Researchers at marketing firm Graphite analyzed writing habits of frontier AI models and found persistent telltale phrases and constructions, despite labs' efforts to make AI writing more human-like. Claude Opus 5.5 overuses phrases like "this matters" (116x more than humans) and "dependable," while OpenAI's Astra favors "another dimension" and corrective framings. Though models have largely eliminated obvious tells like em-dashes, new quirks emerge with each version, suggesting labs have limited control over these deeply embedded linguistic patterns.