The Em Dash Isn't the AI Tell It Used to Be

By early 2025, the em dash had become internet shorthand for AI-generated text — see one in a work email or a student essay, and the accusation wrote itself. The reasoning wasn't baseless: large language models genuinely do reach for em dashes more often than typical human writers. But "more often" and "reliably" turned out to be very different claims, and the gap between them is exactly what 2026 data has started to expose.

Editing a draft and want to spot overused punctuation patterns at a glance? ClearText Editor highlights your text as you write, nothing uploaded anywhere.

Why AI Models Reach for the Em Dash

The explanation isn't stylistic preference — it's a byproduct of how these models get trained. During reinforcement learning from human feedback (RLHF), a model is rewarded for sounding helpful, nuanced, and conversational. The em dash is a cheap way to get there: it lets the model tack extra context onto a sentence without restructuring the grammar around it, which is computationally simpler than composing two well-formed sentences. The model isn't choosing a tone — it's finding the path of least resistance through next-token prediction, and that path runs through the em dash more often than human writing does.

What the 2026 Numbers Actually Show

Recent analysis of AI-generated versus human-edited text puts em dashes in roughly 18.5% of AI outputs, against 7.1% of human-written text — a real, measurable gap, but only about 2.6 times more common, not the smoking gun the internet treated it as. A single em dash in an email proves nothing; a document that leans on them in every paragraph is a weaker but not worthless signal.

Where the Real Tells Moved

As the em dash got famous, two things happened at once: AI labs tuned models to vary their punctuation more, and human writers got self-conscious enough to edit dashes out of their own writing. Both trends weakened the signal. What hasn't gone away is a set of structural and phrasing habits that are harder for a writer to notice and edit out of their own prose, because they read as "fine" rather than "AI."

Hedging verbs like ensuring, highlights, supports, and reflects show up disproportionately as models try to sound measured without committing to a specific claim. Multi-word phrases carry an even stronger signal — "rather than" is one of the most consistent tells identified in 2026 analysis, turning up thousands of times more often in AI-generated samples than in human-edited equivalents. And the single most recognizable structural pattern is a near-template: "X plays a crucial role in shaping Y," a sentence shape so common in AI output that it's become its own kind of tell, regardless of what fills in X and Y.

There's an uncomfortable asymmetry here: human writers have used em dashes for centuries, long before any model existed to imitate them — and now some of those same writers are self-editing their own natural punctuation out of fear of being flagged. Meanwhile the stronger, more reliable tells are quieter structural habits most readers would never consciously notice. The mark everyone learned to watch for was never the strongest signal; it was just the most visible one.

What This Means in Practice

Treating any single punctuation mark or phrase as proof of AI authorship is a mistake in both directions — it produces false accusations against human writers with a particular style, and it's trivially easy to route around once a specific tell becomes well known. The more durable approach, for anyone actually trying to sound distinctly human rather than game a detector, is the boring one: vary sentence length and structure, let some sentences stay short and declarative, and read a draft back out loud. AI output tends to flatten toward a narrow band of rhythm and phrasing; human writing — edited by an actual person — rarely does.

---

The honest summary: em dashes are measurably more common in AI writing, but only by about 2.6×, which is far from proof on its own — plenty of human writers use them constantly, and plenty of AI output doesn't. The tells that have replaced it as of 2026 are quieter: hedging verbs, "rather than," and formulaic role-in-shaping constructions that most readers would never flag by eye. The signal moved from the mark everyone was watching to the habits nobody was.