A new study from Graphite reveals that artificial intelligence models continue to use distinct writing habits.

A new study from the marketing firm Graphite reveals that artificial intelligence models still use distinct writing habits, despite dropping older clues like em-dashes. Researchers compared 10,000 human-written articles published before ChatGPT with AI-rewritten versions to spot unique model quirks. Graphite found 13,000 phrases that appeared at least twice as often in AI text than in human text.

Greg Druck, Graphite's chief AI officer, stated that Claude models grow closer to human word distribution over time, while GPT models move further away. For example, Anthropic's Claude Opus 5.5 relies heavily on phrases like "this matters" and the word "dependable." Meanwhile, OpenAI's Astra often uses corrective framing and phrases such as "another dimension." While labs successfully reduced em-dash usage, the overall number of tells remains steady as new ones replace old habits.

Druck notes that labs struggle to fully control these persistent patterns in giant models with billions of parameters.