> ## Content Index
> Fetch the complete content index at: https://www.metatalks.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# AI models lose familiar writing habits and pick up new ones, Graphite finds
- URL: https://www.metatalks.ai/ai-models-writing-habits-graphite/
- Published: 2026-10-06T10:23:00.000Z
- Updated: 2026-10-06T10:22:59.000Z
- Author: Al
- Tags: News, AI Detection, #newswire

**Across nine models, the study identified nearly 13,000 words, phrases and patterns used at least twice as often as in human articles.**

Marketing firm Graphite found 2,548 recurring writing tells in Claude Opus 5.5, just 4% fewer than in Opus 5\. Its [October 1 update](https://graphite.io/five-percent/research/ai-tells-opus-5-5-update?ref=metatalks.ai) also found that em-dash use had fallen 99% between those versions. Thousands of other expressions remained disproportionately common in its writing.

Opus 5.5 repeatedly explained why things mattered. It used “this matters” at 116 times the human rate and the pattern “why X matters” at 92 times that rate. Compared with Opus 5, it also used “can help you” eight times as often, while its predecessor used “is genuinely” 26 times as often.

OpenAI’s Astra favored defining things through contrasts. In Graphite’s [original study](https://graphite.io/five-percent/research/ai-tells?ref=metatalks.ai), “not simply” appeared at 157 times the human rate and “rather than relying” at 187 times. Opus 5 favored superlatives, while Gemini 3.1 Pro favored formal transitions. Across that comparison, 65% of the tells were confined to one model family.

Successive versions also differed substantially. Only 45% of the tells found across Astra and its predecessor, Sol, appeared in both. Familiar punctuation had also become a poor guide: Astra used em dashes 88% less often than the human samples, while Gemini 3.1 Pro had almost stopped using them.

The models used a wider vocabulary within individual articles, longer words and less variation in sentence length than the human writers. Graphite stresses that vocabulary variety within an article does not establish greater variety of writing styles across articles.

The total number of tells did not follow one trend. Between the earliest and latest versions in the original study, it fell 29% for Claude and 32% for Gemini, but rose 48% for GPT.

Graphite separately compared the models’ overall word distributions with human writing. In the follow-up, that statistical distance fell 19% from Opus 5 to Opus 5.5, but rose 8% from Sol to Astra. By this measure, Claude moved closer to the human sample and GPT moved further away, even though both had reduced familiar writing tells.

Claude’s closer match in word use did not make it the least ornate writer. In a separate assessment of metaphor-heavy or embellished prose, Opus 5.5 scored 37% below Opus 5, but remained about 1.6 times the human score; Astra was at about 1.2 times. Opus 5 assigned these ratings, so they may carry its biases.

Graphite distinguishes the number of overused expressions from the overall distribution of words. A model can lose familiar habits without moving closer to human word use, or move closer while retaining thousands of characteristic expressions. Neither measure alone establishes that its prose is better.

The researchers compared 10,000 articles from before ChatGPT’s launch with model-written articles based on summaries of them. They adjusted for text length and excluded infrequent expressions. The initial study covered nine model versions; the update added Opus 5.5.

All the generated articles used one fixed prompt. Different instructions or genres could produce different habits, and changes in human writing since 2022 could affect the comparison. Graphite advises against judging authorship from individual tells: these are patterns across samples, not probabilities that a particular sentence came from AI.