Research Identifies AI Writing Tells to Avoid

Writers aiming to produce content indistinguishable from human work should be aware of specific linguistic patterns identified in a new study by Graphite. The research focused on Anthropic's Claude Opus 5.5 model, highlighting phrases and word choices that are significantly more common in AI-generated text compared to human writing.

The study analyzed 10,000 articles created by Opus 5.5 alongside 10,000 articles written by humans on similar topics. Researchers defined "tells" as words, phrases, or patterns occurring at least twice as often in AI output. While Opus 5.5 shows fewer tells than previous versions, it still exhibits distinct characteristics.

Particularly notable are three-word phrases that appeared exclusively or overwhelmingly in AI-generated content. Examples include "helps to understand," "article explains what," and "this matters because." These phrases were found to be used hundreds of times in AI text but rarely, if ever, in human writing.

Single words like "dependable," "clearer," "matters," "quietly," and "practical" were also flagged as more frequent in AI output. Two-word phrases such as "matters because," "especially valuable," and "this matters" also emerged as common AI indicators.

Beyond specific phrases, stylistic elements like greater lexical diversity, longer words, and increased comma usage were observed more in AI text. Although the overuse of em dashes, a previous AI tell, has decreased significantly, AI models still tend to employ more "mannered prose"—flowery language and metaphors—than human writers, despite a reduction in its use compared to earlier AI versions.