The em dash is not the problem

Em dashes. Trailing participles. Synonym substitution. The rule of three.

Yet another LinkedIn post making the rounds this week proposes a checklist as a way to detect AI-generated writing: count the em dashes, count the trios, count the parallelisms, and the number tells you whether a person or a machine wrote the draft.
Every item on the list is real. Journalists, editors, speechwriters and columnists have used them for decades – because they work. They give prose rhythm, emphasis, compression, release.

To call them AI “tells” is to mistake AI’s training data for AI’s signature.

AI was trained on published professional writing. Of course it reads like it. Strip every technique AI has learned to imitate and you end up writing badly, to prove that you are human.

There is also a practical problem. Surface tells are easily prompted away. Ask a model to avoid em dashes, vary sentence length and drop the neat threes, and it will. The checklist produces false positives on human editors and false negatives on AI that has been told not to leave fingerprints.

“Count the em dashes” has a shelf life measured in months.

Meanwhile, the AI “tells” that do real damage to organisational writing are not being counted at all.

Take AI’s habit of producing risk-free prose. Every sentence reads as if it has been pre-approved by a committee that does not exist. Our commitment to stakeholder engagement reflects the values that have always defined who we are. That line could appear in any annual report, for any company, in any sector. Strip the word “stakeholder” and it could be a wedding speech. For a brand, this is not just a style failure. It tells readers that no one is really standing behind the message.

The second is more slippery: generic specificity. AI writes “a busy Singapore coffee shop”. A real writer either names the actual place, finds the detail that matters, or cuts the café because it is doing no work. The nouns are concrete but interchangeable. The piece feels grounded, but it isn’t.

The same smoothing shows up in tone. AI can produce anger, concern, excitement, admiration, urgency. What it struggles to sustain is movement. Anger flattens into concern. Awe flattens into appreciation. Criticism arrives padded and pre-apologised for. Three paragraphs in, the emotional temperature barely shifts. Real writing sharpens, pauses, doubles back, goes dry, withholds, presses harder.

Then comes the compulsive “both-sides reflex”. AI has been trained to sound reasonable, and reasonable often comes out as padded neutrality. Even when the author has a clear point of view, a dutiful “of course, others would argue” pops up. This used to pass for balance. In AI-assisted opinion writing, it reads as a retreat: the sentence backing away from the claim it was meant to make.

These tells are harder to count than em dashes. They cannot be reduced to a number per 500 words. They show up not in any single sentence but in the cumulative impression the piece makes – the sense that the prose has been smoothed, hedged, balanced, and pre-approved into saying nothing in particular.

This is what costs organisations. The damage we see most often in editorial work is not bad AI prose. It is generic AI prose that has been waved through because it passed every surface check and now sits on a brand’s website, in a CEO’s keynote or in an investor letter, saying nothing in particular.
The em dash is not the problem. It never was.

If the prose is not making a claim that anyone in the organisation is prepared to own, we are debating the wrong thing.