Study finds AI writing tells vary by model

A study by marketing firm Graphite found 13,000 phrases that appeared at least twice as often in AI-generated writing as in human-written samples. The firm defined such phrases as a “tell.”

To compare the writing, researchers used 10,000 articles published before ChatGPT’s release as a human control group. They had AI models rewrite those articles from summaries, aiming to reduce source bias, then compared words, phrases, and sentence patterns.

Graphite reported that Claude Opus 5.5 used “dependable” 23 times more often than human samples. The phrase “this matters” appeared 116 times more often, while “why X matters” appeared 92 times more often. Graphite also found that corrective framing phrases were more than 100 times more common in OpenAI’s Astra output than in human writing.

Some punctuation habits have changed: Graphite’s samples showed Opus 5.5 used em dashes 99% less often than Opus 5, and Astra used them 88% less often than human samples. Gemini 3.1 Pro had almost eliminated the mark, according to the study.

Graphite’s chief AI officer Greg Druck told TechCrunch that Claude models were moving closer to human word distribution over time, while GPT models were moving further away. He said familiar tells may disappear as other habits emerge, with each model version having its own patterns.