The enduring fingerprints of AI writing

Graphite analyzed 10,000 pre-ChatGPT articles against AI-generated rewrites to identify 13,000 phrases that appear at least twice as often in machine output. While models have largely abandoned the em-dash, they remain tethered to specific structural quirks. Claude Opus 5.5, for instance, exhibits a peculiar obsession with emphasizing significance, using the phrase “this matters” 116 times more frequently than human authors. Its reliance on the word “dependable” and specific comparative sentence structures remains a primary identifier of its generated content.

OpenAI’s Astra utilizes an entirely different vocabulary of tells. The model frequently leans on “corrective framing,” often defining topics as “not simply X” or suggesting alternatives “rather than relying on X.” These constructions appear over 100 times more often in Astra’s prose than in human samples. According to Greg Druck, Graphite’s chief AI officer, models are not becoming less predictable; they are simply shifting their patterns. While Anthropic and OpenAI have touted more natural communication in recent releases, Druck remains skeptical that labs can fully scrub these linguistic biases from systems with billions of parameters. As these models grow, the specific tells may change, but the underlying tendency to fall back on predictable, repetitive phrasing persists.