As LLM-produced text becomes increasingly common, people are seeking ways to identify it. Researchers have found that although initial indicators have disappeared, AI models still exhibit specific patterns in their writing.
A recent study by the marketing firm Graphite analyzed the writing trends of various advanced AI models, identifying unique preferences for certain words and phrases. Even though previous markers, such as reliance on em-dashes, have been reduced, these models continue to show a tendency for contrast-rich expressions. Notably, the study uncovered 13,000 phrases that appeared at least twice as frequently in AI-generated content compared to human-written text, which they categorized as “tells.”
Graphite’s chief AI officer, Greg Druck, shared insights with TechCrunch, stating that “Claude models are inching closer to human-like word distributions over time, whereas GPT models seem to be diverging.”
To analyze AI writing effectively, the study employed a comprehensive design. Graphite compiled a collection of 10,000 articles created before the launch of ChatGPT as a control group for human writing. Then, various AI models were tasked with rewriting these articles from summaries to minimize bias. This approach enabled a thorough comparison of word usage and sentence structures between human and AI outputs.
The findings indicated that the word “dependable” emerged as a prominent feature in Claude Opus 5.5’s writing, appearing 23 times more than in human samples. Although this model has moved away from the phrase “it’s not X, it’s Y,” it often frames expressions like “it is more than an X, it’s a Y.”
Notably, Opus frequently emphasizes importance, using “this matters” 116 times more than human writers, and the phrase “why X matters” is also used 92 times more often.
In contrast, OpenAI’s Astra displays a different set of characteristics. This model tends to reference “another dimension” of its subjects and often mitigates its assertions, suggesting that something “may provide” or “can provide” benefits. One of Astra’s defining features is what Graphite describes as “corrective framing,” where topics are qualified with phrases like “not simply X” or “rather than relying on X,” which appeared over 100 times more frequently in Astra’s texts than in human writing.
Interestingly, all major AI labs seem to acknowledge the previous overuse of em-dashes. According to Graphite’s analysis, Opus 5.5 has reduced its use of this punctuation by 99%, while Astra has cut back by 88%, and Gemini 3.1 Pro has nearly eliminated it altogether.
Despite these changes, Graphite noted that the overall number of tells remains relatively stable. “They may be eliminating the more prominent tells,” Druck stated, “but new ones keep emerging, with each model exhibiting distinct characteristics.”
The persistence of these markers is intriguing, particularly given the labs’ emphasis on achieving natural-sounding writing styles. Anthropic pointed out that the Opus 5.5 model “communicates more naturally” than its predecessors, reporting that early adopters found its output clearer and easier to digest.
OpenAI has voiced similar sentiments regarding its GPT-6 models, promising users greater clarity, reduced jargon, and fewer unusual phrasings.
However, Druck voiced skepticism about the labs’ ability to fully eradicate identifiable phrases and structures. “My overarching theory is that the labs might not have as much control over these aspects as one might think,” he remarked. “These are massive models with billions of parameters, and they have a limited number of tests they can conduct, allowing certain elements to slip through.”
