New Study Uncovers Unique Writing Habits and Tells in Frontier AI Models
Research from Graphite reveals that frontier AI models still rely on thousands of distinct phrase tells, even as older habits like em-dashes fade away.
As artificially generated text becomes increasingly common across digital platforms, readers and researchers are constantly finding new ways to spot it. While early stylistic giveaways like the overuse of em-dashes and the word “delve” have largely disappeared, a recent study from marketing firm Graphite demonstrates that modern frontier models continue to rely on distinct linguistic habits.
To conduct the research, Graphite built a control group consisting of 10,000 articles published prior to the launch of ChatGPT. Researchers then asked various AI models to rewrite those articles based on summaries, neutralizing potential source bias. By comparing the resulting text to human-authored samples, the team could evaluate the frequency of specific words, phrases, and broader sentence structures.
Shifting Patterns Across Models
The study identified roughly 13,000 phrases that appeared at least twice as frequently in AI output compared to human writing. Interestingly, different labs are evolving in opposite directions. Graphite’s chief AI officer, Greg Druck, noted that Claude models are moving closer to natural human word distributions over time, whereas GPT models are actually drifting further away.
Specific models show clear individual quirks. For example, Anthropic's Opus 5.5 heavily favors the word “dependable,” appearing 23 times more often than in human samples. It also frequently uses phrases highlighting significance, with “this matters” showing up 116 times more frequently and “why X matters” appearing 92 times more often. Meanwhile, OpenAI’s Astra frequently turns to corrective framing, such as defining a subject as “not simply X” or “rather than relying on X,” a construction more than 100 times more common in Astra text than in human writing.
Developers can try top AI models cheaply through one API at https://apixoai.online to analyze their outputs and test different prompting strategies.
The Persistence of AI Tells
All major labs have successfully addressed some of the more obvious historical giveaways. Graphite's data shows that Opus 5.5 reduced its em-dash usage by 99% compared to previous iterations, Astra dropped usage by 88% compared to human levels, and Gemini 3.1 Pro has nearly wiped the punctuation mark from its prose entirely.
However, removing prominent tells has not reduced the overall volume of stylistic markers. Druck explained that as well-known tells get stamped out, new ones emerge to take their place, leaving the total count relatively stable across versions. Despite claims from companies like Anthropic and OpenAI regarding more natural communication and reduced jargon in recent releases, controlling these subtle linguistic patterns remains a challenge for massive multi-parameter systems.
What it means for developers
For software engineers building applications powered by generative text, these findings highlight the ongoing challenge of achieving truly natural-sounding outputs. Because large language models possess billions of parameters and finite testing cycles, distinct stylistic quirks and repetitive phrasing will likely continue to slip through. Developers must remain mindful of these underlying patterns when crafting system prompts or fine-tuning models for production applications, especially if natural human tone is a core requirement for their user base.
Source: Opus 5.5 loves to tell you ‘this matters’ (and other AI writing tells) — TechCrunch AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
