Study Finds Thousands of Writing Tells Still Separate AI Models From Humans

AI models may be getting better at avoiding familiar writing habits, but researchers say newer systems still leave behind distinctive patterns in their prose.

A study from marketing firm Graphite compared human-written articles with text generated by frontier AI models and found thousands of words, phrases and sentence constructions that appeared more frequently in AI-generated writing.

The researchers started with 10,000 articles published before ChatGPT’s release and asked different models to recreate them from summaries. This allowed Graphite to compare model output against human writing while reducing the influence of the original wording.

Opus 5.5 has its own recurring phrases

According to Graphite, Claude Opus 5.5 used the word “dependable” 23 times as often as the human samples. The phrase “this matters” appeared 116 times more often, while constructions built around “why X matters” appeared 92 times more frequently.

Graphite also found that Opus 5.5 had moved away from some previously recognisable AI constructions, while developing other recurring patterns. Anthropic said when announcing Claude Opus 5.5 that the model communicates more naturally than earlier versions and produces clearer writing.

Different models leave different signals

OpenAI’s Astra showed a different set of tendencies in Graphite’s analysis. The model frequently used phrases such as “another dimension” and hedged claims with constructions including “may provide” and “can provide.”

The researchers also identified what they called corrective framing, where a model defines something as “not simply” one thing or presents an alternative “rather than relying on” something else. Graphite found such constructions were far more common in Astra-generated samples than in human writing.

OpenAI has also emphasised writing quality in newer models. When introducing GPT-6 Sol and Luna, the company said users should expect greater clarity, less jargon and fewer unusual turns of phrase.

Old AI tells are fading, but new ones keep appearing

One of the clearest changes across frontier models involves em dashes, once widely seen as a recognisable sign of AI-generated prose. Graphite found Opus 5.5 used them 99% less often than Opus 5, while Astra used them 88% less frequently than human writers in the study. Gemini 3.1 Pro had nearly eliminated them from its output.

But removing familiar signals has not eliminated model-specific writing habits. Graphite found thousands of phrases that appeared at least twice as frequently in AI-generated text as in human samples.

The findings suggest that AI writing patterns are evolving rather than disappearing. Individual words and constructions may change between model generations, making any single “AI tell” unreliable on its own. At the same time, broader statistical patterns remain visible when you compare large volumes of text.