Study Identifies AI Writing Tells in Claude Opus 5.5

New research conducted by Graphite, a growth agency specializing in AI-driven content, has identified specific linguistic patterns that can indicate writing generated by artificial intelligence, particularly focusing on Anthropic's latest language model, Claude Opus 5.5. The study compared 10,000 articles generated by Opus 5.5 with 10,000 human-written articles on identical topics to pinpoint "tells" – words, phrases, and patterns that appear at least twice as frequently in AI writing compared to human writing. While Opus 5.5 exhibits fewer tells than its predecessors, it still contains over 2,500 distinct markers. The research aims to help writers avoid these AI-specific linguistic habits to produce content that is indistinguishable from human authorship.

The study categorized these tells into single words, two-word phrases, and three-word phrases, revealing distinct tendencies in Opus 5.5's output. Among the most frequent single-word tells, "dependable" appeared 23 times more often than in human writing, followed by "clearer" (14 times), "matters" (13 times), "quietly" (11 times), and "practical" (11 times). These words, while common, are used with significantly higher frequency by the AI model. The research suggests that the AI's propensity to use these words might stem from its training data and its objective to convey information in a seemingly straightforward and reliable manner.

When examining two-word phrases, the AI's tendency to break down complex concepts becomes evident. Phrases such as "matters because" were found to be 146 times more prevalent in Opus 5.5's writing than in human text. Other notable two-word tells include "especially valuable" (136 times more frequent), "this matters" (116 times more frequent), "article explains" (91 times more frequent), and "builds trust" (83 times more frequent). These phrases often serve to connect ideas or introduce explanations, a function that Opus 5.5 performs with a distinct statistical bias.

The most striking differences were observed in the three-word phrase category, where the identified phrases appeared "infinitely more times" in AI writing than in human-written content. This means that while Opus 5.5 utilized these phrases hundreds of times, human writers did not employ them at all in the analyzed dataset. The top three-word tells identified in the study include "helps to understand" and "article expla" (the latter being an incomplete phrase likely cut off in the source material, but indicative of a pattern). These findings underscore the subtle yet measurable linguistic fingerprints left by advanced AI models, even as they strive for more natural and human-like communication.