In short: Claude Opus 5.5 cut its use of em dashes by 99%, but a new study still counts 2,548 AI writing tells in its text. Cutting em dashes isn’t enough to make writing read as human.
Table of Contents
Graphite, a digital marketing agency, gave ten different AI models and a set of human writers the same assignment: read a summary of a pre-ChatGPT article, published before November 2022 so no model could have been trained on anything resembling it, and write a full article from that summary using one fixed prompt. Across 9,974 matched topics, the resulting dataset let researchers measure something more specific than “does this sound like AI”: exactly which words, phrases, and sentence patterns show up in model output at least twice as often as they do in comparable human writing. VentureBeat’s report on the findings breaks down what changed, and what didn’t, between Anthropic’s Claude Opus 5 and Opus 5.5.
Why the Em Dashes Number Only Tells Part of the Story
Opus 5 used em-dashes at a rate of 2.92 per 1,000 words. Opus 5.5 dropped that to 0.015 per 1,000 words, a 99% reduction, close enough to zero that em-dash counting, the single most common manual “is this AI-written” heuristic circulating online, would now fail to flag Opus 5.5 output almost every time. That’s a genuine, measurable change in how the model writes. But Graphite’s researchers, led by CEO Ethan Smith and chief AI officer Gregory Druck, didn’t stop at em-dashes.
Their full list of tracked patterns still identifies 2,548 words, phrases, and sentence constructions that Opus 5.5 uses at least twice as frequently as human writers covering the same topics: specific transition phrases, hedging constructions, and sentence-opener patterns that never got the public attention em-dashes did, because nobody wrote a viral post about them.
The overall word-distribution divergence score, Graphite’s broader measure of how different a model’s word choices are from human writing across the whole vocabulary, improved by 19%, moving from 0.064 to 0.052. That’s real progress. But it’s progress on a score that’s still well above zero, which is the quiet point underneath the headline number: fixing the one tell people learned to look for doesn’t fix the ones nobody’s looking for yet.
Why this matters more for a small business blog than a research paper
Most coverage of AI detectability research is aimed at academics or enterprise content teams running their own classifiers. For a small business publishing its own blog, whether written by a person, an AI tool, or some mix of both, the practical question is narrower: does it matter if a reader, or a search engine, can tell. On the reader side, the answer is reputational rather than technical. Content that reads as formulaic, the same sentence-opener patterns, the same hedge words, the same transition phrases repeating across every post, erodes trust in a blog’s expertise over time even if no single post gets flagged, the same way a human writer who overuses one phrase eventually gets noticed by a regular reader.
On the search and AI-answer-engine side, there’s no confirmed evidence that Google or any major search engine penalizes content purely for being AI-written; Google’s own guidance has consistently focused on whether content is helpful and original, not on its authorship method. What the engines do reward, and what formulaic AI patterns tend to work against, is content with genuinely distinct structure and specific, concrete claims rather than interchangeable phrasing that could belong to any article on the topic. A business blog chasing both human trust and search visibility ends up needing the same fix either way: fewer of the patterns Graphite’s study is measuring, regardless of whether a human or a model produced the first draft.
The methodology is worth understanding before trusting the numbers
Graphite’s approach is designed to isolate writing style from research ability. Every model works from the same input, so differences in output reflect how each one writes, not what it knows. That matters for anyone deciding whether this problem is getting better on its own over time. The 99% em-dash reduction and 19% divergence improvement between Opus 5 and 5.5 suggest it is, gradually, model generation by model generation, but the 2,548 remaining tells show there’s a long tail that hasn’t moved nearly as fast.
It’s also worth separating two different kinds of tell the study captures: word- and phrase-level repetition, which is what the 2,548 figure measures, and structural repetition at the sentence or paragraph level, like uniform sentence length, which a word-frequency study like this one isn’t designed to catch at all. A model can reduce its favorite words and phrases significantly, the way Opus 5.5 clearly has, while still writing every sentence in a section at roughly the same length and rhythm, a pattern human editors notice intuitively long before any word-frequency count would flag it.
What a reasonable self-edit pass actually catches
Most publicly circulated advice on spotting AI writing, overused transition words, suspiciously perfect grammar, a lack of personal anecdotes, is accurate as far as it goes but incomplete, because it’s checking for the tells that were obvious two model generations ago. A more useful self-edit pass for anyone publishing AI-assisted content checks for the structural patterns Graphite’s research keeps surfacing: headers that follow the same template across an entire post, sentences that cluster tightly around the same length with no short or long outliers, and transition or hedge phrases repeating more than once or twice in a single piece.
None of these require a paid detection tool, which is worth saying plainly since most of the free “AI detector” sites that rank for this topic are neither accurate nor necessary for a business just trying to self-edit its own content. Reading a draft specifically looking for repeated structural patterns, not just typos, catches most of what actually reads as formulaic.
A Quick Em Dashes Self-Audit Checklist
| Pattern Graphite flags | What it looks like in practice | Quick fix |
|---|---|---|
| Repetitive sentence openers | Several paragraphs in a row starting with “This means,” “It’s worth,” or similar | Rewrite at least one opener per paragraph to start with a concrete detail instead |
| Uniform sentence length | Most sentences in a section land within the same 15-25 word band | Split one sentence into two, or let a short sentence stand alone |
| Formulaic headers | Headers that all follow “What X Means” or “The Real Y” | Vary header style between statements, questions, and short phrases |
| Hedge-word stacking | “Genuinely,” “actually,” or “real” appearing more than once or twice per post | Cut the hedge word; the sentence usually works without it |
| Generic transitions | “It’s worth noting that…” or “In today’s digital landscape…” opening a paragraph | Replace with the actual point the paragraph is making |
None of these checks require specialized software. A five-minute read-through of a finished draft, looking specifically for these five patterns rather than reading for meaning, catches most of what separates formulaic AI output from something that reads like it was written by someone who knows the topic firsthand.
The practical bar isn’t “undetectable,” it’s distinct
Chasing an AI detection score down to zero misses the actual goal. A business blog doesn’t need to pass as human-written for its own sake; it needs to read as specific, useful, and worth a reader’s time, which happens to be largely the same target as avoiding the patterns Graphite is measuring. AllCloudHost’s own coverage of how AI chat tools surface in Search Console data runs into a related version of this question: the tools are improving quickly, the underlying patterns worth watching for keep shifting, and the right response for a small business is a habit of periodic re-checking, not a one-time fix.
Opus 5.5 closing its most famous tell in one release and still carrying 2,548 others is a reminder that whatever “sounds like AI” means right now will keep moving, and any editing checklist built around today’s list of tells needs revisiting once the next model generation changes which patterns are common.