Graphite identifies thousands of AI-writing tells across models

Graphite says the study compares human writing with ten AI models across 9,974 aligned topics.[1] The human articles predate ChatGPT, and various frontier models rewrote the same human-written articles from summaries to reduce source bias.[1][2] That setup lets Graphite compare model outputs against a shared set of human source material, rather than relying only on unrelated samples. This is a study of language patterns across a collection; its results should be read at that level, not as a test of a particular article or author.[4] Graphite is a marketing firm, a useful point of context when assessing the source of the research.[2]

Graphite defines a tell as a word, phrase or frame that appears at least twice as often in AI writing as in human writing.[1] Across its broader study, Graphite identified 13,000 phrases meeting that threshold.[2] For Opus 5.5 specifically, it found 2,548 tells, 4% fewer than the 2,666 it found in Opus 5.[1] Examples in Graphite’s list include “this matters,” with a value of 116x, “dependable” at 23x and “looking ahead the” at 40x.[1] These are frequency comparisons, not a guarantee that any one phrase identifies AI authorship.

Graphite says Opus 5.5 qualifies claims less often than Astra.[1] It also reports that Opus 5.5’s word-distribution divergence from human writing is 19% lower than Opus 5’s, while Astra’s divergence is 8% higher than GPT-5.6 Sol’s.[1] Opus 5.5 uses em dashes 99% less often than Opus 5, and Astra uses them 95% less often than GPT-5.[1] Graphite puts Opus 5.5’s mannered-prose score 37% below Opus 5’s, but still about 1.6 times the human score.[1] The comparisons point to different patterns across models, not one universal AI-writing style.

The study does not establish whether readers would mistake an individual AI-generated article for human work, or whether its content is accurate, original or useful.[4] A speaker associated with the findings also said they do not show a steady progression toward better or worse writing.[4] That limits what teams can conclude from a model’s score or a phrase appearing in a draft: the research measures differences in language patterns, not the quality or provenance of a single piece. Graphite says its tells data and raw article data are available to download for further examination.[1]