AI writing tells: What the Graphite study found and why they shift

A new study from the marketing firm Graphite examined AI writing tells across frontier models, according to reporting by TechCrunch on 1 October 2026. Graphite identified 13,000 phrases that appeared at least twice as often in AI content as in human content, which is how the firm defines a tell. The broader finding is that familiar signals can disappear while other habits remain. Instead of relying on one famous word or punctuation mark, current models show recurring preferences in wording, contrast and sentence construction.

Graphite built its comparison around a human control group of 10,000 articles published before the release of ChatGPT. The firm then had different AI models rewrite those articles from summaries, a step intended to reduce source bias. Researchers compared how often particular words and phrases appeared in the AI versions and the human writing, while also examining broader patterns of sentence construction. This design allowed Graphite to look beyond isolated vocabulary and study repeated structures that appeared more frequently in model output.

For Claude Opus 5.5, Graphite found that the strongest tell was dependable, which appeared 23 times more often than in the human samples. The model now avoids the construction it is not X, it is Y, yet it still tends to frame ideas as more than an X, it is a Y. It also used this matters 116 times more often than human writing and why X matters 92 times more often. These findings show how a model can drop one conspicuous pattern while retaining other forms of emphasis and contrast that make its prose statistically distinctive.

OpenAI Astra showed a different set of habits. Graphite found that it often described another dimension of a topic and used hedging phrases such as may provide or can provide. Its biggest tell was corrective framing, where an idea is defined as not simply X or presented as an alternative rather than relying on X. Those constructions appeared more than 100 times as often as in human writing. Em dash use also changed sharply across models. Opus 5.5 used it 99 percent less than Opus 5, Astra used it 88 percent less than human samples, and Gemini 3.1 Pro almost eliminated it.

Graphite says the overall number of tells is mostly holding steady. Greg Druck, chief AI officer at Graphite, told TechCrunch that the tells are not decreasing. In his account, labs can remove widely noticed habits, but other patterns appear, and each model version develops its own quirks. Anthropic said with the Opus 5.5 release that the model communicates more naturally and that early users found its writing clearer and easier to follow. OpenAI made similar claims about greater clarity, less jargon and fewer odd turns of phrase in newer Sol and Luna versions.

For students, teachers and working writers, the useful lesson is not that one word or one punctuation mark proves where a passage came from. The Graphite results instead show that AI output can develop repeated rhetorical habits, and those habits can change between model versions. Reading with that in mind can encourage closer attention to how a passage makes its case. A reader can notice repeated contrast, routine hedging or familiar framing and then ask whether the reasoning is precise, whether the emphasis is justified and whether the wording actually fits the subject.

The same awareness can improve editing without turning revision into a hunt for forbidden words. A writer can review a draft for repeated structures, automatic contrasts, habitual hedges and stock transitions, then decide whether each one earns its place. Graphite found that well known tells can be reduced while new ones emerge, so the more durable editing habit is to judge sentences by clarity, purpose and variety rather than by a fixed blacklist. For students and professional writers alike, the practical value of AI writing tells is stronger attention to patterns in prose and more deliberate revision.