Opus 5.5 Can't Stop Saying "This Matters" — and Other Dead Giveaways of AI Writing

Meta description: Graphite study finds persistent AI writing tells in Claude Opus 5.5, Astra and Gemini 3.1 Pro despite reduced em-dash use across frontier models.

Tags: AI writing, Graphite, Claude Opus 5.5, Astra, Gemini 3.1 Pro


AI writing tells have moved from obvious giveaways to model-specific fingerprints. A study from marketing firm Graphite found 13,000 phrases that appear at least twice as often in AI-generated content as in human writing, showing that frontier models still leave detectable patterns even after widely known signs such as em-dashes and "delve" have faded.

Background: Early detection habits became unreliable

The first public discussion about spotting machine text focused on shallow stylistic markers. Em-dashes and the verb "delve" became shorthand for AI-generated prose, but the study treats those signals as outdated. Graphite’s definition of a tell is broader: a phrase must occur at least twice as frequently in AI content as in human content. That threshold produced 13,000 candidate phrases, making clear that the problem is not a handful of words but a large, changing set of habits.

Development: Graphite built a matched comparison corpus

To measure the problem at scale, Graphite assembled 10,000 articles published before the release of ChatGPT and used them as the human-generated control group. The researchers then asked different AI models to rewrite those articles from summaries, a design intended to reduce source bias. The resulting paired samples allowed comparisons of word frequency, phrase frequency and sentence construction across humans and each frontier model.

Latest findings: Claude, Astra and Gemini show different quirks

The broader pattern is contrast-heavy constructions, but each model’s details differ. Graphite said Claude Opus 5.5’s biggest tell is the word "dependable," which appears 23 times more often than in human samples. The model has moved away from the "it’s not X, it’s Y" construction, but still leans on "is more than an X, it’s a Y." Opus also emphasizes importance: "this matters" appears 116 times more often, while "why X matters" appears 92 times more often.

OpenAI’s Astra has a different profile. It favors phrases such as "another dimension," hedges with "may provide" or "can provide," and shows its strongest tell in what Graphite calls "corrective framing." That pattern defines a topic as "not simply X" or frames it as an alternative, "rather than relying on X." Those constructions were more than 100 times more common in Astra-generated prose than in human writing.

The study also found progress on one famous signal. In Graphite’s samples, Opus 5.5 used em-dashes 99% less often than Opus 5. Astra now uses them 88% less than human samples, while Gemini 3.1 Pro has almost completely eliminated the punctuation from its writing.

Current stage: Tells persist even as known patterns fade

Graphite’s chief AI officer Greg Druck told TechCrunch that Claude models are getting closer to human word distribution over time, while GPT models are getting further away. He also said the overall number of tells is mostly holding steady. Models remove the most well-known patterns, but new ones appear, and every model version has its own quirks.

That persistence is notable because labs have promoted more natural prose. Anthropic’s Opus 5.5 release said the model communicates more naturally than prior models and that early users found its writing clearer and easier to follow. OpenAI’s GPT-6 versions of Sol and Luna were introduced with claims of more clarity, less jargon and fewer odd turns of phrase.

The current stage is therefore not a simple cat-and-mouse game with one fixed checklist. AI writing tells have become broader, model-specific and adaptive. Future detection work will likely need to track phrase frequency and sentence patterns across model versions, not just search for familiar words such as em-dashes or "delve."