Graphite catalogs AI writing tells across ten models
Graphite had ten models rewrite 10,000 pre-ChatGPT articles and found 13,000 phrases that appear at least twice as often in AI copies; Opus 5.5 moved closer to human writing while GPT-6 Astra drifted further away.
Hunting em dashes won't catch Claude anymore. Opus 5.5 uses about one per 67,000 words. Graphite, a marketing firm, took 10,000 pre-ChatGPT articles and had ten AI models rewrite them from summaries. Then it counted every phrase that shows up at least twice as often in the copies and found 13,000. Graphite calls these tells. Of those, 2,548 appear in Opus 5.5. Its vocabulary got closer to human writing than Opus 5's. GPT-6 Astra went the other way and drifted further from it. To check whether a text was made with AI, search for these: - "incredibly" (Gemini 3.1 Pro, 18x) - "dependable" (Opus 5.5, 23x) - "why ... matters" (Opus 5.5, 92x) - "is more than a ..., it's a ..." (Opus 5.5, 98x) - "this matters" (Opus 5.5, 116x) - "rather than relying" (GPT-6 Astra, 187x) - "the ... is not simply" (Astra, 576x) I wonder how soon a model's writing will be impossible to tell apart from a person's.