Anthropic's Claude Opus 5.5 has very nearly stopped using the em dash, the punctuation mark that became shorthand for machine-written prose. A follow-up study published October 1 by the research firm Graphite measured the model at 0.015 em dashes per 1,000 words, against 2.92 for Opus 5 β a 99% reduction. The rest of the model's fingerprint survived almost untouched.
Key takeaways
- Opus 5.5 uses 0.015 em dashes per 1,000 words versus 2.92 for Opus 5, a 99% drop.
- Graphite still counted 2,548 distinct tells in Opus 5.5 output, only 4% fewer than the 2,666 it found in Opus 5.
- The phrase "this matters" appears 116 times more often in Opus 5.5 prose than in a corpus of articles written before ChatGPT shipped.
Graphite defines a tell as a word, phrase or sentence frame that a model uses at least twice as often as human writers do. By that measure Opus 5.5 carries 2,548 of them, down from 2,666 in Opus 5. Deleting the single most recognizable habit in AI prose moved the total by four percent.
Which phrases give Opus 5.5 away
The study's standout individual word is "dependable," which Opus 5.5 reaches for 23 times more often than the human control group. Its strongest structural habit is a compulsion to flag significance: "this matters" shows up 116 times more often than in human text, and the frame "why _ matters" 92 times more often.
The model has also abandoned the "it's not X, it's Y" construction that defined earlier Claude releases. Graphite chief AI officer Gregory Druck told TechCrunch that a close relative took its place, with the model now preferring to say something is more than an X, it's a Y.
How Graphite built the corpus
The method comes from Graphite's September 16 report, which assembled 10,000 articles published before ChatGPT's release as a human baseline, then had nine frontier large language models rewrite each one from a summary. That produced 90,000 machine articles on topically matched subjects, which strips out the topic bias that confounds most style comparisons.
Comparing length-normalized usage rates surfaced 12,877 unique tells across the nine models, with each model carrying between 2,355 and 3,746. Roughly 65% belonged to a single model family. Other measures separated the two groups just as cleanly: AI articles averaged 5.2 to 5.7 characters per word against 4.9 for humans, and their sentence lengths varied far less.
Where rival models leave their marks
OpenAI's GPT-6 Astra leans on corrective framing, defining a subject by what it is not with constructions such as "not simply X" or "rather than relying on X." Graphite clocked that habit at about 12 times the human rate. Google's Gemini 3.1 Pro instead favors intensifiers such as "absolutely" and "highly" at 13 times the human rate, alongside formal transitions like "furthermore."
On Graphite's mannered-prose score, Opus 5 and Gemini 3.1 Pro sat near 2.5 times the human baseline while Astra measured 1.2 times. Opus 5.5 pulled its own score down to 10.57 from 16.75, a 37% cut that still leaves it about 1.6 times human. Its overall word distribution drifted toward human text as well, with unigram divergence falling from 0.064 to 0.052.
Why the tells keep coming back
Druck told TechCrunch that Claude's word distribution has edged closer to human writing across versions while OpenAI's models have moved further away. He also said the total count of tells is holding steady rather than falling: labs strip out the best-known ones and fresh habits surface in their place, different for every release. He attributed that to scale, arguing that models with billions of parameters outrun any finite set of tests a lab can run before shipping.
Both labs had claimed progress on exactly this front. Anthropic's Opus 5.5 announcement said the model communicates more naturally than its predecessors and that early users found its writing clearer and easier to follow. OpenAI made a similar pitch for the GPT-6 versions of Sol and Luna, promising more clarity, less jargon and fewer odd turns of phrase.
What it means for AI detection
The practical lesson for publishers and detection vendors is that single-signal heuristics age fast, because any tell famous enough to be useful is also famous enough to be sanded off during post-training. Graphite says the tell data and the raw article corpus are available to download, which hands classifier builders a per-model vocabulary and hands the labs a checklist. Our earlier report on Pew's finding that 35% of post-ChatGPT web pages show signs of AI authorship indicates how much text the question now covers.
FAQ
Is the em dash still a reliable AI detector?
No. Graphite measured Opus 5.5 at 0.015 em dashes per 1,000 words, 99% below Opus 5, and found that Gemini 3.1 Pro has almost eliminated the mark. GPT-6 Astra uses it 88% less often than the pre-ChatGPT human corpus, so its absence now says more about recent training than its presence does.
How many AI writing tells did Graphite find in total?
The September report identified 12,877 unique words, phrases and frames used at least twice as often by AI as by humans across nine models. Individual models accounted for 2,355 to 3,746 each, and about 65% of the tells were specific to one model family.
Can the data be checked independently?
Yes. Graphite states that the tell data and the raw article corpus can be downloaded from the report page. The October follow-up applies the same method to Claude Opus 5.5 and lists its per-phrase ratios against the human baseline.
