Graphite: Opus 5.5 uses 'this matters' 116x more than humans as AI writing tells persist

A study from marketing firm Graphite found that frontier AI models still carry distinctive, measurable writing habits, even as labs have suppressed the best-known ones. Graphite identified 13,000 phrases that appear at least twice as often in AI-generated text as in human text, which is how the study defines a "tell." TechCrunch reported the findings on October 1, 2026.

Study design

Graphite used 10,000 articles published before ChatGPT's release as the human control group. Researchers then had each AI model rewrite those articles from summaries, an approach intended to limit source bias. That produced matched human and model samples for comparing word frequency and sentence construction.

Model-by-model findings

Claude Opus 5.5 (Anthropic):

• Its biggest tell is "dependable," which appears 23 times more often than in human samples.
• "This matters" appears 116 times more often than in human writing, and "why X matters" appears 92 times more often.
• The model now avoids the "it's not X, it's Y" construction but still uses "is more than an X, it's a Y."

Astra (OpenAI):

• It frequently describes "another dimension" of a topic and hedges with "may provide" or "can provide."
• Its biggest tell is what Graphite calls "corrective framing," such as "not simply X" or "rather than relying on X." Graphite says these constructions were more than 100 times more common in Astra's prose than in human writing.

Em-dashes largely gone

All three labs appear to have addressed em-dash overuse, though the comparison baselines differ across the reported figures:

• Opus 5.5 used the em-dash 99% less often than Opus 5.
• Astra used it 88% less than human samples.
• Gemini 3.1 Pro has "almost completely eliminated" it, per the report. No exact figure was given.

Tells shift rather than shrink

Graphite chief AI officer Greg Druck said the overall number of tells is mostly constant. "They are managing to remove the most well-known tells, but other ones pop up. And every model version has its own," he told TechCrunch.

The trend differs by vendor. "Claude models are actually getting closer to the human word distribution over time," Druck said. "And for the GPT models, it's getting further away." Graphite did not publish the underlying metric in the material TechCrunch reported.

Lab claims versus measurements

Anthropic said in the Opus 5.5 release that the model "communicates more naturally than prior models," and that early users found its writing "clearer and easier to follow." OpenAI made similar claims for the GPT-6 versions of Sol and Luna, saying users could expect "more clarity, less jargon, [and] fewer odd turns of phrase." These are company claims about perceived quality. Graphite's study measures something different: statistical divergence from human phrase frequency.

Druck is skeptical that labs can fully remove these patterns. "The labs are less able to control some of these things than you might expect," he said, noting that large models can only be tested against a finite set of checks, "and things slip through."

Caveats

Graphite is a marketing firm, not an independent lab. The source report does not say whether the study was peer reviewed or whether the full methodology and dataset are public. The rewrite-from-summary design also measures model behavior on one task type, so results may not generalize to chat, code comments, or creative writing. The report does not cover model versions beyond those named.

What this means

The study suggests that suppressing individual stylistic markers is a patch rather than a fix. Labs can train out a specific, widely publicized pattern like the em-dash, but the underlying tendency toward formulaic rhetoric reappears in other forms. Each model version carries its own signature.

For builders, that has two practical consequences. First, detection approaches built on a fixed word list will go stale with each release. Second, teams shipping AI-assisted content may need per-model, per-version style filters or editing passes rather than a general "AI-sounding" blocklist.

The Claude versus GPT divergence is the most notable claim, if it holds up under independent replication. It would indicate that the labs' approaches to writing-style tuning are producing different outcomes, and that marketing language about "natural" writing is not a reliable proxy for measured human-likeness.