Opus 5.5 loves to tell you ‘this matters’ (and other AI writing tells)

As artificial intelligence-generated writing becomes increasingly prevalent, researchers are seeking methods to distinguish AI prose from human writing. A recent study by marketing firm Graphite examined the writing habits of frontier models, identifying patterns in the words and phrases they favor. While early indicators such as em-dashes have disappeared, models still exhibit telltale habits.

Graphite discovered 13,000 phrases that were at least twice as common in AI-generated content compared to human-generated content, which they define as "tells."

Graphite conducted a study using a corpus of 10,000 articles published before the release of ChatGPT, serving as the control group for human-generated content. The researchers then had various AI models rewrite the articles from summaries to minimize source bias. By comparing the frequency of certain words and phrases in AI writing to human writing, they could analyze broader patterns in sentence construction.

One notable tell in AI writing is the frequent use of the word "dependable," which appeared 23 times more often in Claude Opus 5.5's writing than in human samples. While Opus 5.5 has moved away from the "it's not X, it's Y" sentence structure, it still tends to describe things as "more than X, it's a Y." Additionally, Opus 5.5 often emphasizes the importance of something by stating "this matters," a phrase that occurred 116 times more often than in human writing.

OpenAI's Astra model has its own set of telltale signs, including describing "another dimension" of a topic and hedging claims by saying an action "may provide" or "can provide" a benefit. Its biggest tell is the "corrective framing," where a topic is defined as "not simply X" or offered as an alternative to X, appearing over 100 times more often in Astra-generated prose than in human writing.

The research reveals that frontier labs have made efforts to address the overuse of em-dashes in their models. Opus 5.5 now uses em-dashes 99% less frequently than Opus 5.0, and Astra uses them 88% less than human samples. Gemini 3.1 Pro has also nearly eliminated em-dashes from its writing. However, individual tells may change, but the overall number remains relatively stable, according to Graphite.

The firm's chief AI officer, Greg Druck, expressed skepticism about the labs' ability to completely eradicate telltale constructions or phrases, stating that these models' complexity makes it challenging to control all aspects of their writing.