AI systems can produce em dashes because punctuation is part of the text patterns they learn and predict. That is the limited, evidence-based answer. The public sources below do not prove a single cause for one model's punctuation habits, and an em dash is not proof that AI wrote a sentence.
There is a reasonable general explanation for why a language model can output an em dash. There is not enough public evidence to name one exact cause or a universal frequency.
What model makers sayOpenAI explains that language models learn relationships in text and predict likely next tokens; punctuation can be one of those tokens.
What we can inferA model may output an em dash when that mark fits its prompt and surrounding text. This does not tell us why one exact response used it.
What it isThe Unicode Standard identifies the em dash as U+2014 and describes it as a break in the flow of a sentence.
When it helpsThe Microsoft Style Guide uses it for a break, clarification, amplification, or parenthetical remark, while warning that overuse can hurt readability.
No supported multiplierWe found no cited first-party benchmark supporting a claim that one named model uses em dashes ten times more often than another.
No supported single causeTraining data, post-training, system instructions, the user's prompt, and sampling settings can all affect output. The cited public material does not isolate which factor caused a specific dash.
An em dash is not reliable proof of AI writing. It is ordinary punctuation used by human writers and language models. Even broader AI detectors can mislabel human text, according to OpenAI's detector guidance. Treat the count as a writing metric, not an authorship verdict.
The live metric counts only the exact U+2014 em dash. It reports the count and em dashes per 1,000 English word units. It does not estimate whether the text is AI-written.
The rule is fixed and visible, so two people pasting the same text get the same result.
em-dash density = U+2014 count ÷ English word-unit count × 1,000
—. Do not count en dashes, hyphens, or minus signs.The goal is not to ban a useful punctuation mark. The goal is to stop the draft from using one mark as an all-purpose connector.
Use it for a real interruptionKeep one when the aside changes the rhythm, adds a sharp contrast, or gives the sentence a useful turn.
Limit the densityOne strong em dash can work. Several in one paragraph usually means the sentence structure is doing too little.
Use a periodIf the second half is a full thought, split the sentence and commit to it.
Use a commaIf the dash only marks a small pause, a comma is usually enough.
Use a colonIf the second half explains or names the first half, a colon is cleaner.
Use nothingIf the dash only adds drama, delete it and tighten the sentence.
A natural edit usually removes the pile-up, splits one thought, and keeps only the detail that matters.
Our platform is fast — incredibly fast — and designed to scale — effortlessly — so your team can focus on what matters most — building great products.
Our platform is fast and built to scale. Your team can spend less time on infrastructure and more time building the product.
This is a writing cleanup tool, not a way to trick AI detectors. Em dashes are not AI evidence by themselves, and no punctuation edit can prove who wrote a text. Use the checker to make a draft clearer, less formulaic, and easier for a human reader to trust.
Em-dash use is one of the browser checker's 12 literal signals. The writing-pattern guide explains why no single signal proves authorship. If the checker also finds triads, stock vocabulary, hedging, or formal transitions, review those in context.
For the full hub, read the AI Writing Patterns guide. To use the main browser tool, open Blader Humanizer and paste your draft.
These first-party and standards sources support the narrow claims on this page:
U+2014 and its sentence-break use.Limit: none of these sources establishes a universal em-dash rate, a ten-times model comparison, or a claim that adding digitized books caused a particular model's dash frequency. Those stronger claims were removed.
Paste a draft, run Detect Patterns, then review and apply only the changes that preserve your meaning.