ×1 “first time in his life” ×1 |
| Kimi K3 | 0.18 | 13.40 | 1.35 | 9.68 | Mara (3%) | “something else entirely” ×2 “one last time” ×1 “first time since” ×1 |
| Gemini 4 Argon | 0.23 | 35.04 | 0.75 | 5.02 | Kaelen (3%) | “said, her voice dripping” ×2 “sighed, a sound like” ×1 “gaze locked onto” ×1 |
| Qwen3.8 Max | 0.29 | 12.34 | 3.21 | 6.55 | Mara (8%) | “three days since” ×2 “like a living thing” ×2 “would one day” ×1 |
| Mistral Large 4 | 0.43 | 25.94 | 2.04 | 9.47 | Elias (11%) | “three days since” ×2 “one last time” ×1 “like distant thunder” ×1 |
| DeepSeek V4 Pro | 0.57 | 27.62 | 1.58 | 6.34 | Mara (3%) | “casting long shadows” ×2 “something else, something” ×1 “one last time” ×1 |
| DeepSeek V4.1 Flash | 0.70 | 19.54 | 2.14 | 4.12 | Elias (2%) | “inside, the air was thick” ×2 “first time in my life” ×2 “brow furrowed in concentration” ×1 |
| Cydonia 24B | 2.95 | 36.08 | 1.94 | 4.82 | Elara (5%) | “casting long shadows” ×7 “like a second skin” ×3 “air hung thick” ×3 |
| Venice Uncensored | 4.69 | 48.55 | 0.59 | 0.72 | Elara (3%) | “air hung thick” ×5 “took a deep breath” ×4 “eyes wide with fear” ×3 |
| Llama 4 Maverick | 5.02 | 52.57 | 0.79 | 0.08 | Lena (8%) | “words hung in the air” ×6 “air was heavy with the scent” ×5 “casting long shadows” ×5 |
The clichés that turned up most
- “casting long shadows”×16
- “first time in my life”×9
- “one last time”×8
- “air hung thick”×8
- “words hung in the air”×7
- “took a deep breath”×7
- “air was thick with the scent”×6
- “eyes never leaving”×6
- “voice barely above a whisper”×6
- “eyes wide with fear”×6
How we measured it, and its limits
Use this data
The dataset
Free to cite and republish under CC BY 4.0, with a link to this page.
Download the dataset (CSV)About KudoWrite
KudoWrite is a browser app for writing long fiction with AI. A chapter is one page of paragraphs: the writer types each one, the AI expands it, and a story bible and a running summary of the whole book keep up as they write, so chapter 30 is written with chapter 1 in context.
KudoWrite adds no content filter of its own: romance, horror, dark and explicit themes are the writer's call, within the law (a model's own behaviour still applies). No subscription: a free credit to start, then one-off credit packs from $3.50. Made by a solo founder in France. Contact: admin@kudowrite.com.
Other reports
Your voice, not the model's
In KudoWrite you sketch each beat, the AI writes it, and you steer it with a one-line instruction or edit it by hand, so the clichés stop with you. The story bible and a running summary keep chapter 30 consistent with chapter 1.
Start writing free