Opus 5.5 loves to tell you ‘this matters’ (and other AI writing tells)
techcrunch ·
A new study by marketing firm Graphite reveals that while early AI writing clichés like em-dashes have been largely eliminated, frontier language models still exhibit distinct linguistic habits. Researchers analyzed 10,000 pre-ChatGPT articles and compared them against rewrites generated by various AI models to identify 13,000 phrases that appear significantly more often in AI content. Claude Opus 5.5 was found to favor the word "dependable" and frequently uses constructions such as "this matters" or "is more than an X, it’s a Y." In contrast, OpenAI’s Astra model tends to hedge claims with phrases like "may provide" and employs "corrective framing" structures over 100 times more frequently than human writers. Despite labs' efforts to make prose feel more natural, these tells remain persistent because each model version develops its own unique quirks. The study suggests that developers may struggle to fully control these patterns due to the immense scale of billions of parameters in modern models. Consequently, while specific tells change, the overall volume of detectable AI markers remains steady across different versions.
techcrunch