The Em Dash

An em dash is now enough to make some readers suspicious. Writers have begun removing it from their work, sometimes leaving small imperfections behind instead, as though a typo might serve as proof of human presence.

People are not imagining the pattern. Em dashes really have become more common since generative AI entered everyday writing. But spotting one in a text is something else. There are signals. None of them is conclusive.

The suspicion is already producing stranger effects. In a recent Wall Street Journal piece on AI and book publishing, WME agent Angeline Rodriguez described encountering something “honestly badly written” as “such a breath of fresh air now.” Beneath the Journal’s Instagram post, one commenter wrote: “The better the writing, the more it is suspect.”

Once a signal is recognised, it can also be removed. A writer who has always used em dashes can stop, while someone relying heavily on AI can simply ask it never to use one. Human writers are learning how not to look like machines at the same time that machines are learning how to look less like themselves.

Others beneath the same post responded differently. “Do not dumb down your writing,” one wrote. Another wrote: “That’s why I write my manuscripts on a typewriter.”

Maybe we have been looking for the human in the wrong place.

Who Wrote What?

“AI-written” can describe very different kinds of involvement. One person may hand over the subject and accept what comes back. Another may bring the idea, research and argument, using AI only to test or articulate what is already theirs.

One person can originate the ideas and make the important decisions while much of the final wording comes through AI. Another can take the ideas and argument from AI, then rewrite every sentence until it sounds unmistakably like them. The second may look far more human on the page. Which one is more of an author?

Words are not ideas, ideas are not decisions. This is why “70 per cent AI” tells us less than it appears to. A person can write every sentence without originating much of what the piece says. Another can contribute relatively little of the final language while determining almost everything the piece is trying to do.

Editors have always changed other people’s texts, sometimes extensively. Speechwriters put words into someone else’s mouth. Ghostwriters can produce whole books. Translators reconstruct a work in another language. Researchers and assistants contribute things a byline never reveals. Authorship has long been distributed. AI did not invent that ambiguity. It made it visible, scalable and accessible to almost everyone.

What once required several people can now happen inside one conversation. Research can turn into editing, editing into rewriting, and rewriting into something close to ghostwriting. The boundaries are not always announced when they are crossed.

Nor does the relationship have to begin again with every prompt. Previous work, a style guide, a private lexicon and repeated corrections can teach a model what someone prefers and, more interestingly, what they repeatedly reject. It is possible to build something closer to an editorial system than a collection of prompts.

Uploading a few pages does not make anyone Virginia Woolf. But it does make style a less reliable witness.

A Borrowed Voice

A text can resemble someone without coming from them. A text can come from someone and no longer resemble them. A model can be guided towards another writer’s style. Heavy rewriting can also pull a writer away from their own, smoothing out some of what made the voice recognisable in the first place.

“This sounds like AI” is a less useful verdict than it first appears. By what standard? Which one? Doing what? The model that supplied an entire argument and the one that corrected three sentences have both been “used”. A blank prompt and months of accumulated editorial instructions are described by the same verb.

Where Is the Author?

An em dash is easy to inspect. So is a sentence that feels too smooth, too familiar or somehow not quite right. The less visible parts of writing are harder to inspect: knowing what is worth pursuing, what does not belong, or when the piece needs to change direction.

Before the sentence was finished. James Joyce’s working draft for the “Circe” episode of Ulysses, 1920.

None of this makes the words irrelevant. A writer has always been more than the person with the idea, but writing is also what happens in the making of the sentence itself. The person with the idea, the person shaping the argument and the one producing the language no longer have to be the same person — or even people.

We began with punctuation, trying to decide whether a machine had written a sentence. What we end up asking is what “writing” means when thinking and writing can be separated.

But can they?

A writer does not always have the thought first and find the words afterwards. Sometimes the sentence is where the thought is found.

If so, what exactly are we handing over when we ask AI to write it?

A Few Numbers

4.23% → 11.58% → 20.3%

In a study of nearly 70,000 medRxiv preprints, the share of Discussion sections containing at least one em dash rose from 4.23% before ChatGPT to 11.58% afterwards, reaching 20.3% in 2025. The researchers describe the em dash as a population-level marker, not a way to identify AI use in an individual text.

1 in 300

In a 2025 study, five people who frequently used LLMs for writing were asked to distinguish human-written texts from work produced by GPT-4o, Claude and o1. Their majority judgment misclassified just one of 300 texts. They were not looking only for punctuation or vocabulary; their explanations also referred to qualities such as originality, formality and clarity.

66.5 percentage points

A 2026 study found that heavy AI rewriting could reduce measurable signals of individual authorship by as much as 66.5 percentage points in personal blog writing. Even instructions to preserve the writer’s voice did not prevent much of that loss.

Human Authored, with conditions

The Authors Guild expanded its Human Authored certification in 2026, while still allowing some uses of AI. It later revised the definition after acknowledging that terms such as “brainstorming” and “outlining” could cover AI involvement substantial enough to replace human creative work.