I was scrolling through a dataset at midnight when a pattern jumped out at me: the same phrase, again and again, punctuating machine prose. You feel a small chill—that moment when a clever trick stops being clever. I told myself I had to write about it before someone else monetized the discovery.
I’m going to walk you through what researchers found, what it means for anyone who writes or reads online, and why the old red flags—those em dashes and clunky rhythms—are no longer the hard evidence they once were. You’ll get names, numbers, and the one almost-comic phrase that keeps betraying certain models.
A startup just raised $9 million to detect AI writing
That funding round—reported by TechCrunch and tied to a company called Pangram—put a dollar figure on the demand for detectors: $9 million (€8.3 million). Investors are buying tools that promise to separate human text from machine output.
Graphite, a marketing firm, released a study that maps roughly 13,000 words and phrases it believes act as signals of AI authorship. You should treat the work as industry analysis, not definitive science, but the breadth of the list is striking.
How do AI detectors spot AI writing?
Detectors compare statistical fingerprints: phrase frequency, syntax patterns, and repeated phrasings across millions of tokens. Graphite’s approach was blunt—scan corpora, rank phrases by how much more often they appear in model outputs than in human samples, then flag the outliers.
That method surfaces anomalies—phrases that spike in a model’s text but are rare in human writing. It’s not forensic; it’s probabilistic. A high score suggests a model likely wrote the passage, not that it certainly did.
In a dataset, “this matters” appears 116 times more in Claude outputs
Graphite’s strongest callout was the phrase “this matters”, which turns up 116 times more often in Anthropic’s Claude Opus 5.5 outputs than in the human baseline. That single phrase has become a signature.
I’ve spent hours reading model outputs. These tells are a fingerprint the models leave behind. When a phrase like that repeats, it’s telling you about the model’s training and its safety or style prompts—small traces of its internal choreography.
Opus 5.5 has cut em dashes by 99% compared with Opus 5
Graphite also reports that Claude’s Opus 5.5 is shifting: it uses em dashes far less than the previous version. The model appears to be correcting for earlier giveaways.
That change matters because it shows the feedback loop at work. Engineers tweak models, detectors flag patterns, and models adapt to minimize those flags. The landscape of writing has become shifting sand—what marks AI today may vanish tomorrow, while new telltales appear.
Can AI remove the markers that give it away?
In short: sometimes. Claude’s outputs are reportedly getting closer to human norms, according to the study. But changes create new anomalies—if you remove one fingerprint, you leave another print elsewhere.
The study suggests Opus 5.5 uses fewer qualifiers and more superlatives, while OpenAI’s Astra tends toward different legalistic phrasing.
Astra uses “does not establish” 275 times more than Claude
Graphite found that OpenAI’s Astra favors the phrase “does not establish”, and it appears massively more often in Astra outputs than in Claude’s. The ratio Graphite cites is startling: Astra’s use of that phrase is roughly 275 times more common than Claude’s.
That divergence tells you two things. One: models inherit stylistic tendencies from their training data and safety guards. Two: detectors that generalize from one model to another will miss model-specific fingerprints unless they keep retraining on fresh samples.
What this means for writers, publishers, and readers
You don’t need to panic, but you should pay attention. Companies like Anthropic and OpenAI are iterating fast, and marketing groups like Graphite are hunting for signals investors will pay for. The market response—money flowing into detection startups—changes incentives for everyone creating or moderating text.
For editors, the take is simple: trust your judgment, use detectors as one input among many, and pay attention to style and context. For readers, remember that a phrase can be a clue, not a verdict.
I’ll keep watching the patterns, testing the claims, and calling out when a supposed tell is just the noise of style. Are you ready to bet on which tell will be the next to vanish?