Graphite study finds persistent writing patterns in frontier AI models

A study by marketing firm Graphite found thousands of phrases appearing more frequently in AI-generated writing than in human articles, TechCrunch reported, even as newer models reduced familiar habits such as em-dash use.

Researchers used 10,000 articles published before ChatGPT’s release as a human control group. They asked AI models to rewrite articles from summaries, then compared word frequencies and sentence constructions across matched samples.

Graphite identified 13,000 phrases that appeared at least twice as often in AI samples as in human writing, its threshold for a distinguishing pattern.

Claude Opus 5.5 used the word “dependable” 23 times more frequently than the human samples. The phrase “this matters” appeared 116 times more often, while constructions explaining why something matters occurred 92 times more often, according to the study.

OpenAI’s Astra showed different patterns, including corrective framing and language presenting alternatives to an existing approach. Graphite found those constructions more than 100 times as frequently in Astra-generated text as in the human comparison group.

Em-dash use declined sharply. Opus 5.5 used the mark 99% less often than Opus 5, Astra used it 88% less than human samples, and Gemini 3.1 Pro nearly eliminated it, TechCrunch reported.

Graphite chief AI officer Greg Druck said Claude’s word distribution was moving closer to the human baseline over time, while GPT models were moving farther away. He said removing well-known patterns had not eliminated the broader tendency for models to develop distinctive phrasing.

Druck suggested that the size and complexity of these models may limit developers’ ability to control every writing habit. The reported comparisons describe aggregate patterns in Graphite’s samples.