Researchers have cataloged nearly 13,000 distinct words, phrases and syntactic patterns that appear at least twice as often in text produced by leading AI models as in human writing. The findings, released this week by marketing firm Graphite, signal that even as obvious giveaways fade, machine-generated prose continues to carry measurable signatures.

The study examined 10,000 human articles pulled from Common Crawl and published before November 2022. Summaries of those pieces, generated by GPT-4.1, were then fed to nine different large language models. Each model expanded the summaries into full articles, creating a parallel corpus of 90,000 AI-generated texts on identical topics. From that data, analysts isolated 12,877 linguistic tells.

One phrase stands out. “This matters” appears 116 times more frequently in output from Anthropic’s Claude Opus 5.5 than in the human baseline. Related constructions such as “why X matters” show up 92 times more often in the same model. Claude also favors “dependable,” using it 23 to 26 times as often as humans do. Gizmodo highlighted the pattern, noting that the days of fixating on em dashes have ended.

OpenAI’s latest flagship, referred to in the study as Astra or GPT-6 Astra, exhibits its own habits. The model leans on corrective and negative framing. Phrases such as “not simply,” “rather than,” and “does not establish” proliferate. One analysis found “does not establish” 275 times more common in Astra text than in Claude outputs. The model also describes “another dimension” of topics and hedges with “may provide” or “can provide.”

Gemini 3.1 Pro, from Google, shows different tendencies. It favors formal transitions like “furthermore” and has almost eliminated em dashes from its style. Across the models tested, em dash usage has plummeted. Claude Opus 5.5 deploys them 99 percent less than its predecessor. Astra uses them 88 percent less than human samples. The once-reliable punctuation marker no longer serves as a reliable indicator.

Yet the total volume of tells has not shrunk. Graphite’s chief AI officer Greg Druck observed that while labs remove well-known markers, new ones surface to take their place. Astra, for instance, carries 48 percent more tells than its predecessor GPT-5.6 Sol. Only 45 percent of the tells overlap between the two versions. Tells shift. They do not disappear.

Sixty-five percent of the identified tells prove unique to a single model family. Claude prefers superlative phrasing such as “single most” and contrastive structures like “less like a _ and more like.” GPT models have moved away from earlier “salesy” vocabulary such as “unlock” and “streamline,” replacing it with negation-heavy definitions of concepts by what they are not. Gemini leans on structured, academic transitions.

The convergence surprises. All nine models now write more like one another than like people. A shared “AI-ese” has taken shape, characterized by longer sentences, reduced punctuation variety, lower burstiness in sentence length, and repetitive logical scaffolding. Human prose displays greater stylistic range shaped by personal experience and intent. AI output clusters tightly.

These patterns echo earlier academic work. A 2025 study in biomedical literature, reported by The New York Times, tracked excess use of words such as “delves,” “pivotal,” “showcasing” and “intricate” in paper abstracts after ChatGPT’s release. Researchers estimated at least 13.5 percent of 2024 biomedical abstracts showed signs of AI assistance, with higher rates in computer science reaching 22.5 percent according to a related analysis published in Nature Human Behaviour.

More recent examinations reinforce the persistence. A University College Cork study released in late September, covered by the university’s news site, applied literary stylometry to creative prose. AI stories formed tight clusters with uniform patterns. Human stories scattered widely. “Even when ChatGPT tries to sound human, its writing still carries a detectable fingerprint,” one researcher noted.

Graphite’s methodology improves on prior efforts by holding topic constant across human and AI versions. Previous frequency-based approaches compared pre- and post-2022 corpora without that control, risking topic drift as a confounding factor. The new parallel design isolates stylistic differences more cleanly.

But questions remain about real-world application. The study used prompted expansion of summaries rather than open-ended generation or editing of human drafts. Separate research on arXiv, including a September 2026 paper titled “AI Writers Have a Consistent Stylometric Footprint, but AI Editors Do Not,” shows that AI editing produces a different trace than pure generation. Lexical diversity and entropy behave differently. Detection systems tuned only on generated text may miss hybrid content.

Industry insiders already wrestle with these signals. Publishers, universities and platforms deploy detectors that combine perplexity, burstiness, n-gram frequency and supervised classifiers. Many now report false positives on non-native English writing or highly edited text. A professor at one university recently flagged 60 AI-written essays in a single assignment, according to a Times of India report from October 2, only after combining hidden prompt traps with improved detection tools.

Model makers have responded. Frontier labs tune systems to reduce conspicuous habits. Em dash suppression offers one clear example. Yet the Graphite data suggests the underlying statistical divergence remains. Entropy, lexical diversity, function word distribution and discourse-level choices such as moral ambiguity in plots or temporal complexity continue to separate human from machine.

A parallel arXiv study from August 2026 on narrative features in fiction found that discourse-level signals alone achieve over 93 percent accuracy in distinguishing human from AI stories. AI outputs favor tidy single-track plots and over-explain themes. Humans introduce greater moral ambiguity and nonlinear time. These patterns hold even without surface stylistic cues.

So the signals evolve. Detectors must keep pace. Watermarking schemes from providers such as Google and OpenAI offer one path when models cooperate. Statistical and neural detectors handle the rest. But adversarial paraphrasing, style transfer and human editing erode performance. No single method suffices.

Graphite frames its work as practical guidance for marketers and content teams. The implications stretch further. Academic integrity, journalistic standards, legal evidence and creative attribution all hinge on reliable provenance. As models grow more fluent, the tells grow subtler. They have not vanished.

Claude still says “this matters” far too often. Astra corrects and negates in distinctive rhythms. Gemini transitions with formality that feels machined. Thousands of such markers accumulate. The aggregate fingerprint persists. And researchers keep counting.