AI writing can become less conspicuous in one respect while retaining distinctive patterns elsewhere. That is the picture emerging from a Graphite study reported by TechCrunch, which compared human-written articles with model-generated versions of the same material. The marketing firm’s researchers found thousands of expressions that occurred substantially more often in AI output, even as familiar markers such as heavy em-dash use became less common.
The results concern patterns across samples, not a verdict on the authorship of any individual sentence. They offer a more specific account of how different models write than the idea that all generated prose has one recognisable voice. Graphite’s chief AI officer, Greg Druck, says each model version has its own habits.
Matched articles provide the comparison
Graphite began with 10,000 articles published before ChatGPT’s release as its human-written comparison group. Researchers then used summaries to have different models rewrite those articles. The intention was to compare corresponding material while reducing the influence of wording borrowed directly from the originals.
With samples representing the same underlying articles, the team examined how frequently words and phrases appeared in each group. It also looked at broader sentence constructions. Graphite defined a writing tell as a phrase occurring at least twice as often in AI content as in the human comparison material.
Using that definition, it found 13,000 phrases. The threshold describes a difference in frequency within the study’s samples; it does not mean that each phrase belongs exclusively to AI. A phrase can meet the threshold while still appearing in human writing, which is important when interpreting the results.
Druck told TechCrunch that Claude models appeared to be moving closer to the human word distribution over time, while GPT models were moving further away. That is Graphite’s interpretation of the comparison, not a finding that either family’s writing is uniformly better or worse for every purpose.
Different models favour different expressions
According to the study, Opus 5.5 used the word “dependable” 23 times more often than the human samples. Its tendency to explicitly announce significance was more pronounced: “this matters” appeared 116 times more frequently, while “why X matters” appeared 92 times more frequently.
The model also retained contrast-based constructions even while avoiding some better-known versions of them. Graphite’s account suggests that removing one conspicuous pattern does not necessarily remove the underlying preference for explaining a topic through comparison or correction.
OpenAI’s Astra showed a different profile. TechCrunch reports that it favoured references to another aspect of a subject and qualified descriptions of potential benefits. Graphite also found corrective framing—explaining what a topic should not be reduced to before presenting an alternative—more than 100 times as often in Astra output as in the human material.
These are reported tendencies in generated samples. They should not be treated as rules that every output follows. Nor does the frequency comparison establish that an ordinary writer using similar phrasing must have used one of the models.
Familiar tells decline while others persist
The punctuation results show how much an obvious marker can change. In Graphite’s samples, Opus 5.5 used em-dashes 99% less often than Opus 5. Astra used them 88% less often than the human comparison group, and Gemini 3.1 Pro had almost eliminated them.
Despite those reductions, Druck says the overall number of tells was broadly holding steady. Well-known habits disappeared while other patterns emerged. That finding complicates any simple account in which a model becomes progressively indistinguishable from human writing merely by avoiding a list of recognisable words or punctuation marks.
TechCrunch places the results alongside vendors’ claims about clearer, more natural writing. Anthropic described improvements to Opus 5.5’s communication, while OpenAI made similar claims for GPT-6 Sol and Luna. Those statements describe intended product improvements; Graphite’s measurements examine a different question about the distribution of language.
Druck doubts that laboratories can precisely control every stylistic tendency in large models. The study’s useful contribution is therefore not a definitive checklist for detecting AI. It is a set of model-specific comparisons showing that less conspicuous prose can still contain unusually persistent habits, and that those habits change between versions.
