The Em Dash Problem: What Actually Gives Lazy AI Writing Away
The em dash is the most famous giveaway of AI writing, what researchers call a “tell,” and it's no longer a reliable one. In a September 2026 analysis of 90,000 AI-written articles by the growth agency Graphite (nine models, 10,000 articles each, all given the same prompts), GPT-6 Astra used em dashes about one-eighth as often as human writers, Gemini 3.1 Pro nearly stopped using them, and Claude Opus 5 landed right around the human rate (Graphite, 2026).
If your whole method for spotting lazy AI writing is hunting for one punctuation mark, you're chasing a moving target. There's a giveaway that doesn't move, and it's far more useful to you.
The most reliable giveaway of lazy AI writing isn't a punctuation mark. It's sameness: writing that could have been produced for any business.
AI Writing Tells Keep Changing
A study of 15.1 million biomedical abstracts shows how fast an AI fingerprint can spread. In 2024 abstracts, the word “delves” showed up about 28 times more often than expected, and the pattern was strong enough that the researchers estimated at least 13.5% of those abstracts had been processed with AI (Kobak, González-Márquez, Horvát, and Lause, 2025). For a while, “delve” was the tell everyone knew.
Then the tells moved. Graphite found that 65% of the tells it measured are unique to one model family, and that between the two latest versions of the same model, 55 to 72% are unique to one version (Graphite, 2026). A checklist of AI giveaways from last year is already out of date. That's also why advice to ban a list of “AI words” only works until the next update. That part is my inference from the data, not something Graphite tested. Graphite's own researchers add a warning: they don't recommend judging authorship by individual tells at all, and point to AI detectors instead.
What Doesn't Change Is Sameness
Here's what holds up across models and versions: AI writing sounds like other AI writing. In Graphite's analysis, every model's word choices were closer to at least one other model's than to human writing, and AI sentence lengths varied much less than human ones (Graphite, 2026). One caveat: every model got the same prompt, so some of that similarity comes from the task.
A peer-reviewed experiment found the same pattern in creative writing. Writers who got AI story ideas produced stories rated as more creative and better written, but those stories were more similar to each other than stories written by humans alone (Doshi and Hauser, 2024). Each writer improves, the weaker ones most of all. Everyone sounds a little alike.
Picture three competing businesses asking AI for a friendly post about their spring special. They can easily end up with three near-identical posts. For a local business, that's the real risk. Not a punctuation mark. Sounding like everyone else. One caution: none of the research above tested a local business's own posts, so treat it as a guide, not a guarantee.
What Only You Can Add
Graphite also looked at what human writers do that its AI didn't, in its test setup. People speak in the first person, tell small anecdotes, and draw on personal experience, and they use exclamation marks more than a hundred times as often as each of the three models Graphite tested (Graphite, 2026). Most of that list is your advantage. Your own experience and your small stories are things a tool can't know about your business unless you tell it.
A Three-Minute Sameness Check
You don't need to hunt for tells. Run any AI draft through these four steps before it goes out. They're my own recommendations, built from the findings above, not a tested formula.
Swap the stock lines. Find any sentence that could sit on any other business's page and replace it with something specific: a street, a job, a price, a customer's first name with their permission.
Cut the stock closer. Formulaic closing lines and filler like “it is important to note” are among the best-known tells (Graphite, 2026), and they're usually safe to delete. Graphite found tells as a group are less prominent in each model family's latest version, so this step won't catch everything on its own.
Change the rhythm. AI sentences tend to run the same length. Add one very short sentence and one longer, looser one.
Add one thing only you know. A real moment from this week, in your own words. That's the detail no tool can generate for you.
One Thing to Check Right Now
Pull up the last AI-assisted post you published and try just the first step. Find one sentence that could sit on any other business's page and replace it with one specific detail only your business would know. That's the whole fix, and it takes about three minutes.
Frequently Asked Questions
Are em dashes really a sign of AI writing?
Not reliably anymore. Graphite's 2026 analysis found em dash use ranging from about one-eighth of the human rate in one major model to about the human rate in another, and one model nearly stopped using them. That means humans use them too, and a missing em dash proves nothing.
Why does so much AI writing sound alike?
Different models share patterns that set them apart from human writing. Graphite found every model's word choices closer to at least one other model's than to human writing, though a shared prompt contributed. A peer-reviewed story-writing experiment also found AI-assisted stories were more similar to each other than human-only stories.
Should I still avoid em dashes in my own content?
We do, as a house rule, because some readers still treat them as a tell. But removing em dashes won't make generic writing specific. The specifics are what matter.
What's the fastest way to make AI writing sound like me?
Add details only you know, vary your sentence lengths, cut the stock closing line, and read it out loud. Those steps target the patterns the research above flags.
Here's the Plan
You don't need to hunt for tells. You need every post to carry something only your business could have written.
If you'd rather have someone set AI up to sound like your business and keep it that way: in Graphite's analysis, every major model leaned on its own recognizable patterns. Book a free, zero-pressure strategy session with me today. If you want us to build it for you, great. If not, you walk away with a free roadmap.
MAKE MY AI CONTENT SOUND LIKE ME
This piece is part of our AI series, “AI Isn't the Problem. How You Use It Is.” Start with the overview: The Marble and the Chisel: Why Bad AI Content Usually Isn't AI's Fault.
