The familiar giveaways of AI-generated prose may be fading, but the underlying pattern has not vanished. A study by marketing firm Graphite found 13,000 phrases that appeared at least twice as often in AI-written samples as in human text, according to TechCrunch. Instead of sharing one permanent vocabulary, current models appear to develop their own recurring words, rhetorical moves, and sentence structures.

Graphite built its comparison around 10,000 articles published before ChatGPT’s release, using them as a human-written control set. Researchers summarized those articles and asked multiple frontier models to rewrite them from the summaries, an approach intended to reduce differences caused by subject matter. They then compared matched human and machine samples for word frequency, phrases, and broader construction patterns.

Matched writing samples pass through a prism revealing recurring language patterns.
Researchers compared human articles with model rewrites produced from summaries of the same source material.

Claude Opus 5.5 showed some particularly strong preferences in Graphite’s results. The word “dependable” appeared 23 times more often than in the human samples. The phrase “this matters” appeared 116 times more frequently, while variations of “why X matters” occurred 92 times more often. The model had largely moved away from the familiar “it’s not X, it’s Y” pattern but continued using related contrast structures.

OpenAI’s Astra produced a different fingerprint. Graphite found that it frequently introduced “another dimension,” hedged benefits with constructions such as “may provide” or “can provide,” and relied heavily on what the firm called corrective framing. Patterns that defined a subject as something other than a simpler alternative appeared more than 100 times as often in Astra’s output as in human writing, TechCrunch reported.

Punctuation illustrates how quickly these signatures can change. Frontier labs appear to have reacted to public jokes about excessive em dashes: Opus 5.5 used them 99 percent less often than Opus 5 in Graphite’s samples. Astra used them 88 percent less often than human writers, and Gemini 3.1 Pro had almost eliminated them. Yet Graphite said the total number of detectable tendencies remained broadly steady as different habits replaced the famous ones.

Old punctuation clues fade as new rhetorical patterns form around evolving models.
Widely recognized tells such as em dashes declined sharply, but other model-specific habits took their place.

Greg Druck, Graphite’s chief AI officer, told TechCrunch that Claude’s word distribution has moved closer to human writing over time, while GPT models have moved farther away in the firm’s analysis. He suggested that labs may have less control over every stylistic behavior than users assume because large models contain billions of parameters and can be tested against only a finite set of cases.

The results describe statistical tendencies across matched corpora, not a definitive authorship test for any single passage. A human can naturally use a model-favored phrase, and a model can avoid it. The more practical conclusion is that simplistic rules—such as treating every em dash or one overused word as proof of AI—will age badly as model versions change.

For editors, teachers, and platforms, the study points toward a moving-target problem. A model can become more natural by suppressing its best-known quirks while remaining unusually repetitive in less obvious ways. Detecting AI writing, if attempted at all, therefore requires version-aware analysis of many signals and cautious interpretation—not a blacklist of fashionable words.