Persistent AI 'Tells' Reveal Model Quirks Despite Lab Efforts
A new study from Graphite identifies thousands of phrases that distinguish frontier LLM prose from human writing, even as developers attempt to scrub common giveaways.
While AI labs claim their latest models communicate more naturally, a study from marketing firm Graphite—as TechCrunch first reported—suggests that telltale linguistic habits remain deeply embedded in LLM output. By comparing a control group of 10,000 pre-ChatGPT articles against AI-rewritten versions, Graphite identified 13,000 phrases that appear at least twice as often in AI content as in human writing.
Reporting from TechCrunch highlights that while frontier models have largely eliminated early tells like the em-dash—with Gemini 3.1 Pro nearly removing it and Astra using it 88% less than humans—new patterns have emerged. For instance, Claude Opus 5.5 frequently uses the word "dependable" (23 times more often than humans) and is heavily prone to explaining why things matter, using the phrase "this matters" 116 times more often than human writers.
OpenAI's Astra exhibits different patterns, such as describing "another dimension" of a topic or using "corrective framing" (e.g., "not simply X"), which TechCrunch reports is over 100 times more common in Astra's prose than in human samples. Astra also tends to hedge claims with phrases like "may provide" or "can provide."
Greg Druck, Graphite's chief AI officer, told TechCrunch that while Claude models are trending closer to human word distribution, GPT models are moving further away. Despite claims from Anthropic regarding Opus 5.5 and OpenAI regarding GPT-6 versions Sol and Luna that their models offer more clarity and fewer odd phrases, Druck suggests that the scale of these models makes it difficult for labs to fully control these linguistic slips.