Claude Opus 5.5 uses the phrase "this matters" 116 times more often than human writers, according to a new study from marketing firm Graphite released today. The research identified 13,000 phrases that appear at least twice as frequently in AI-generated content compared to human writing, revealing what the firm calls "tells" that distinguish machine prose from human prose. Despite efforts by AI labs to eliminate obvious giveaways like em-dashes and the word "delve," models continue to rely on distinctive patterns and constructions that betray their origin.
The study analyzed 10,000 articles published before ChatGPT's release as a human control group, then had various AI models rewrite those articles from summaries to create matching samples. Claude Opus 5.5's most pronounced tell is the word "dependable," which appears 23 times more frequently than in human samples, while the phrase "why X matters" occurs 92 times more often. OpenAI's Astra model favors describing "another dimension" and hedges claims by saying an action "may provide" or "can provide" benefits. The model's strongest tell is what Graphite calls "corrective framing," using constructions like "not simply X" or "rather than relying on X" more than 100 times as often as human writers. Meanwhile, frontier labs have successfully addressed em-dash overuse: Opus 5.5 reduced its usage by 99% compared to Opus 5, Astra employs it 88% less than humans, and Gemini 3.1 Pro has nearly eliminated the punctuation mark entirely.
"It turns out that Claude models are actually getting closer to the human word distribution over time," Graphite's chief AI officer Greg Druck told TechCrunch. "And for the GPT models, it's getting further away." According to the research, while individual tells change between model versions, the overall number remains mostly stable. "It's not like the tells are decreasing," Druck said. "They are managing to remove the most well-known tells, but other ones pop up. And every model version has its own."
The persistence of telltale patterns is surprising given that both Anthropic and OpenAI have emphasized natural writing in recent releases. Anthropic claimed Opus 5.5 "communicates more naturally than prior models" with clearer, easier-to-follow writing, while OpenAI promised its GPT-6 versions would deliver "more clarity, less jargon, [and] fewer odd turns of phrase." But Druck remains skeptical about complete elimination of these linguistic fingerprints. He suggests the labs have less control over these patterns than might be expected, explaining that giant models with billions of parameters can only undergo a finite number of tests, allowing tells to slip through. The findings suggest contrast-heavy sentence constructions remain a fundamental feature of how frontier models generate prose, with each version developing its own unique quirks even as older tells are stamped out.
The study indicates that as AI-generated content becomes ubiquitous, humans will continue to have methods for detecting it, though the specific tells will evolve with each new model release. While labs can eliminate well-known patterns once they're identified, replacement tells emerge to take their place, creating a persistent gap between machine and human writing styles. For organizations concerned about AI-generated content, the sheer volume of distinguishing phrases offers multiple detection pathways even as individual markers change. The tension between labs' stated goal of human-like prose and the structural limitations of how these models actually compose text may prove difficult to fully resolve, leaving detection an ongoing challenge rather than a settled question.