As synthetic text fills online blogs, media publications, and corporate documentation, readers are looking for subtle style patterns that expose AI-generated prose. While early giveaways like overused punctuation marks and specific buzzwords have largely faded from recent software updates, new research shows that top artificial intelligence models still rely on distinct phrasing habits. A comprehensive style study conducted by marketing firm Graphite analyzed writing patterns across leading frontier models, revealing that every major system retains unique verbal habits.
Graphite’s research team evaluated thousands of synthetic articles against human-written content to identify specific words and sentence structures that appear far more frequently in generated text. The study flagged roughly 13,000 phrases that occur at least twice as often in synthetic content compared to human writing. Chief AI Officer Greg Druck noted that while newer Claude models are moving closer to natural human word distributions over time, recent GPT iterations are actually drifting further away from natural human writing habits.
Each major model displays a distinct set of style quirks. Claude Opus 5.5 relies heavily on the word “dependable,” using it 73 times more often than human writers. The model also repeatedly uses framing expressions like “this matters” or “why X matters,” which appear up to 116 times more frequently in synthetic prose. While Opus 5.5 successfully avoids negative contrast setups like “it’s not X, it’s Y,” it frequently resorts to alternative comparisons such as “more than an X, it’s a Y”.
OpenAI’s latest model, Astra, exhibits a completely different set of writing habits. Astra frequently uses the word “dimension” when describing abstract concepts and tends to qualify bold claims with cautious phrasing like “may provide” or “can provide”. Its most prominent style quirk involves corrective framing structures, where topics are defined as “not simply X” or presented as an alternative “rather than relying on X”. Graphite’s analysis found these corrective setups over 100 times more often in Astra’s generated text than in human writing.
On a positive note, AI labs successfully trained recent models to drastically reduce their reliance on overused punctuation. Claude Opus 5.5 uses em dashes 99% less often than previous model versions, while Astra uses them 88% less frequently than standard human baselines. Models like Gemini 3.1 Pro have almost entirely eliminated em dashes from their default writing output.
Despite these targeted punctuation fixes, the overall volume of writing quirks remains steady across successive model releases. As developers patch famous buzzwords, new habits emerge in updated model versions. AI research labs struggle to completely eliminate stylistic quirks because massive training datasets and complex parameter layers make fine-grained stylistic control difficult to enforce.
Understanding these recurring writing habits helps editors and content managers spot synthetic text before publication. While AI labs continue updating their models to produce clearer, more natural prose, subtle phrasing patterns continue to reveal the machine behind the screen. Spotting these persistent style habits allows readers to evaluate online content with a sharper, more critical eye.