The Em Dash Is Not Proof Of AI Writing

Somewhere in 2025 the internet convicted a punctuation mark. Post anything containing an em dash and someone will announce that a machine wrote it, usually with the confidence of a person who has just solved something.

It is a bad heuristic applied with total certainty, which is the worst combination. And it has real costs: writers who have used the mark for twenty years are now being accused, and students are being flagged over typography.

What The Measurements Actually Show

The frequency difference is real and much smaller than the discourse implies. Controlled comparisons put a widely used model at roughly 10 em dashes per 1,000 words against a matched human baseline near 3 per 1,000. Three times as often sounds damning until you notice the overlap: plenty of human writers, particularly newsletter writers and anyone raised on nineteenth-century prose, sit comfortably above the model average.

A signal with that much overlap is close to useless as a per-document test. It says something about a corpus of a million documents and almost nothing about the paragraph in front of you.

The situation also moved. Since late 2025 several assistants suppress em dashes by default or on request, and custom instructions can switch them off entirely. So the heuristic now flags a subset of models, a subset of settings, and every human who happens to like the mark.

The Tells That Hold Up Better

If you actually need to judge whether text was drafted by a machine and never edited, punctuation is the wrong layer to look at.

Cadence uniformity. This is the strongest one. Sentence after sentence landing in the same 18 to 24 word range, paragraph after paragraph running three sentences. Human writing lurches. It has a four-word sentence after a forty-word one because the writer got impatient.

Symmetry that reality does not have. Three benefits, each with three sub-points, each of equal length. Real topics are lopsided, because some parts matter more.

Hedged non-claims. “Results may vary depending on your specific requirements.” A sentence that cannot be wrong is a sentence that says nothing, and unedited drafts are full of them.

Absence of the specific. No figures, no dates, no names, no incident anyone lived through. This is the tell that actually matters, and it is the one we cover in small tells that make writing look AI generated.

The pivot paragraph. “However, it is important to consider the other side.” Placed exactly where a structure would put it, containing no actual counterargument.

Notice that all five are about substance and rhythm rather than characters. That is also roughly where the platforms landed: Google’s guidance on AI-generated content makes no claim about detecting authorship and objects to content produced without value, not to content produced with a tool. That is not a coincidence. Character-level tells are trivially removable with a find-and-replace, which is why they are the ones people fixate on and the ones worth the least.

What This Looks Like In Practice

Before:

Local processing offers meaningful advantages for teams handling sensitive documents. Performance is generally acceptable on modern hardware, though results may vary depending on your specific configuration. It is worth considering whether this approach aligns with your requirements.

After:

Local processing keeps the draft on the machine. It also needs about 2 GB free for a small model and falls over on anything pre-2013 without AVX2. We tested it on the four oldest laptops in the office and two of them could not run it at all.

The second version has no em dashes either, and nobody would mistake it for a machine draft. Not because of punctuation, but because it commits to numbers, admits a failure, and varies its sentence length. Those three properties are hard to fake and easy to produce if you actually know the subject.

Why This Blog Bans Em Dashes Anyway

Full disclosure, since it would be strange not to: Wrivio’s rewrite prompt forbids em dashes in its output, and this blog does not use them.

That is a house style decision, not a claim that the mark is a machine artifact. A comma, a colon, or two sentences almost always produce a cleaner result, and the mark is heavily overused as a way to avoid deciding what the relationship between two clauses actually is. We picked a rule and kept it, and the model follows the rule because a rewriter that quietly imposes its own punctuation habits on your writing is doing something you did not ask for.

A Wrivio Context for this could say:

Rewrite this to vary sentence length deliberately. Break at least one long sentence into two, and combine at least one pair of short adjacent sentences where they express a single idea. Delete any sentence that hedges without adding a condition. Keep every figure, name, and date exactly as written. Do not add examples, statistics, or claims that are not in the original.

Press Ctrl+Shift+Space, paste a paragraph, and read the diff. Watch that a hedge you deliberately included, because the claim genuinely is uncertain, did not get deleted along with the empty ones.

The Thing To Take From This

If you are being accused, the answer is not to purge your punctuation. It is to include something only you could know. A specific number, a date, an outcome that went badly. Machine-shaped text is not identifiable by what characters it uses; it is identifiable by having nothing at stake.

And if you are the one accusing, consider that detectors and typography heuristics are wrong often enough that acting on them can be genuinely unfair, which is the argument in what AI detectors actually prove.

Common Questions

Do AI models use more em dashes than humans?

On average yes, by roughly a factor of three in controlled comparisons, but the distributions overlap heavily. That makes it a weak signal for judging any individual document.

Can I stop ChatGPT using em dashes?

Yes. Since late 2025 custom instructions can suppress them, and several models now avoid them by default, which further weakens the mark as a detection heuristic.

What is the strongest tell in 2026?

Cadence uniformity: long runs of similarly sized sentences and paragraphs. It is much harder to remove than punctuation because it comes from how the text was generated rather than what it contains.

Should I remove em dashes from my writing?

Only if you want to. It is a legitimate style choice, and it is not evidence of anything either way.

Download Wrivio for Windows to vary the rhythm of a draft that reads too evenly, with a diff showing every sentence that changed.