Marketing firm Graphite conducted a study analyzing the writing habits of frontier AI models by comparing them against a control group of 10,000 articles written before the launch of ChatGPT. Researchers identified 13,000 phrases that appear at least twice as often in AI-generated content compared to human writing.
The study highlights specific linguistic patterns for different models. Claude Opus 5.5 frequently uses the word 'dependable' and the phrase 'this matters,' which appears 116 times more often than in human text. Meanwhile, OpenAI’s Astra model often employs 'corrective framing' and phrases like 'may provide' or 'can provide' to hedge its claims.
Data shows that AI labs have successfully reduced the use of em-dashes across recent model versions. Opus 5.5 uses this punctuation 99% less than its predecessor, while Gemini 3.1 Pro has nearly removed it entirely. Despite these adjustments, the total number of identifiable AI writing 'tells' remains stable as new patterns emerge.
Graphite’s chief AI officer Greg Druck suggests that labs struggle to fully control these stylistic quirks due to the massive scale of models with billions of parameters. While Anthropic and OpenAI have claimed improvements in natural communication and clarity in their recent releases, researchers argue that specific phrasing habits persist.
