Graphite study finds AI writing habits persist as familiar tells fade
An analysis of 10,000 pre-ChatGPT articles and AI rewrites found thousands of phrases more common in model-generated prose, even as em-dash use fell.
AI models may be dropping familiar writing habits, but Graphite says distinctive patterns remain. Its analysis identified 13,000 phrases at least twice as common in AI-generated samples as in human writing.
As reported by TechCrunch, Graphite compared 10,000 articles published before ChatGPT with AI rewrites produced from summaries. The approach was intended to reduce the influence of the original wording when comparing human and model output.
In Graphite’s results, Claude Opus 5.5 used “dependable” 23 times more often than human writers. The phrase “this matters” appeared 116 times more often, while “why X matters” was 92 times more common.
The study also identified patterns that varied by model. Opus 5.5 favoured constructions such as “more than an X, it’s a Y”, while OpenAI’s Astra showed more “corrective framing”, including phrases such as “not simply X” and “rather than relying on X”.
Em-dash use fell sharply in Graphite’s samples across several models. The company said Opus 5.5 used the punctuation 99 per cent less often than Opus 5, while Astra used it 88 per cent less often than human samples. Despite those changes, Graphite said the overall number of stylistic tells remained broadly steady.
The findings are based on rewrites of articles and do not establish how frequently these patterns appear across all AI writing. The supplied report does not include the full study or independent validation.
