Graphite updated its study of stylistic tics in language models on 1 October 2026, adding Claude Opus 5.5. The result cuts both ways: the model has almost abandoned em-dashes, yet the count of phrases it overuses fell by only 4%.
Key takeaways
- 9,974 aligned topics across ten models, with a human baseline from articles predating ChatGPT
- A tell is a word, phrase or frame appearing at least twice as often in AI as in human writing
- Opus 5.5 carries 2,548 tells, just 4% fewer than Opus 5
- Em-dash use fell 99%, from 2.92 to 0.015 per 1,000 words
- Word-distribution divergence moved 19% closer to human writing and the mannered-prose score dropped 37%
How the tells were counted
The method is simple, which is why it is hard to dispute. The Graphite team — Gregory Druck, Jose Luis Paredes and Ethan Smith — compared word and phrase frequencies in model output against articles written before ChatGPT. The models got summaries of the same 9,974 topics and rewrote them, which strips out topic effects.
Symbol meaning
- …
- frequency of the word or phrase in model output
- …
- frequency in articles predating ChatGPT
- …
- threshold: at least twice as often as humans
The em-dashes went, the mannerisms stayed
The showiest change involves the mark that became an informal signature of AI text. Opus 5.5 places an em-dash 0.015 times per 1,000 words, against 2.92 in Opus 5.
Removing one signal does not alter the rest of the profile: 2,548 tells?Tell: A word, phrase or sentence frame that shows up at least twice as often in model output as in human writing. remain, 4% fewer than before.
Every model has its own tic
As TechCrunch reports, the overused phrases differ enough between models to work as a signature on their own.
| Model | Tell | More often than humans |
|---|---|---|
| Opus 5.5 | "this matters" | ×116 |
| Opus 5.5 | "why X matters" | ×92 |
| Opus 5.5 | "dependable" | ×23 |
| GPT-6 Astra | corrective framing ("not simply X") | over ×100 |
GPT-6 Astra sticks to corrective framing?Corrective framing: A rhetorical move where a sentence first rejects a simplified version ("not simply X") before stating the real claim., while Opus 5.5 keeps announcing that a topic matters. Astra scores better on mannered prose?Mannered prose score: A measure of how saturated the text is with ornate, unnaturally polished turns of phrase. Lower means closer to ordinary writing., while Opus 5.5 leans harder on superlatives and helpfulness.
Why it matters
A model's style is not cosmetic but a durable imprint of training. Labs can remove the one signal that became a meme, yet vocabulary distribution shifts far more slowly. That has practical consequences: detecting generated text need not hinge on one punctuation mark, and editors gain a measurable criterion, not a hunch. For the models, sounding human is harder than fixing one parameter.
What's next?
- Graphite runs the study in editions — the first on 16 September 2026, the Opus 5.5 update a month later
- The method is repeatable on a fixed base of 9,974 topics, so each new model version can be measured on the same scale
Sources
- Graphite — AI Tells: Opus 5.5 Update
- TechCrunch — Opus 5.5 loves to tell you 'this matters' (and other AI writing tells)
