As AI-generated text becomes part of everyday digital communication, researchers are identifying the subtle language habits that can still separate machine-written prose from human writing.
A new analysis by Graphite examined how leading language models rewrite the same source material. The research compared AI outputs with 10,000 articles published before ChatGPT's release, creating a benchmark for naturally human editorial patterns.
Thousands of recurring language signals
Graphite identified around 13,000 phrases that appeared at least twice as often in AI-generated text as in human-written samples. Rather than relying on once-familiar clues such as frequent em dashes or words like "delve," the study found that newer models develop more nuanced and model-specific habits.
Claude Opus 5.5, for example, uses "dependable" far more frequently than human writers in the sample set. It also repeatedly highlights significance through phrases such as "this matters" and "why X matters." Meanwhile, OpenAI Astra often uses corrective structures including "not simply X" and "rather than relying on X," alongside cautious expressions such as "may provide."
Writing styles continue to evolve
The analysis suggests AI developers have successfully reduced some highly visible stylistic signals. Em-dash use has fallen sharply across several models, while Gemini 3.1 Pro reportedly uses the punctuation mark only rarely.
Yet reducing one recognizable pattern does not necessarily make AI text indistinguishable from human work. As models are updated, older habits may fade while new phrasing preferences emerge. Graphite's findings indicate that AI writing is becoming more polished, but it still reflects the statistical patterns shaped by each model's training and design.
For editors, educators and content teams, these insights could support more informed use of AI tools and encourage stronger human review. The future of AI writing may be defined not only by fluency, but by systems that adapt more naturally to context, voice and audience expectations.
